Intelligent recognition method and system for biodegradable mulch degradation stage based on multispectral image fusion

Abstract

This invention provides an intelligent identification method and system for the degradation stage of biodegradable mulch film based on multispectral image fusion, belonging to the field of image processing technology. The method includes: acquiring multispectral images of biodegradable mulch film in the field, the multispectral images containing reflectance data from three bands (visible light, near-infrared, and shortwave infrared) acquired simultaneously; constructing a multispectral band observation matrix based on the multispectral images; calculating the local spectral variation coefficient of each pixel neighborhood in the multispectral images, using the local spectral variation coefficient as a criterion for spatial region division, and introducing a preset variation threshold determined statistically based on historical field multispectral data. This invention improves the anti-interference capability of mulch film degradation index calculation and the accuracy of degradation stage matching determination.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method and system for intelligent identification of the degradation stage of biodegradable mulch film based on multispectral image fusion. Background Technology

[0002] When identifying the degradation stage of biodegradable mulch films in the field based on multispectral images, existing methods usually use a uniform dimensionality reduction or feature extraction strategy to process the entire multispectral image. However, in actual field scenarios, there are essential differences in the spectral response characteristics between areas with significant mulch film degradation features (such as film surfaces with obvious degradation traces) and background noise areas (such as bare soil, crop residues, shadows, etc.).

[0003] Applying the same mapping method to the entire image indiscriminately leads to two problems: if a mapping method that emphasizes feature preservation is used, interference signals in background noise regions are also preserved and amplified, contaminating subsequent degradation index calculations; if a mapping method that emphasizes noise suppression is used, key spectral degradation information in regions with significant degradation features will be lost due to excessive smoothing, resulting in decreased accuracy in determining the degradation stage. Existing technologies lack a mechanism to select dimensionality reduction mapping strategies based on regional spectral characteristics, making it difficult to achieve an effective balance between preserving degradation features and suppressing background noise. Summary of the Invention

[0004] This invention provides a method and system for intelligent identification of degradation stages of biodegradable mulch film based on multispectral image fusion, which improves the anti-interference ability of mulch film degradation index calculation and the accuracy of degradation stage matching judgment.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: Firstly, a method for intelligent identification of the degradation stage of biodegradable mulch film based on multispectral image fusion, the method comprising: Multispectral images of biodegradable mulch film in the field were acquired. The multispectral images contained reflectance data of three bands, visible light, near infrared and shortwave infrared, which were acquired simultaneously. A multispectral band observation matrix was constructed based on the multispectral images. By calculating the local spectral variation coefficient of each pixel neighborhood in the multispectral image, the local spectral variation coefficient is used as the criterion for spatial region division, and a preset variation threshold determined based on historical field multispectral data is introduced. When the local spectral variation coefficient is greater than the preset variation threshold, the neighborhood of the current pixel is determined to be a region with significant mulch film degradation characteristics. For the multispectral band observation matrix of the region with significant mulch film degradation characteristics, a high-dimensional feature preservation mapping of the orthogonal decomposition family is performed to map the multispectral band observation matrix to the main degradation feature vector set. When the local spectral variation coefficient is less than or equal to the preset variation threshold, the neighborhood of the current pixel is determined to be a background noise region. For the multispectral band observation matrix of the background noise region, a low-dimensional smooth dimensionality reduction mapping of the statistical inference family is performed to map the multispectral band observation matrix to the background suppression feature vector set. The main degradation feature vector set and the background suppression feature vector set are aligned and fused at the feature level according to the original pixel spatial coordinates to obtain the regional differential fused feature tensor. The degradation index of the mulch film is calculated based on the regionally differentiated fusion feature tensor. The degradation index of the mulch film is matched and determined with multiple preset stage threshold intervals to obtain the corresponding biodegradable mulch film degradation stage label.

[0006] Furthermore, multispectral images of the biodegradable mulch film in the field are acquired. These multispectral images contain reflectance data from three bands: visible light, near-infrared, and shortwave infrared, acquired simultaneously. A multispectral band observation matrix is ​​constructed based on these multispectral images, including: Multispectral images of biodegradable mulch films in the field were acquired, and raw reflectance data of the visible light band, near-infrared band, and short-wave infrared band were extracted from the multispectral images respectively. Spatial coordinate alignment and radiometric normalization were performed on the raw reflectance data of the three bands to eliminate the dimensional differences and spatial pixel offsets generated when the sensors in different bands were collected, and the aligned multi-band reflectance data were obtained. Traverse the aligned multi-band reflectivity data, and perform vectorization stacking operation on the visible light reflectivity data, near-infrared reflectivity data and short-wave infrared reflectivity data in the same pixel spatial coordinates according to the preset band arrangement order to obtain the result of the vectorization stacking operation. Based on the results of the vectorized stacking operation, a multispectral band observation matrix is ​​constructed with pixel spatial coordinates as row indices and reflectance data of three bands as column elements.

[0007] Furthermore, by calculating the local spectral variation coefficient of each pixel neighborhood in the multispectral image, the local spectral variation coefficient is used as the criterion for spatial region division, and a preset variation threshold determined based on historical field multispectral data is introduced, including: Using the current pixel in the multispectral band observation matrix as the center, extract the multispectral band observation matrix data within a neighborhood window of a preset size; The ratio of the standard deviation to the mean of the reflectance data for the visible light, near-infrared and short-wave infrared bands within the neighborhood window is calculated respectively to obtain the local spectral variation coefficient of each band, and the comprehensive local spectral variation coefficient of the current pixel neighborhood is determined based on the local spectral variation coefficient of each band. A preset variation threshold is introduced, which is determined by statistical analysis of historical field multispectral data. The preset variation threshold characterizes the statistical boundary that distinguishes the degradation characteristics of plastic film from background noise.

[0008] Furthermore, when the local spectral variation coefficient exceeds a preset variation threshold, the neighborhood of the current pixel is determined to be a region with significant mulch film degradation characteristics. For the multispectral band observation matrix of this region, a high-dimensional feature-preserving mapping of an orthogonal decomposition family is performed, mapping the multispectral band observation matrix to the main degradation feature vector set, including: The multispectral band observation matrix of the region identified as having significant characteristics of plastic film degradation was extracted and used as input data for the high-dimensional feature preservation mapping. The input data is processed by feature mapping based on orthogonal decomposition family, which decomposes the multispectral band observation matrix into multiple orthogonal feature components. The eigenvalues ​​characterize the retention intensity of each orthogonal feature component for the spectral differences of mulch film degradation, and their values ​​are constrained by the covariance structure of the multispectral band observation matrix. Multiple orthogonal feature components are sorted according to the magnitude of their eigenvalues, and orthogonal feature components with eigenvalues ​​greater than a preset contribution threshold are selected as the main degradation feature components. The selected main degradation feature components and their corresponding original pixel spatial coordinates are recombined into vectors to obtain the main degradation feature vector set.

[0009] Furthermore, when the local spectral variation coefficient is less than or equal to a preset variation threshold, the neighborhood of the current pixel is determined to be a background noise region. For the multispectral band observation matrix of the background noise region, a low-dimensional smoothing dimensionality reduction mapping using a statistical inference family is performed, mapping the multispectral band observation matrix to a background suppression feature vector set, including: The multispectral band observation matrix of the region identified as background noise is extracted and used as input data for low-dimensional smooth dimensionality reduction mapping. The input data is processed by a family of statistical inferences to achieve smooth dimensionality reduction. The statistical distribution characteristics of the pixel reflectance data in the neighborhood are calculated. The smoothing weight represents the suppression strength of background noise, and its value is constrained by the variance of the reflectance data in the neighborhood window. Based on statistical distribution characteristics and smoothing weights, the multispectral band observation matrix of the background noise region is weighted, smoothed and aggregated to eliminate high-frequency spectral fluctuations and obtain low-dimensional smoothed feature components. The low-dimensional smooth feature components and their corresponding original pixel spatial coordinates are reduced in dimension and recombined to obtain a background suppression feature vector set.

[0010] Furthermore, the main degradation feature vector set and the background suppression feature vector set are aligned and fused at the feature level according to the original pixel spatial coordinates to obtain the region-differentiated fused feature tensor, including: Based on the pixel position identifiers carried in the main degradation feature vector set and the background suppression feature vector set, the corresponding original pixel spatial coordinate indexes are extracted to obtain the spatial coordinate mapping set; Based on the spatial coordinate mapping set, the feature elements of the main degradation feature vector set and the background suppression feature vector set are reordered by gridding to eliminate the spatial discrepancy caused by the region division and obtain the coordinate aligned feature matrix. Based on the coordinate-aligned feature matrix, a feature space recombination process based on tensor algebra family is performed to map the coordinate-aligned feature matrix along the channel dimension to obtain the initial fused feature tensor. The recombination weight represents the aggregation strength of spatial continuity, and its value is constrained by the original number of channels of the multispectral band observation matrix. Based on the initial fusion feature tensor, topological interpolation smoothing aggregation based on manifold embedding families is performed on the feature vectors at the boundaries of different regions within the tensor to eliminate spectral faults caused by feature-level splicing, thus obtaining a regionally differentiated fusion feature tensor.

[0011] Furthermore, based on the initial fused feature tensor, topological interpolation smoothing aggregation based on manifold embedding families is performed on the feature vectors at the boundaries of different regions within the tensor to eliminate spectral faults caused by feature-level splicing, resulting in a regionally differentiated fused feature tensor, including: Based on the initial fusion feature tensor, feature vectors at the boundaries of different regions within the tensor are extracted to obtain a set of boundary feature vectors; Based on the set of boundary feature vectors, a dimensionality reduction mapping process based on manifold embedding family is performed to map the high-dimensional boundary feature vectors to the low-dimensional manifold space, resulting in a set of coordinate points in the manifold space. The manifold curvature represents the degree of continuous change of the spectral features of the boundary region, and its value is constrained by the local gradient of the initial fused feature tensor. Based on the set of coordinate points in the manifold space, the geodesic distance between adjacent feature points is calculated, and topological interpolation is performed on the feature vectors between adjacent regions based on the geodesic distance to fill the spectral faults generated by feature-level splicing, resulting in an interpolated smooth feature vector set. Based on the interpolated smooth feature vector set, the interpolation result is mapped back to the original high-dimensional feature space, and tensor recombination is performed with the non-boundary region features in the initial fused feature tensor to eliminate spatial discretization bias and obtain the regionally differentiated fused feature tensor.

[0012] Furthermore, the degradation index of the mulch film is calculated based on the regionally differentiated fusion feature tensor. The degradation index is then matched with multiple preset stage threshold intervals to obtain the corresponding biodegradable mulch film degradation stage labels, including: Based on the regional differential fusion feature tensor, the multispectral band feature components contained in the regional differential fusion feature tensor are extracted to obtain the multispectral feature component set. Based on the set of multi-band feature components, a feature weighted aggregation process based on a family of variational optimizations is performed to map the set of multi-band feature components to a one-dimensional scalar space to obtain the initial mulch film degradation index. The aggregation weight characterizes the sensitivity of each band to the degradation state, and its value is constrained by the channel variance of the set of multi-band feature components. Based on the initial mulch film degradation index, multiple preset stage threshold intervals determined by statistical analysis of historical field multispectral data are introduced, and a numerical interval matching operation is performed to obtain the mulch film degradation index interval matching result. Based on the interval matching results of the mulch film degradation index, the mulch film degradation index falling into a specific interval is mapped to the corresponding status identifier to obtain the corresponding biodegradable mulch film degradation stage label.

[0013] Furthermore, based on the initial mulch film degradation index, multiple preset stage threshold intervals determined by statistical analysis of historical field multispectral data are introduced, and a numerical interval matching operation is performed to obtain the mulch film degradation index interval matching results, including: Based on the initial mulch film degradation index and multiple preset stage threshold intervals, the upper and lower boundary values ​​of each stage threshold interval are extracted to obtain the stage threshold boundary set. Based on the set of stage threshold boundaries, the initial mulch film degradation index is compared with the values ​​of each upper and lower boundary to determine the target threshold interval to which the initial mulch film degradation index belongs, and the interval assignment result is obtained. Based on the interval attribution determination results, the initial mulch film degradation index at the boundary of adjacent stage threshold intervals is subjected to smooth transition mapping processing based on fuzzy membership families to eliminate the identification mutation caused by stage jumps and obtain the corrected interval attribution determination results. Based on the corrected interval assignment results, the target threshold interval to which the initial plastic film degradation index belongs is determined, and the plastic film degradation index interval matching results are obtained.

[0014] Secondly, a smart identification system for the degradation stage of biodegradable mulch film based on multispectral image fusion includes: The acquisition module is used to acquire multispectral images of biodegradable mulch films in the field. The multispectral images contain reflectance data of three bands, visible light, near infrared and shortwave infrared, which are acquired simultaneously. A multispectral band observation matrix is ​​constructed based on the multispectral images. The calculation module is used to calculate the local spectral variation coefficient of each pixel neighborhood in the multispectral image, use the local spectral variation coefficient as the criterion for spatial region division, and introduce a preset variation threshold determined by statistical analysis of historical field multispectral data. The determination module is used to determine that the neighborhood of the current pixel is a region with significant mulch film degradation characteristics when the local spectral coefficient of variation is greater than a preset variation threshold. For the multispectral band observation matrix of the region with significant mulch film degradation characteristics, a high-dimensional feature preservation mapping of the orthogonal decomposition family is performed to map the multispectral band observation matrix to the main degradation feature vector set. When the local spectral coefficient of variation is less than or equal to the preset variation threshold, the neighborhood of the current pixel is determined to be a background noise region. For the multispectral band observation matrix of the background noise region, a low-dimensional smooth dimensionality reduction mapping of the statistical inference family is performed to map the multispectral band observation matrix to the background suppression feature vector set. The fusion module is used to perform feature-level alignment and fusion processing on the main degradation feature vector set and the background suppression feature vector set according to the original pixel space coordinates to obtain the regional differential fusion feature tensor. The processing module is used to calculate the mulch film degradation index based on the regionally differentiated fusion feature tensor, and to match and determine the corresponding biodegradable mulch film degradation stage label based on the mulch film degradation index and multiple preset stage threshold intervals.

[0015] The above-described solution of the present invention has at least the following beneficial effects: By introducing the local spectral variation coefficient as the criterion for spatial region division, the multispectral image is adaptively divided into regions with significant mulch film degradation characteristics and background noise regions. The high-dimensional feature preservation mapping of the orthogonal decomposition family and the low-dimensional smooth dimensionality reduction mapping of the statistical inference family are respectively applied to the two types of regions, achieving a differentiated processing effect of accurately preserving the regions that should be preserved and fully suppressing the regions that should be suppressed.

[0016] Based on this, feature-level alignment and fusion are performed, so that the final constructed regionally differentiated fusion feature tensor not only fully preserves the high-dimensional spectral details of the degradation front region, but also effectively filters out the interference noise of the background region, thereby improving the anti-interference ability of the film degradation index calculation and the accuracy of degradation stage matching judgment. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the intelligent identification method for the degradation stage of biodegradable mulch film based on multispectral image fusion provided in an embodiment of the present invention.

[0018] Figure 2 This is a schematic diagram of an intelligent identification system for the degradation stage of biodegradable mulch film based on multispectral image fusion, provided by an embodiment of the present invention. Detailed Implementation

[0019] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art.

[0020] like Figure 1 As shown, embodiments of the present invention propose an intelligent identification method for the degradation stage of biodegradable mulch films based on multispectral image fusion. The method includes the following steps: Step 1: Obtain multispectral images of biodegradable mulch film in the field. The multispectral images contain reflectance data of three bands: visible light, near infrared and shortwave infrared, which were collected simultaneously. A multispectral band observation matrix is ​​constructed based on the multispectral images. Step 2: Calculate the local spectral variation coefficient of each pixel neighborhood in the multispectral image, use the local spectral variation coefficient as the criterion for spatial region division, and introduce a preset variation threshold determined based on historical field multispectral data. Step 3: When the local spectral variation coefficient is greater than the preset variation threshold, the neighborhood of the current pixel is determined to be a region with significant mulch film degradation characteristics. For the multispectral band observation matrix of the region with significant mulch film degradation characteristics, a high-dimensional feature preservation mapping of the orthogonal decomposition family is performed to map the multispectral band observation matrix to the main degradation feature vector set. Step 4: When the local spectral variation coefficient is less than or equal to the preset variation threshold, the neighborhood of the current pixel is determined to be a background noise region. For the multispectral band observation matrix of the background noise region, a low-dimensional smooth dimensionality reduction mapping of the statistical inference family is performed to map the multispectral band observation matrix to the background suppression feature vector set. Step 5: Align and fuse the main degradation feature vector set and the background suppression feature vector set according to the original pixel spatial coordinates to obtain the regional differential fused feature tensor. Step 6: Calculate the mulch film degradation index based on the regional differentiated fusion feature tensor, and match the mulch film degradation index with multiple preset stage threshold intervals to obtain the corresponding biodegradable mulch film degradation stage label.

[0021] In this embodiment of the invention, a multispectral band observation matrix is ​​constructed by simultaneously acquiring reflectance data from three bands: visible light, near-infrared, and short-wave infrared. Spatial regions are divided based on the comparison between local spectral variation coefficients and preset variation thresholds. Then, high-dimensional feature preservation mapping of orthogonal decomposition families is performed on regions with significant mulch film degradation characteristics, while low-dimensional smoothing dimensionality reduction mapping of statistical inference families is performed on background noise regions. Finally, feature-level alignment and fusion processing are performed according to the original pixel spatial coordinates to obtain regionally differentiated fused feature tensors to calculate the mulch film degradation index and perform matching judgment. Therefore, this invention overcomes the technical problems in existing field multispectral image processing where mulch film degradation characteristics are easily confused with complex background noise, and global unified processing easily leads to the loss of degradation details or severe background interference. Thus, it achieves the goal of accurately preserving the high-dimensional spectral differences in mulch film degradation while effectively suppressing background noise and eliminating spatial discrepancies and spectral faults caused by feature fusion.

[0022] In a preferred embodiment of the present invention, step 1 above may include: Step 1.1: Acquire multispectral images of the biodegradable mulch film in the field. Extract raw reflectance data from the visible, near-infrared, and short-wave infrared bands of the multispectral images. Specifically, this involves using a multispectral imaging sensor mounted on a mobile platform or fixed observation frame in the field to capture vertically downward images of the ridges covered with biodegradable mulch film, obtaining raw multispectral image data containing the entire mulch film area. The multispectral imaging sensor simultaneously acquires signal response values ​​from three independent spectral channels—visible, near-infrared, and short-wave infrared—in a single exposure. The center wavelengths of each channel are set to 550 nm for the visible band, 850 nm for the near-infrared band, and 1650 nm for the short-wave infrared band, respectively. The full width at half maximum (FWHM) of each channel are 40 nm, 50 nm, and 30 nm, respectively. Read the digital count value corresponding to each pixel position from the raw multispectral image data channel by channel. The digital count value is a dimensionless integer value directly output by the sensor, representing the intensity of radiant energy received by the pixel in the corresponding band. Let the total number of rows of pixels in the multispectral image be... The total number of columns of pixels is For the visible light band, the acquired raw digital count values ​​constitute a OK Two-dimensional matrix of columns ,in Indicates the first Line 1 The digital count value of the visible light band of the column pixels. The value range is 1 to , The value range is 1 to Similarly, for the near-infrared band, a two-dimensional matrix of the original digital count values ​​is obtained. For the shortwave infrared band, obtain a two-dimensional matrix of the original digital count values. .

[0023] Then, using the radiometric calibration coefficients set at the factory of the multispectral imaging sensor, the raw digital count values ​​for the three bands were converted into surface reflectance data. The radiometric calibration coefficients were obtained by taking images under various radiance conditions in front of the multispectral imaging sensor in laboratory settings using a standard diffuse reflectance reference plate with known reflectance. A linear response relationship between the digital count values ​​and the incident radiance was established, and the conversion coefficient from digital count values ​​to surface reflectance for each band was finally obtained by measuring the actual reflectance of the reference plate under the same illumination conditions. The algebraic formula for the conversion calculation is: ; In the formula, For the first Line 1 The visible light band surface reflectance of a column pixel is dimensionless and ranges from 0 to 1. The dark current offset of the visible light band sensor is obtained by taking a dark frame image with the lens cap on and obtaining the average value of the dark field digital count. This is the radiation calibration gain coefficient for the visible light band, expressed in watts per square meter per steradian per micrometer per digital count. The surface downlight irradiance in the visible light band at the time of shooting is measured in watts per square meter per micrometer, and is obtained by a synchronously installed diffuse irradiance sensor. The solar zenith angle, measured in radians, is calculated using a solar position algorithm based on the date, time, and geographical latitude and longitude of the image taken. It also represents the near-infrared band surface reflectance. and shortwave infrared band surface reflectance The conversion calculation uses the same formula structure. Simply substitute the dark current offset, radiation calibration gain coefficient, and downlink irradiance parameter for each band. The output results are three... OK Two-dimensional reflectivity matrix of columns: Reflectivity matrix of visible light band Near-infrared band reflectivity matrix and the reflectivity matrix of the shortwave infrared band .

[0024] Step 1.2 involves performing spatial coordinate alignment and radiometric normalization on the raw reflectance data for the three bands. This eliminates dimensional differences and spatial pixel offsets caused by sensor acquisition in different bands, resulting in aligned multi-band reflectance data. Specifically, this includes performing spatial coordinate alignment. Since the visible, near-infrared, and short-wave infrared bands may be acquired by three independent detector arrays in a multispectral imaging sensor, there are slight sub-pixel-level displacement differences in the physical installation positions of the three detector arrays on the focal plane. Furthermore, the magnification and distortion parameters of the optical lenses for each band are not entirely consistent, leading to variations in the reflectance data at the same location. The pixel coordinates of the target points in the three-band images show deviations. To correct this deviation, a geometric calibration board with a black and white checkerboard pattern printed on its surface was used for sensor geometric calibration before the actual field shooting. The calibration board was placed flat on the surface of the field mulch film, and multispectral images containing the complete checkerboard pattern were captured. The Harris corner detection algorithm was used to extract the coordinates of the checkerboard corner points in the three-band images, establishing correspondence pairs between the corner points in the three-band images. Based on at least 12 representative, uniformly distributed corner point correspondence pairs, a perspective transformation model was used to estimate the spatial transformation matrix between the three-band images. The algebraic formula for the perspective transformation model is: ; ; In the formula, and The original pixel coordinates of the band to be corrected; The nine elements are a 3x3 perspective transformation matrix, where... The value is fixed at 1, and the other eight parameters are obtained by solving the least squares method based on the correspondence of corner points; and These are the transformed pixel coordinates.

[0025] Using the visible light image as a spatial reference, the near-infrared and short-wave infrared images are resampled using their respective perspective transformation matrices until they are strictly aligned with the pixel grid of the visible light image. During the resampling process, a bilinear interpolation algorithm is used to calculate the reflectance values ​​at sub-pixel positions to avoid jagged spatial artifacts introduced by nearest-neighbor interpolation. After spatial coordinate alignment, the reflectance matrices of the three bands have a completely consistent number of pixel rows. and number of pixel columns Furthermore, the same pixel coordinate position corresponds to the exact same ground target point in the three bands. Radiometric normalization is performed. Since there are natural differences in the range and statistical distribution characteristics of reflectance values ​​in different bands, and the field lighting conditions may fluctuate slightly during the shooting process, it is necessary to uniformly perform radiometric normalization on the reflectance data of the three bands to eliminate the difference in dimensions and finally obtain the aligned multi-band reflectance data.

[0026] Step 1.3: Traverse the aligned multi-band reflectance data. Perform vectorized stacking operations on the visible light reflectance data, near-infrared reflectance data, and short-wave infrared reflectance data located in the same pixel spatial coordinates, according to a preset band arrangement order. This involves determining the preset band arrangement order: the first channel is the visible light band, the second channel is the near-infrared band, and the third channel is the short-wave infrared band. , , The order of these bands is determined as follows: the visible light band is consistent with human visual perception and is placed first as the reference channel for easy manual verification; the near-infrared band is closely adjacent to the visible light band on the spectral axis, which is beneficial for calculating gradient characteristics; the short-wave infrared band is sensitive to the chemical bond vibrations of the mulch film material and is placed last as an independent channel for flexible subsequent weighting.

[0027] Perform pixel-by-pixel vectorization stacking operation, traversing the spatial coordinates of all pixels in the image, for the... Line 1 For each column pixel, extract the band reflectance value corresponding to that position from the three normalized reflectance matrices: from Extract the normalized visible light reflectance value, denoted as . ;from Extract the near-infrared normalized reflectance value, denoted as . ;from Extract the normalized reflectance value of the shortwave infrared radiation, denoted as . These three scalar values ​​are concatenated into a three-dimensional column vector according to a preset band order. The formula for this operation is as follows: ; In the formula, Given a 3x1 column vector, the superscript... This represents the vector transpose operation. This is the visible light channel value of that pixel. This is the near-infrared channel value for that pixel. This is the shortwave infrared channel value for that pixel.

[0028] During the traversal, the order is row-by-row followed by column-by-column: fixed row index. Increasing from 1 to Fixed column index in each row Increasing from 1 to This process generates a 3D spectral vector for each pixel location sequentially. Each generated 3D spectral vector is then associated with its corresponding original pixel spatial coordinates and recorded as a tuple. The tuple structure is as follows: This is to obtain the result of the vectorized stacking operation.

[0029] Step 1.4, based on the results of the vectorization stacking operation, construct a multispectral band observation matrix with pixel spatial coordinates as row indices and reflectance data of the three bands as column elements. Specifically, this includes: processing the obtained vector set... Included Expand each of the three-dimensional column vectors row by row in pixel traversal order to construct a column vector with a total of 3D columns. A two-dimensional multispectral band observation matrix with 3 columns. , will the The 3D column vector corresponding to each pixel The three elements are filled into the matrix in sequence. The Line, pixel index with spatial coordinates The mapping relationship between them is as follows: ; in, The value is , The value is , The value is .

[0030] matrix The structure is as follows: ; in, The total number of pixels, matrix The element in row 1 column equal to the The visible light reflectance value of the nth pixel, the nth element in row 2 equal to the The near-infrared reflectance value of the pixel, the th element in row 3 equal to the The short-wave infrared reflectance value of each pixel, i.e.:

[0031] in, and Depend on Through reverse mapping relationship and Confirm, symbol This indicates rounding down, and simultaneously constructs a matrix... Matching pixel coordinate index vector , its first The element stores the first... The original pixel space coordinates corresponding to the row data : ; Multispectral band observation matrix The row index is obtained through the coordinate index vector. Establish a one-to-one correspondence with the pixel space coordinates of the original image, matrix Each row represents the complete spectral observation value of a pixel in the visible, near-infrared and short-wave infrared bands, and the three columns correspond to the visible reflectance, near-infrared reflectance and short-wave infrared reflectance channels, respectively.

[0032] In this embodiment of the invention, the raw reflectance data of the visible light, near-infrared and short-wave infrared bands are extracted separately, and spatial coordinate alignment and radiometric normalization are performed on them. Then, they are vectorized and stacked in a preset order to construct a multispectral band observation matrix with pixel spatial coordinates as row indexes and band reflectance data as column elements. This overcomes the technical problems of dimensional differences and spatial pixel offsets that are easily generated when different band sensors collect data, as well as the lack of a unified structured expression for multi-source spectral data. In this way, the spatial and radiometric dimensional deviations between multi-band data are eliminated, and the high alignment and standardization of multi-source spectral data are achieved.

[0033] In a preferred embodiment of the present invention, step 2 above may include: Step 2.1: Using the current pixel in the multispectral band observation matrix as the center, extract the multispectral band observation matrix data within a neighborhood window of a preset size. This specifically includes: extracting the multispectral band observation matrix data from the constructed multispectral band observation matrix. In the process, each pixel is selected sequentially according to its row index as the current pixel to be processed. Let the spatial coordinates of the current pixel in the original image be... Corresponding matrix row index in Based on mapping relationship It is confirmed that, among them, The default neighborhood window size is a 15-pixel x 15-pixel square window centered on the current pixel, representing the number of image columns. This window covers the row direction. to The window consists of 15 rows and covers 15 columns (j-7 to j+7) in the column direction, containing a total of 225 pixels. The 15-pixel × 15-pixel window size has been verified through field measurements, achieving a balance between spatial locality and statistical stability: a window that is too small, such as 5 pixels × 5 pixels, contains only 25 sample points, resulting in an excessively large variance in the estimation of spectral statistics and instability in the calculation of local spectral variation coefficients; a window that is too large, such as 31 pixels × 31 pixels, contains 961 sample points, but it will cross the boundary between the mulch film area and the background area, causing spectral features of the two types of areas to overlap, weakening the accuracy of area division.

[0034] For the neighborhood window located at spatial coordinates Each pixel, where The value is , The value is Determine whether the coordinates are within the valid range of the image. or or or At this point, the pixel at that coordinate position is outside the boundary and does not participate in subsequent calculations; it is removed from the neighborhood window sample set. The matrix corresponding to the remaining valid pixels is then... Row indexes are determined one by one, row index The calculation formula is: ; From the matrix Extract row index The corresponding row vector yields the reflectance values ​​of the pixel in the three bands: ; in, This represents the reflectivity value in the visible light band. This represents the reflectance value in the near-infrared band. This represents the reflectivity value in the shortwave infrared band.

[0035] Summarize the three-band reflectance values ​​of all valid pixels within the neighborhood window to construct the current pixel. Corresponding neighborhood window multispectral data submatrix Let the number of effective pixels within the neighborhood window be... ,but The size is . The The first row in the corresponding window There are 100 effective pixels. The first column stores the reflectance value in the visible light band, the second column stores the reflectance value in the near-infrared band, and the third column stores the reflectance value in the short-wave infrared band. The value is .

[0036] Submatrix The data is split into three column vectors, each corresponding to a sequence of reflectance observations for the three bands within a neighborhood window: ,in, The The element in the first column of the row.

[0037] ,in, The The element in the second column of the row.

[0038] ,in, The The element in the 3rd column of the row.

[0039] These three column vectors , and That is, the current pixel Multispectral band observation matrix data within the neighborhood window.

[0040] Step 2.2: Calculate the ratio of the standard deviation to the mean of the reflectance data for the visible light, near-infrared, and short-wave infrared bands within the neighborhood window to obtain the local spectral variation coefficient for each band. Based on the local spectral variation coefficients of each band, determine the comprehensive local spectral variation coefficient of the current pixel's neighborhood. Specifically, this includes: For the visible light band neighborhood reflectance observation sequence, calculate its mean. This mean is calculated by summing the visible light band reflectance values ​​of all valid pixels within the neighborhood window, then dividing the sum by the total number of valid pixels S. The quotient is the mean reflectance of the visible light band. Calculate the standard deviation of the visible light band neighborhood reflectance observation sequence. The calculation process is as follows: For each valid pixel within the neighborhood window, first subtract the mean obtained in the previous step from its visible light band reflectance value to obtain the deviation of that pixel. Then, square the deviation to obtain the squared deviation value. Sum the squared deviation values ​​of all S valid pixels and divide by the total number of valid pixels S to obtain the average of the squared deviations. The square root of this average value is taken to obtain the standard deviation of the reflectance in the visible light band. The local spectral variation coefficient in the visible light band is defined as the ratio of the standard deviation to the mean. This is calculated by dividing the calculated standard deviation by the mean, and the quotient is the local spectral variation coefficient. When the mean of the visible light band is zero, the division operation is skipped, and the local spectral variation coefficient is directly assigned to zero to avoid division-by-zero anomalies.

[0041] For the near-infrared band neighborhood reflectance observation sequence, the same steps are followed: sum the reflectance values ​​of all effective pixels in the near-infrared band and divide by the total number of effective pixels, S, to obtain the mean reflectance of the near-infrared band. Then, subtract the mean reflectance value from the near-infrared reflectance value of each effective pixel to obtain the deviation. Square the deviations, sum them, divide by S, and take the square root to obtain the standard deviation of the near-infrared band. Divide the standard deviation of the near-infrared band by the mean reflectance value to obtain the local spectral variation coefficient of the near-infrared band. When the mean reflectance value of the near-infrared band is zero, the local spectral variation coefficient of the near-infrared band is set to zero. For the short-wave infrared band neighborhood reflectance observation sequence, the same steps are followed. Calculate the sum of the reflectance values ​​of all effective pixels in the short-wave infrared band and divide by the total number of effective pixels, S, to obtain the mean reflectance of the short-wave infrared band. Calculate the deviation of the short-wave infrared reflectance value of each effective pixel from this mean, square the deviations, sum them all, divide by S, and take the square root to obtain the standard deviation of the short-wave infrared band. The local spectral variation coefficient of the short-wave infrared band is obtained by dividing the standard deviation of the short-wave infrared band by the mean of the short-wave infrared band. When the mean of the short-wave infrared band is zero, the local spectral variation coefficient of the short-wave infrared band is also assigned a value of zero.

[0042] Obtain the local spectral variation coefficients for each of the three bands. , and back, The coefficient of variation of the local spectrum in the visible light band. The coefficient of variation of the local spectrum in the near-infrared band. The local spectral variation coefficient is the coefficient of variation of the current pixel neighborhood. The combined local spectral variation coefficients were calculated using a weighted root mean square synthesis method to fully reflect the joint contribution of the spectral variations of the three bands at different scales: ; in, , and These are the weighting coefficients for the three bands, satisfying... The weighting coefficients are determined based on the differences in sensitivity of each wavelength band to the spectral response of the biodegradable mulch film. During the degradation process of the biodegradable mulch film, the breakage of polymer molecular chains causes changes in the vibrational characteristics of chemical bonds. The near-infrared and short-wave infrared bands are more sensitive to changes in functional groups such as carbon-hydrogen bonds, carbonyl groups, and hydroxyl groups than the visible light band. Therefore, the weighting coefficients are set as follows: , , .

[0043] Traverse all At each pixel location, neighborhood window extraction and comprehensive local spectral variation coefficient calculation are repeatedly performed to obtain an image of size [size missing]. Comprehensive local spectral variation coefficient distribution map . The middle is located in the first Line 1 Column element values Equal to pixels The comprehensive local spectral variation coefficient calculated centered on .

[0044] Step 2.3 introduces a preset variation threshold determined based on historical field multispectral data. This threshold represents the statistical boundary distinguishing between mulch film degradation characteristics and background noise. Specifically, it involves retrieving multispectral image data from the historical field multispectral data archive, collected over 45 sampling dates across the past three planting seasons, covering the entire lifecycle of the mulch film from initial installation to complete degradation. Five standard quadrats (0.5m x 0.5m each) are selected for each sampling date, totaling 225 standard quadrats. Two trained agronomic professionals independently visually interpret all 225 standard quadrats, labeling the pixel areas within each quadrat as areas with significant mulch film degradation characteristics or background noise areas. Areas with significant mulch film degradation characteristics are defined as areas where visually identifiable color changes, cracks, holes, or fragmentation occur on the mulch surface; background noise areas are defined as bare soil surfaces, crop plants, crop residues, irrigation water marks, or areas obscured by cloud cover. When the judgments of the two labelers differ, a third senior agronomic expert arbitrates the decision.

[0045] For each standard quadrat, calculate the combined local spectral variation coefficient for each pixel according to the methods described in steps 2.1 and 2.2. The set of comprehensive local spectral variation coefficient values ​​of all pixels in the areas with significant characteristics of plastic film degradation were statistically analyzed. The set of combined local spectral variation coefficient values ​​of all pixels within the background noise region. ,calculate The statistical distribution characteristics, including the mean of the degradation characteristic class. and standard deviation Calculate the same The statistical distribution characteristics, including the mean of the background noise class. and standard deviation Based on the statistical distribution characteristics of the two types of samples, the maximum inter-class variance criterion is used to determine the preset variation threshold. Define the between-class variance function. ,in Candidate mutation threshold variable: ; in, To achieve a comprehensive local spectral variation coefficient greater than The proportion of pixels to all pixels, The overall local spectral variation coefficient is less than or equal to The proportion of pixels to all pixels, and respectively with The mean of the combined local spectral variation coefficients of the two classes of pixels, divided by a boundary. Traversal For all candidate values ​​from 0.01 to 1.00 with a step size of 0.005, calculate the inter-class variance for each candidate value, and select the one that makes the variance between classes equal. Reaching the maximum value As a preset mutation threshold .

[0046] After the above historical data statistical optimization process, the preset variation threshold was finally determined to be: This value characterizes the optimal statistical boundary between the mulch film degradation characteristics and background noise in the local spectral variation coefficient space. When the comprehensive local spectral variation coefficient of a pixel neighborhood is greater than 0.235, the pixel neighborhood is determined to be a region with significant mulch film degradation characteristics; when the comprehensive local spectral variation coefficient is less than or equal to 0.235, the pixel neighborhood is determined to be a background noise region. A preset variation threshold is used. This will serve as a unified criterion for spatial region division, and will be applied to the generated comprehensive local spectral variation coefficient distribution map. Each pixel in the image is compared with a threshold one by one to achieve adaptive determination of region type at the pixel level across the entire image.

[0047] In this embodiment of the invention, a preset-size neighborhood window of data is extracted with the current pixel as the center. The ratio of the standard deviation to the mean of the reflectance data of the three bands is calculated to determine the comprehensive local spectral variation coefficient. A preset variation threshold determined based on historical field multispectral data is introduced as a statistical boundary. Therefore, this overcomes the technical problems of single pixels being easily affected by random noise in complex field environments, resulting in spectral feature distortion, as well as the confusion between the degradation characteristics of plastic film and the complex background spectrum and the lack of objective quantitative distinction standards. Thus, it effectively smooths local random noise and accurately quantifies the degree of neighborhood spectral fluctuation.

[0048] In a preferred embodiment of the present invention, step 3 above may include: Step 3.1 involves extracting the multispectral band observation matrix of regions identified as having significant mulch film degradation characteristics. This matrix serves as input data for the high-dimensional feature preservation mapping. Specifically, it includes comparing the generated comprehensive local spectral variation coefficient distribution map with a preset variation threshold pixel-by-pixel to determine the region type of all pixels in the image. The preset variation threshold is determined by statistical optimization using the maximum inter-class variance criterion from historical field multispectral data. The statistical sample covers manually labeled data from 225 standard quadrats across three complete planting seasons. The final value of the preset variation threshold is 0.235, which serves as the optimal statistical boundary for distinguishing mulch film degradation characteristics from background noise. For any pixel at any spatial coordinate in the original image, the comprehensive local spectral variation coefficient at that location is read and compared with 0.235. If the comprehensive local spectral variation coefficient is greater than 0.235, the pixel is marked as a pixel in a region with significant mulch film degradation characteristics and assigned a corresponding region type identifier; if the comprehensive local spectral variation coefficient is less than or equal to 0.235, it is marked as a pixel in a background noise region.

[0049] After traversing all pixels to complete the region determination, the total number of pixels marked as regions with significant characteristics of mulch film degradation is counted, and the original row and column coordinates corresponding to each significant region pixel are recorded. Based on the mapping relationship between pixel spatial coordinates and row indices of the multispectral band observation matrix, the row index corresponding to each significant region pixel in the multispectral band observation matrix is ​​calculated using the following formula: ; In the formula, For the first The row index corresponding to each pixel with significant degradation features Let be the row coordinate of that pixel. The column coordinates of that pixel. This represents the total number of columns in the original multispectral image. The value ranges from 1 to the total number of pixels with significant degradation features.

[0050] Based on the calculated row indices, complete data for each corresponding row is extracted sequentially from the original multispectral band observation matrix. Each row contains the reflectance values ​​of the pixel in the visible, near-infrared, and shortwave infrared bands. All extracted row data are stacked sequentially according to pixel order to construct a multispectral band observation matrix specific to areas with significant mulch film degradation characteristics. The number of rows in this matrix equals the total number of pixels with significant degradation characteristics, and the number of columns is fixed at 3, corresponding to the three spectral bands. Simultaneously, a coordinate index vector is constructed to match this observation matrix. Each position in the vector stores the original spatial coordinates of the pixel to which the corresponding row data belongs, ensuring that the spatial location information is completely preserved during subsequent feature processing. The resulting multispectral band observation matrix for areas with significant mulch film degradation characteristics serves as the input data for the high-dimensional feature preservation mapping.

[0051] Step 3.2: Perform feature mapping processing based on orthogonal decomposition family on the input data to decompose the multispectral band observation matrix into multiple orthogonal feature components. The eigenvalues ​​represent the retention intensity of each orthogonal feature component for the spectral differences in mulch film degradation. Their values ​​are constrained by the covariance structure of the multispectral band observation matrix. Specifically, this includes: performing data centralization processing on the input multispectral band observation matrix of saliency regions of mulch film degradation to eliminate the interference of differences in the mean reflectance of each band on the decomposition results; calculating the mean reflectance of each band corresponding to each column of the observation matrix; and subtracting the mean of the corresponding column from each element in the matrix to obtain the centered observation matrix.

[0052] The covariance matrix is ​​calculated based on the centered observation matrix. The covariance matrix is ​​used to quantify the linear correlation between reflectance data from different wavebands. The calculation formula is as follows: ; In the formula, It is a 3x3 covariance matrix. The total number of pixels with significant degradation features. This is the multispectral band observation matrix of the centered degradation feature region, with superscript indicating the region. This represents the matrix transpose operation.

[0053] Perform orthogonal eigenvalue decomposition on the covariance matrix, decomposing it into a combination of eigenvalues ​​and eigenvectors. The decomposition formula is as follows: ; In the formula, It is a three-row, three-column orthogonal eigenvector matrix, where each column corresponds to the direction vector of an orthogonal eigencomponent, and the vectors in each column are pairwise orthogonal. It is a diagonal matrix, and the elements on the diagonal are the eigenvalues ​​of the corresponding orthogonal eigencomponents. The magnitude of the eigenvalues ​​represents the retention strength of the corresponding orthogonal eigencomponents for the spectral differences in mulch film degradation. Their values ​​are completely determined by the covariance structure of the multispectral band observation matrix and are constrained by the band correlation and dispersion of the data. After orthogonal decomposition, three mutually orthogonal eigencomponents are obtained. Each eigencomponent corresponds to an eigenvalue and an eigenvector. All eigencomponents together constitute the orthogonal basis space of the spectral data of the degradation characteristic region.

[0054] Step 3.3: Sort multiple orthogonal feature components according to the magnitude of their eigenvalues, and select orthogonal feature components with eigenvalues ​​greater than a preset contribution threshold as the main decomposition feature components. Specifically, this includes: arranging all decomposed eigenvalues ​​in descending order from largest to smallest, simultaneously adjusting the order of the corresponding feature vectors to ensure a one-to-one correspondence between the feature vectors and the sorted eigenvalues; and calculating the variance contribution rate of each orthogonal feature component. The variance contribution rate represents the proportion of information carried by that feature component to the total information, and the calculation formula is: ; In the formula, For the sorted number The variance contribution rate of each feature component For the sorted number 1 eigenvalue, It is the sum of all eigenvalues.

[0055] A preset contribution threshold was introduced as the screening criterion. This threshold was optimized based on the accuracy requirements of identifying mulch film degradation features and the band dimensionality characteristics of multispectral data. After repeated testing and verification with multiple sets of field samples, the specific value of the preset contribution threshold was finally set to 0.90. This threshold achieves an optimal balance between feature dimension compression and degradation information retention, ensuring that the screened principal components can cover the vast majority of effective spectral information related to mulch film degradation, while eliminating low-contribution random noise components. The variance contribution rate of each feature component was accumulated sequentially according to the descending order of feature values. The accumulated variance contribution rate was compared with the preset contribution threshold of 0.90, and all orthogonal feature components whose cumulative contribution rate first reached and exceeded 0.90 were selected as the main degradation feature components to be retained. In the three-band multispectral data scenario of this embodiment, three orthogonal feature components were obtained after orthogonal decomposition. After the corresponding feature values ​​were arranged in descending order, the calculated variance contribution rates of the three components were approximately 0.72, 0.22, and 0.06, respectively. After accumulating sequentially, the cumulative contribution rate of the first feature component is 0.72, which is less than 0.90; the cumulative contribution rates of the first two feature components are the sum of 0.72 and 0.22, which is 0.94, which is greater than 0.90. Therefore, the two orthogonal feature components with the highest ranking are selected and retained as the main degradation feature components, and the third low contribution component is removed as a redundant noise component. The orthogonal feature components that are finally selected and retained are the main degradation feature components.

[0056] Step 3.4 involves recombining the selected principal degradation feature components and their corresponding original pixel spatial coordinates into a set of principal degradation feature vectors. Specifically, this includes projecting the multispectral band observation matrix of the centered degradation feature salient region onto the feature vector space corresponding to the selected principal degradation feature components, thus obtaining the dimensionality-reduced feature vector for each pixel. The projection calculation formula is as follows: ; In the formula, This is the principal feature matrix of the region with significant degradation features, with each row corresponding to the principal degradation feature vector of a pixel. This is the multispectral band observation matrix for the centered region with significant degradation characteristics; It is a matrix composed of column-wise eigenvectors corresponding to the main degradation feature components retained by the screening.

[0057] Each row of the feature vector in the main feature matrix is ​​bound to the original spatial coordinates of the corresponding pixel, and each feature vector includes the row and column coordinates of its corresponding pixel. All feature vectors are reorganized according to the spatial arrangement logic of the original pixels to form a main degradation feature vector set containing spatial location identifiers and high-dimensional degradation feature information. Each element in the main degradation feature vector set retains the spatial coordinate mapping relationship of the original pixels and carries the core degradation spectral features purified by orthogonal decomposition.

[0058] In this embodiment of the invention, by extracting the multispectral band observation matrix of the region with significant mulch film degradation characteristics, performing feature mapping processing based on orthogonal decomposition families to decompose orthogonal feature components constrained by covariance structure, and selecting the main degradation feature components with a contribution value greater than a preset threshold according to the feature value, and finally recombining them with the original pixel spatial coordinates, the technical means overcome the technical problems of high feature redundancy, large computational complexity, and easy loss of key degradation spectral difference information or destruction of the original spatial position correspondence when directly processing high-dimensional multispectral data. Thus, it achieves the goal of effectively removing redundant interference information while maximizing the preservation of the core high-dimensional spectral difference features of mulch film degradation and completely maintaining the mapping relationship between features and the original pixel spatial coordinates.

[0059] In a preferred embodiment of the present invention, step 4 above may include: Step 4.1 involves extracting the multispectral band observation matrix of regions identified as background noise. This matrix serves as the input data for low-dimensional smoothing mapping. Specifically, it includes extracting the corresponding data for all pixels marked as background noise regions based on the full-image pixel region classification results obtained from the local spectral variation coefficient threshold. The preset variation threshold is determined by statistical optimization using the maximum inter-class variance criterion from historical field multispectral data. The statistical sample covers manually labeled data from 225 standard quadrats across three complete planting seasons. The final preset variation threshold value is determined to be 0.235, which is the optimal statistical boundary for distinguishing between mulch film degradation characteristics and background noise. The determination rule is as follows: pixels with a comprehensive local spectral variation coefficient less than or equal to 0.235 are classified as background noise regions and assigned a corresponding region type label; pixels with a comprehensive local spectral variation coefficient greater than 0.235 are classified as regions with significant mulch film degradation characteristics. The total number of pixels in all background noise regions is counted. The original row and column coordinates of each pixel are recorded one by one. Based on the fixed mapping relationship between pixel spatial coordinates and row indices of the original multispectral band observation matrix, the row index of each background noise pixel in the original multispectral band observation matrix is ​​calculated. The calculation formula is as follows: ; In the formula, For the first The row index corresponding to each background noise pixel. This represents the original row coordinates of the pixel. These are the original column coordinates of the pixel. This represents the total number of columns in the original multispectral image. The value ranges from 1 to the total number of pixels in the background noise region.

[0060] The system iterates through the calculated row indices, extracting complete reflectance data for each row from the original multispectral band observation matrix. Each row contains the reflectance values ​​for the pixel in the visible, near-infrared, and short-wave infrared bands. All extracted rows are stacked sequentially in pixel spatial order to construct a multispectral band observation matrix specific to the background noise region. The number of rows in this matrix equals the total number of pixels in the background noise region, and the number of columns is fixed at 3, corresponding to the visible, near-infrared, and short-wave infrared spectral bands respectively.

[0061] A coordinate index vector is constructed simultaneously to match the observation matrix. Each position in the vector stores the original spatial coordinates of the pixel to which the corresponding row of data belongs, ensuring that the spatial position mapping relationship is preserved throughout the subsequent smooth dimensionality reduction process. The resulting multispectral band observation matrix of the background noise region is the input data for the low-dimensional smooth dimensionality reduction mapping.

[0062] Step 4.2: Perform smoothing and dimensionality reduction processing on the input data using a statistical inference family, and calculate the statistical distribution characteristics of pixel reflectance data within the neighborhood. The smoothing weight represents the suppression strength of background noise, and its value is constrained by the variance of the reflectance data within the neighborhood window. Specifically, for each pixel in the background noise region, using its original spatial coordinates as the center, and employing a neighborhood window of the same preset size as when calculating the local spectral variation coefficient, extract the multispectral reflectance data corresponding to all valid pixels within the window. Pixels outside the effective boundaries of the image within the neighborhood window are not included in the calculation; only valid pixels falling within the image are retained for statistical calculation. For all valid pixels within the neighborhood window, calculate the neighborhood mean and neighborhood variance of the reflectance in the visible light, near-infrared, and short-wave infrared bands respectively, using these as the statistical distribution characteristics of the pixel's neighborhood. Taking the visible light band as an example, the neighborhood mean is the arithmetic mean of the visible light reflectance of all valid pixels within the neighborhood, calculated using the following formula: ; In the formula, The mean reflectance of the neighborhood in the visible light band. This represents the total number of valid pixels within the neighborhood window. For the neighboring region The visible light reflectance value of each pixel.

[0063] The neighborhood variance in the visible light band is used to quantify the degree of discrete fluctuation in spectral data within the neighborhood. The calculation formula is as follows: ; In the formula, Let V be the neighborhood reflectance variance in the visible light band.

[0064] Following the exact same calculation logic, the neighborhood mean of the near-infrared band was calculated separately. variance of neighborhood and the neighborhood mean of the shortwave infrared band. variance of neighborhood .

[0065] The comprehensive smoothing weight for the current pixel is calculated based on the neighborhood variance of the three bands. This smoothing weight characterizes the strength of background noise suppression, and its value is constrained by the variance of the reflectance data within the neighborhood window. A larger reflectance variance indicates more severe local spectral fluctuations and more significant noise interference, resulting in a higher smoothing weight and stronger noise suppression. The formula for calculating the comprehensive smoothing weight is as follows: ; In the formula, The overall smoothing weight corresponding to the current pixel. , , The variance contribution coefficients for the three bands are given, and their sum is 1. These coefficients are preset based on the background noise statistical characteristics of each band. The calculated comprehensive smoothing weights are normalized and their values ​​are constrained to the range of 0 to 1.

[0066] Step 4.3: Based on statistical distribution characteristics and smoothing weights, the multispectral band observation matrix of the background noise region is weighted and smoothed to eliminate high-frequency spectral fluctuations and obtain low-dimensional smoothed feature components. Specifically, this includes: combining neighborhood statistical distribution characteristics and comprehensive smoothing weights to perform adaptive weighted smoothing on the original multispectral reflectance data of the center pixel, filtering out high-frequency spectral fluctuations in the background region in the spatial dimension. For each spectral band, the smoothed reflectance value is obtained by weighted fusion of the original reflectance value of the center pixel and the mean reflectance of the neighborhood according to the comprehensive smoothing weights. Taking the visible light band as an example, the smoothing calculation formula is: ; In the formula, This is the smoothed reflectance value for the visible light band. This represents the original visible light reflectance value of the center pixel. This represents the average reflectance in the visible light band within the neighborhood of that pixel. For comprehensive smoothing weights.

[0067] Using the same weighted smoothing calculation method, the smoothed reflectance values ​​for the near-infrared band and the short-wave infrared band were obtained respectively. , This completes the noise suppression processing in the spatial domain.

[0068] After smoothing the spatial dimension, statistical dimensionality reduction aggregation is performed on the smoothed reflectance of the three bands, mapping the three-dimensional band data into single-dimensional low-dimensional smoothed feature components. The background noise region exhibits low spectral variation and strong inter-band correlation; therefore, compressing the feature dimension using weighted average aggregation reduces redundancy in subsequent fusion calculations. The band aggregation weights are determined by the background signal-to-noise ratio (SNR) of each band; bands with higher SNR receive greater weights. The calculation formula for dimensionality reduction aggregation is as follows: ; In the formula, For single-dimensional, low-dimensional, smooth feature components, , , The three weights are used for band aggregation, and their sum is 1. The weight values ​​are pre-optimized and determined based on the statistical characteristics of historical field background spectral data. After two steps of spatial smoothing and band dimensionality reduction, each pixel in the background noise region corresponds to a low-dimensional smooth feature component.

[0069] Step 4.4 involves dimensionality reduction and recombination of the low-dimensional smoothed feature components and their corresponding original pixel spatial coordinates to obtain a background suppression feature vector set. Specifically, this includes binding each low-dimensional smoothed feature component corresponding to a pixel in the background noise region to its original spatial coordinates. Each feature component includes the row and column coordinates of its corresponding pixel, preserving the spatial mapping throughout the process. Following the pixel spatial arrangement logic of the original image, all low-dimensional smoothed feature components bound with coordinate information are recombined in an ordered manner to form the background suppression feature vector set. The order of elements in the vector set strictly corresponds to the pixel spatial order of the original image, and each element contains both the smoothed and dimensionality-reduced background suppression feature value and its corresponding original pixel spatial coordinate index. The background suppression feature vector set completely preserves the spatial mapping of the original pixels.

[0070] In this embodiment of the invention, by extracting the multispectral band observation matrix of the region identified as background noise, performing smoothing dimensionality reduction processing of the statistical inference family to calculate the statistical distribution characteristics of the pixel reflectance data in the neighborhood, and performing weighted smoothing aggregation according to the variance-constrained smoothing weight to eliminate high-frequency spectral fluctuations, and finally performing dimensionality reduction and recombination of the low-dimensional smoothed feature components and the corresponding original pixel spatial coordinates, the technical means overcome the technical problems of high-frequency spectral fluctuation interference caused by complex field backgrounds, the easy blurring of mulch film edge features or loss of spatial location information caused by traditional background processing, and the negative impact of background noise on the calculation of the overall degradation index. Thus, it effectively suppresses and smooths high-frequency random noise in the background region, and completely preserves the spatial coordinate mapping relationship of the background region while reducing the data dimension.

[0071] In a preferred embodiment of the present invention, step 5 above may include: Step 5.1: Based on the pixel location identifiers carried in the main degradation feature vector set and the background suppression feature vector set, extract the corresponding original pixel spatial coordinate indices to obtain a spatial coordinate mapping set. Specifically, each feature element in the main degradation feature vector set and the background suppression feature vector set is accompanied by a corresponding pixel location identifier, which records the original row and column coordinates of the pixel to which the feature element belongs. Iterate through all elements in both feature vector sets, extracting the original pixel spatial coordinate index carried by each element one by one, and simultaneously recording the feature vector data and region type attribute corresponding to each coordinate index. Summarize all extracted coordinate indices and associated attributes to form a spatial coordinate mapping set covering all pixels in the image. Each entry in the spatial coordinate mapping set contains four associated information items: pixel row coordinates, pixel column coordinates, corresponding feature vector, and region type, establishing a one-to-one correspondence between spatial location and feature data, and completely preserving the original spatial attribution of the two types of features.

[0072] Step 5.2: Based on the spatial coordinate mapping set, perform grid-based position rearrangement on the feature elements of the main degradation feature vector set and the background suppression feature vector set to eliminate spatial discrepancies caused by region division, resulting in a coordinate-aligned feature matrix. Specifically, this involves: initializing a two-dimensional feature grid container with the same row and column dimensions as the original multispectral image, based on the index relationship of the spatial coordinate mapping set. Each node in the grid corresponds to an original pixel spatial coordinate, and the total number of rows and columns of the grid remains consistent with the original multispectral image. Traverse all entries in the spatial coordinate mapping set, and fill the corresponding feature vectors into the corresponding node positions of the feature grid according to the row and column coordinates recorded in the entries. Main degradation feature vectors belonging to areas with significant mulch film degradation features are filled into the corresponding coordinate grid nodes, and background suppression feature vectors belonging to background noise areas are filled into the corresponding coordinate grid nodes. After all pixel coordinate grid nodes are filled, a spatially continuous coordinate-aligned feature matrix is ​​formed. The spatial arrangement of the coordinate-aligned feature matrix is ​​strictly aligned with the original multispectral image, with each pixel's feature data corresponding one-to-one with its original coordinates. This eliminates the spatial discrepancy caused by previous independent processing of different regions, enabling the features of the degradation region and the background region to be aligned within a unified spatial grid. Each pixel in the coordinate-aligned feature matrix stores the feature vector obtained from the corresponding region's processing. Degradation region nodes store high-dimensional principal degradation feature vectors, while background region nodes store low-dimensional background suppression feature vectors.

[0073] Step 5.3: Based on the coordinate-aligned feature matrix, perform feature space recombination processing based on tensor algebra families. This involves upscaling the coordinate-aligned feature matrix along the channel dimension to obtain an initial fused feature tensor. The recombination weights characterize the aggregation strength of spatial continuity, and their values ​​are constrained by the original number of channels in the multispectral band observation matrix. Specifically, this includes: performing feature space recombination processing based on tensor algebra families on the coordinate-aligned feature matrix; performing upscaling mapping on the feature matrix along the channel dimension; mapping regional features of different dimensions to the same feature channel space to form an initial fused feature tensor with a three-dimensional structure; and determining the recombination weights for the upscaling mapping. The recombination weights characterize the aggregation strength of spatial continuity, and their values ​​are constrained by the original number of channels in the multispectral band observation matrix. The original channels of the multispectral band observation matrix are visible light, near-infrared, and short-wave infrared. The dimension of the recombination weights matches the original number of channels, ensuring that the upscaled feature space has a physical correspondence with the original spectral space. The number of rows in the recombined weight matrix is ​​consistent with the input feature dimension, and the number of columns is consistent with the set number of unified feature channels. The number of channels is set with reference to the number of original bands to ensure the physical interpretability of the feature space. For a feature vector at any pixel position in the coordinate-aligned feature matrix, the formula for its dimension-up mapping is: ; In the formula, This is the upgraded feature vector, with the dimension consistent with the unified number of channels; This is the original feature vector at the current pixel location. The original feature dimensions of the degradation region and the background region differ. To reorganize the weight matrix, the matrix elements are pre-optimized and determined based on the spectral response characteristics of the original bands, and its dimensionality is constrained by the number of original channels.

[0074] After performing the above-described dimension-up mapping operation point-by-point on all pixel positions in the coordinate-aligned feature matrix, the original two-dimensional feature matrix is ​​expanded into a three-dimensional tensor structure along the channel dimension, i.e., the initial fused feature tensor. The dimension of the initial fused feature tensor is the number of image rows multiplied by the number of image columns multiplied by the unified number of channels, thus unifying the dimension of the feature vectors at all pixel positions.

[0075] Step 5.4: Based on the initial fused feature tensor, perform topological interpolation smoothing aggregation based on manifold embedding families on the feature vectors at the boundaries of different regions within the tensor to eliminate spectral faults caused by feature-level splicing, resulting in a regionally differentiated fused feature tensor. Specifically, this includes: based on the initial fused feature tensor, identifying and extracting feature vectors at the boundaries between regions with significant mulch film degradation characteristics and background noise regions within the tensor, obtaining a set of boundary feature vectors. The boundary region is determined by traversing the neighborhood of all pixels. If the region type of a pixel is inconsistent with the region type of its neighboring pixels, then the pixel is determined to be a region boundary pixel, and its corresponding feature vector is included in the boundary feature vector set.

[0076] A dimensionality reduction mapping based on manifold embedding families is performed on the boundary feature vector set, mapping the high-dimensional boundary feature vectors to a low-dimensional manifold space, resulting in a set of coordinate points in the manifold space. The manifold embedding process preserves the inherent topological structure of the spectral features of the boundary region. The manifold curvature characterizes the degree of continuous change in the spectral features of the boundary region, and its value is constrained by the local gradient of the initial fused feature tensor. The larger the local gradient of the boundary region, the higher the manifold curvature and the more drastic the feature changes. The dimensionality reduction mapping follows the local neighborhood preservation criterion of the manifold space, ensuring that adjacent feature vectors in the high-dimensional space maintain their proximity in the low-dimensional manifold space. Based on the set of coordinate points in the manifold space, the geodesic distance between adjacent feature points is calculated. The geodesic distance is calculated along the manifold surface, which can more accurately reflect the true distance of spectral feature changes. Using the geodesic distance as the interpolation weight, topological interpolation is performed on the feature vectors between adjacent regions, generating transitional feature vectors between the features of the degraded region and the background region, filling in the spectral discontinuities caused by feature-level splicing. The formula for calculating the topological interpolation is: ; In the formula, The transition feature vector obtained by interpolation; and These are the two original feature vectors on either side of the boundary; For interpolation position to Geodesic distance between corresponding manifold coordinate points; For interpolation position to The geodesic distances to the corresponding manifold coordinate points are calculated. The smooth transition feature vectors obtained from interpolation are back-mapped back to the original high-dimensional feature space, replacing the original feature vectors at the corresponding boundary positions in the initial fusion feature tensor. Then, tensor recombination is performed with the non-boundary region features in the initial fusion feature tensor to eliminate the spatial discrepancy bias and spectral abrupt changes caused by region division, ultimately obtaining the regionally differentiated fusion feature tensor.

[0077] In this embodiment of the invention, a spatial coordinate mapping set is extracted based on pixel location identifiers. The feature elements are then reordered by gridding to eliminate spatial discrepancies. Subsequently, a feature space reorganization process based on tensor algebra families is performed to perform channel dimension upscaling. Furthermore, a topological interpolation smoothing aggregation technique based on manifold embedding families is applied to the feature vectors at the boundaries of different regions. This overcomes the technical problems of feature spatial location discrepancies caused by differential independent processing of the mulch film degradation region and the background region, dimension mismatch during multi-source feature fusion, and feature abrupt changes and spectral breaks at the boundaries of different regions due to differences in processing logic. As a result, precise gridding alignment of the main degradation features and background suppression features is achieved in the original pixel space. The spectral breaks at the boundaries of regions are effectively eliminated through topological interpolation smoothing aggregation, ensuring the spatial continuity and spectral consistency of feature fusion. This results in the output of a high-quality, seamlessly connected regionally differentiated fusion feature tensor.

[0078] In a preferred embodiment of the present invention, step 5.4 above may include: Step 5.41: Based on the initial fused feature tensor, extract feature vectors at the boundaries of different regions within the tensor to obtain a set of boundary feature vectors. Specifically, this includes: obtaining the region type identifier corresponding to each pixel position in the initial fused feature tensor. The region type identifiers are divided into two categories: regions with significant mulch film degradation characteristics and background noise regions, maintaining a one-to-one correspondence with the previous region segmentation results. Based on the pixel grid of the initial fused feature tensor, traverse all spatial coordinate positions pixel by pixel, and use an eight-neighbor detection method to determine whether each pixel belongs to a region boundary position. That is, check the region type identifiers of the eight neighboring pixels around the current pixel. If the region type of the current pixel is different from the region type of any neighboring pixel, then the pixel is determined to be a region boundary pixel. For all pixels determined to be region boundaries, extract the corresponding high-dimensional feature vector from the initial fused feature tensor, and simultaneously record the pixel spatial coordinates corresponding to each feature vector. Summarize all extracted boundary position feature vectors and their corresponding coordinates to form a set of boundary feature vectors.

[0079] Step 5.42: Based on the set of boundary feature vectors, perform a dimensionality reduction mapping process based on manifold embedding families to map the high-dimensional boundary feature vectors to the low-dimensional manifold space, obtaining a set of coordinate points in the manifold space. The manifold curvature characterizes the degree of continuous change in the spectral features of the boundary region, and its value is constrained by the local gradient of the initial fused feature tensor. Specifically, this includes: using the set of boundary feature vectors as input, performing a dimensionality reduction mapping process based on manifold embedding families to map the high-dimensional feature vectors to the low-dimensional manifold space, obtaining the corresponding set of coordinate points in the manifold space. The dimensionality reduction mapping uses a locally linear embedding method, preserving the local neighborhood topology of the boundary region features during the dimensionality reduction process, ensuring that pixels with similar spectral features in the high-dimensional space maintain their proximity in the low-dimensional manifold space. Before the dimensionality reduction mapping, the local gradient of the boundary region in the initial fused feature tensor is calculated. The local gradient characterizes the degree of change in the feature vectors on both sides of the boundary region. For any two adjacent boundary pixels, the formula for calculating the local gradient is: ; In the formula, This represents the local gradient at that location. and These are the feature vectors of two adjacent boundary pixels. The L2 norm of a vector. and These represent the spatial coordinates of two pixels. The manifold curvature characterizes the degree of continuous change in the spectral features of the boundary region, and its value is constrained by the local gradient of the initial fused feature tensor. A larger local gradient indicates a more significant difference in features between the two sides of the boundary region, resulting in a higher manifold curvature and a greater degree of bending in the feature changes. The constraint relationship between manifold curvature and local gradient is as follows: ; In the formula, For manifold curvature, The curvature scaling factor is preset based on the characteristic variation of historical field multispectral data. After completing the dimensionality reduction mapping of all boundary feature vectors under the constraint of manifold curvature, each high-dimensional feature vector corresponds to a coordinate point in the low-dimensional manifold space. All coordinate points are summarized to form a set of coordinate points in the manifold space, and the topological structure of the point set is consistent with the intrinsic structure of the original high-dimensional features.

[0080] Step 5.43: Based on the manifold space coordinate point set, calculate the geodesic distance between adjacent feature points, and perform topological interpolation calculation on the feature vectors between adjacent regions based on the geodesic distance to fill in the spectral discontinuities generated by feature-level stitching, obtaining an interpolated smooth feature vector set. Specifically, this includes: calculating the geodesic distance between adjacent feature points based on the manifold space coordinate point set. The geodesic distance is the shortest path distance along the low-dimensional manifold surface, which can accurately reflect the true degree of difference between spectral features and avoid the estimation bias of Euclidean distance in nonlinear manifold space. The geodesic distance is calculated by constructing a manifold adjacency graph and using the shortest path algorithm to traverse the edge weights of the adjacency graph. The edge weights are determined by the Euclidean distance between adjacent manifold coordinate points. For transition pixels in the boundary region, perform topological interpolation calculation based on the geodesic distance to generate a smooth transition feature vector between the degradation region features and the background region features, filling in the spectral discontinuities generated by feature-level stitching. The interpolation uses a geodesic inverse distance weighted method. For any location to be interpolated on the transition zone, the two manifold coordinate points closest to each other on either side are selected as the interpolation reference. The interpolation calculation formula is as follows: ; In the formula, This is the low-dimensional transition feature vector obtained through interpolation. and These are the low-dimensional feature vectors corresponding to the two interpolation reference points, respectively. The distance is the geodesic distance from the location to be interpolated to the first reference point. The geodesic distance from the interpolation location to the second reference point is used to perform the above interpolation calculation on all interpolation locations in the boundary transition zone one by one, generating smooth feature vectors for all transition locations. After summarizing, an interpolation smooth feature vector set is obtained, and each element in the vector set is accompanied by the corresponding pixel spatial coordinate information.

[0081] Step 5.44: Based on the interpolated smooth feature vector set, the interpolation result is back-mapped back to the original high-dimensional feature space, and tensor recombination is performed with the non-boundary region features in the initial fused feature tensor to eliminate spatial discretization bias, resulting in a region-differentiated fused feature tensor. Specifically, this includes: performing a manifold embedding inverse mapping operation on all low-dimensional transition feature vectors in the interpolated smooth feature vector set, mapping the interpolation result in the low-dimensional manifold space back to the original high-dimensional feature space. The inverse mapping process follows the inverse transformation logic of manifold embedding, ensuring that the high-dimensional feature vector after inverse mapping is completely consistent with the dimension and physical meaning of the original feature space, and retaining the smooth transition characteristics obtained from low-dimensional interpolation; replacing the original feature vectors at the corresponding boundary positions in the initial fused feature tensor with the high-dimensional transition feature vectors obtained from inverse mapping according to the corresponding pixel space coordinates. The non-boundary region features in the initial fused feature tensor remain unchanged; only the features in the boundary regions are smoothly replaced. After the replacement is completed, tensor recombination is performed on the feature data of the full tensor in spatial coordinate order. The dimensional arrangement and data storage structure of the tensor are uniformly adjusted to eliminate the spatial discrepancy caused by region division and interpolation. The final regionally differentiated fused feature tensor has the same spatial size and feature dimension as the initial fused feature tensor.

[0082] In this embodiment of the invention, feature vectors at the boundaries of different regions in the initial fused feature tensor are extracted, and a dimensionality reduction mapping process based on manifold embedding family is performed to map them to a low-dimensional manifold space constrained by local gradients. Then, topological interpolation is performed based on the geodesic distance between adjacent feature points to fill spectral faults, and the interpolation results are back-mapped back to the original high-dimensional feature space and tensor recombination with the features of non-boundary regions. Therefore, this method overcomes the technical problems caused by the differential mapping processing of the mulch film degradation area and the background area, such as abrupt feature changes at the boundary, spectral faults caused by direct interpolation in high-dimensional space due to noise interference, and spatial discrepancies caused by simple feature splicing. This method achieves accurate capture of the continuous variation law of spectral features of the boundary region in the low-dimensional manifold space, realizes the natural smooth transition and seamless filling of feature boundaries through topological interpolation, and effectively eliminates the boundary artifacts caused by region division after high-dimensional space recombination, thereby outputting a high-quality regional differential fused feature tensor with spatial continuity and spectral consistency.

[0083] In a preferred embodiment of the present invention, step 6 above may include: Step 6.1: Based on the regional differential fusion feature tensor, extract the multispectral band feature components contained in the regional differential fusion feature tensor to obtain a multispectral feature component set. Specifically, the regional differential fusion feature tensor adopts a three-dimensional structure combining spatial and channel dimensions. The spatial dimension completely corresponds to the row and column dimensions of the original multispectral image, and the channel dimension corresponds to the multispectral band feature components after differential processing and fusion. Each channel forms a one-to-one mapping relationship with the spectral features of the three original bands: visible light, near-infrared, and short-wave infrared. Pixel-by-pixel, traverse all spatial coordinate positions of the regional differential fusion feature tensor. For the pixel in the i-th row and j-th column, extract the feature values ​​of each channel at that position sequentially according to channel order to obtain a set of feature component groups of equal length to the number of channels. This set of feature component groups completely preserves the degradation feature information of the pixel under different spectral dimensions. Simultaneously record the pixel spatial coordinates corresponding to each feature component group to ensure that the mapping relationship between feature data and spatial position is not lost. Summarize the feature component groups corresponding to all pixel positions to form a multispectral feature component set.

[0084] Step 6.2: Based on the multi-band feature component set, perform feature weighted aggregation processing based on a family of variational optimizations to map the multi-band feature component set to a one-dimensional scalar space, obtaining the initial mulch film degradation index. The aggregation weight characterizes the sensitivity of each band to the degradation state, and its value is constrained by the channel variance of the multi-band feature component set. Specifically, this includes: performing feature weighted aggregation using a family of variational optimizations to obtain the initial mulch film degradation index. Based on the multi-band feature component set, first calculate the global feature variance of each feature channel. The channel variance is used to quantify the dispersion of the feature of that band across the entire map. The larger the variance value, the stronger the ability of that band feature to distinguish the mulch film degradation state and the higher its sensitivity to changes in the degree of degradation. The formula for calculating the variance of the c-th feature channel is: ; In the formula, Let be the feature variance of the c-th channel, M be the total number of rows in the image, and N be the total number of columns in the image. Let be the feature value of the pixel in the i-th row and j-th column in the c-th channel. It is the global mean of the feature values ​​of all pixels in the c-th channel.

[0085] The aggregate weights characterize the sensitivity of each band to the degradation state, and their values ​​are constrained by the channel variance of the multi-band feature component set. The corresponding aggregate weights are calculated based on the feature variance of each channel. The weights are positively correlated with the channel variance, and all weights, after normalization, sum to 1. The formula for calculating the aggregate weights is as follows: ; In the formula, is the aggregation weight of the c-th channel, where C is the total number of feature channels.

[0086] After weight optimization based on variational optimization criteria, weighted aggregation is performed on the multi-band feature components of each pixel, mapping the multi-dimensional feature vector to a one-dimensional scalar space to obtain the initial mulch film degradation index for each pixel. The formula for weighted aggregation is as follows: ; In the formula, The initial mulch film degradation index is a single scalar value for the pixel in the i-th row and j-th column. This weighted aggregation process follows the solution logic of a family of variational optimizations. By adaptively matching the contribution weights of each band, it maximizes the index differentiation of different degradation stages and minimizes the random fluctuation of the index within the same degradation stage, so that the generated initial mulch film degradation index can accurately characterize the actual degradation degree of the mulch film.

[0087] Step 6.3: Based on the initial mulch film degradation index, multiple preset threshold intervals for different stages, determined statistically from historical field multispectral data, are introduced. A numerical interval matching operation is then performed to obtain the mulch film degradation index interval matching results. Specifically, this includes: introducing multiple degradation stage threshold intervals determined statistically from historical field multispectral data. Each threshold interval corresponds to a degradation degree level. The upper and lower boundary values ​​of the intervals are determined by the statistical distribution of the index for each degradation stage in historical samples. The statistical samples cover manually labeled data from 225 standard quadrats across three complete planting seasons. The boundary points are optimized using an ordered sample clustering algorithm to maximize the differences in degradation characteristics between classes. By minimizing intra-class exponential random fluctuations, five degradation levels were ultimately identified, with corresponding threshold ranges as follows: initial deployment phase (0-0.18), degradation induction phase (0.18-0.36), rapid degradation phase (0.36-0.57), disintegration phase (0.57-0.79), and complete degradation phase (0.79-1.00). The upper and lower boundary values ​​of all stage threshold ranges were extracted, and duplicate boundary values ​​at the junctions of adjacent stages were deduplicated. The boundary values ​​were then arranged in ascending order to form an ordered set of stage threshold boundaries, clarifying the continuous arrangement of each stage's intervals. The set of stage threshold boundaries contains six boundary values: 0, 0.18, 0.36, 0.57, 0.79, and 1.00. Any two adjacent boundary values ​​in the set correspond exactly to the threshold range of a degradation stage.

[0088] For each pixel, the initial mulch film degradation index is compared sequentially with the boundary values ​​in the stage threshold boundary set to determine the range within which the index falls, thus obtaining a preliminary interval assignment result. The interval assignment uses a left-closed, right-open constraint rule. If the initial mulch film degradation index is greater than or equal to the (k-1)th boundary value and less than the kth boundary value, it is preliminarily determined that the index belongs to the threshold interval corresponding to the kth degradation stage. For the last complete degradation period interval, a bilaterally closed interval rule is used. An index greater than or equal to 0.79 and less than or equal to 1.00 is considered part of the complete degradation period. For example, if a pixel's initial mulch film degradation index is 0.25, and after comparison, it is greater than or equal to 0.18 and less than 0.36, so it is preliminarily determined to belong to the threshold interval corresponding to the degradation induction period; if a pixel's initial mulch film degradation index is 0.62, and after comparison, it is greater than or equal to 0.57 and less than 0.79, so it is preliminarily determined to belong to the threshold interval corresponding to the disintegration and breakage period.

[0089] For the initial mulch film degradation index at the threshold boundary between adjacent stages—that is, samples whose index values ​​fall within a pre-defined transition zone on both sides of the boundary value—a smooth transition mapping process is performed to eliminate the stage jump problem caused by hard threshold determination. The half-width of the transition zone is determined by the statistical fluctuation of the stage boundary in historical data. After statistically analyzing the standard deviation of degradation index fluctuations from multiple repeated observations of the same sample, the half-width of the transition zone is finally set to 0.02. A smooth transition zone is formed by extending 0.02 units to both sides of any stage boundary value as the center. For example, the transition zone range corresponding to a boundary value of 0.36 is 0.34 to 0.38. For indices falling within the transition zone, their membership degree to the two adjacent stages is calculated, and the interval with the higher membership degree is selected as the final assignment interval. The formula for calculating the membership degree within the transition zone is as follows: ; ; In the formula, Let be the membership degree of the index to the k-th stage. Let be the membership degree of the index to the (k+1)th stage. These are the boundary values ​​for the two stages. This represents the half-width value of the transition band. This is the initial mulch film degradation index for a single pixel, which is the target value to be determined for the stage to which it belongs. Taking a boundary value of 0.36 as an example, if the initial mulch film degradation index of a certain pixel is 0.35, it falls within the transition zone corresponding to this boundary. Substituting it into the formula, we can calculate that the membership degree to the degradation induction period is 0.75, and the membership degree to the rapid degradation period is 0.25. Therefore, after correction, the index is determined to belong to the threshold interval corresponding to the degradation induction period.

[0090] After smoothing correction of the boundary area, the initial mulch film degradation index of all pixels corresponds to a unique target threshold range. The mulch film degradation index range matching result is obtained by summarizing the results.

[0091] Step 6.4: Based on the mulch film degradation index interval matching results, the mulch film degradation index falling into a specific interval is mapped to the corresponding state identifier to obtain the corresponding biodegradable mulch film degradation stage label. Specifically, this includes: pre-establishing a one-to-one mapping relationship between stage threshold intervals and biodegradable mulch film degradation stage labels. Each preset threshold interval corresponds to a unique degradation stage identifier. Degradation stages can be divided into different levels according to the degree of degradation from low to high, such as the initial laying stage, degradation induction stage, rapid degradation stage, disintegration and breakage stage, and complete degradation stage. Each level corresponds to different field mulch film morphological characteristics. Based on the mulch film degradation index interval matching results, the target threshold interval matched by each pixel is converted into the corresponding degradation stage label through the mapping relationship. The degradation stage labels of all pixels are arranged in the original spatial coordinate order to form a spatial distribution map of field mulch film degradation stages, intuitively presenting the differences in the degree of degradation in different areas. On this basis, the number of pixels and area ratio corresponding to each degradation stage label within the entire map are counted, and the degradation stage with the highest ratio is selected as the overall degradation stage label corresponding to the entire image. Finally, the biodegradable mulch film degradation stage identification result combining pixel-by-pixel spatial distribution and overall judgment is output.

[0092] In this embodiment of the invention, multispectral band feature components are extracted from the regionally differentiated fusion feature tensor, and feature weighted aggregation processing based on variational optimization family is performed to map them to a one-dimensional scalar space to obtain an initial mulch film degradation index. A preset stage threshold interval determined based on historical data is introduced for numerical interval matching, and finally the index is mapped to the corresponding state identifier. Therefore, this method overcomes the technical problems in traditional mulch film degradation assessment, such as the strong subjectivity of multi-band feature weight allocation, the difficulty in quantifying the true sensitivity of each band to the degradation state, and the susceptibility of single or static threshold judgment to local noise interference, which can lead to abrupt changes or misjudgments in degradation stage identification. This method achieves the scientific quantification of the contribution of each band feature to the degradation state, realizes the accurate dimensionality reduction mapping of high-dimensional complex features to a one-dimensional intuitive degradation index, and achieves high accuracy in the field biodegradable mulch film degradation stage through a data-driven interval matching mechanism.

[0093] In a preferred embodiment of the present invention, step 6.3 above may include: Step 6.31: Based on the initial mulch film degradation index and multiple preset stage threshold intervals, extract the upper and lower boundary values ​​of each stage threshold interval to obtain a set of stage threshold boundaries. Specifically, the preset degradation stage threshold intervals are obtained through statistical optimization based on historical field multispectral data. The statistical samples cover manually labeled data from 225 standard plots across three complete planting seasons, encompassing the entire lifecycle of the mulch film from initial installation to complete degradation. An ordered sample clustering algorithm is used to maximize inter-class degradation feature differences and minimize intra-class index random fluctuations, ultimately determining five degradation levels, each corresponding to a continuous index threshold interval. The lower and upper boundary values ​​of the five degradation stage threshold intervals are extracted sequentially. After deduplication of overlapping boundary values, they are arranged in ascending order from smallest to largest, forming an ordered set of stage threshold boundaries. This set contains six boundary values: 0, 0.18, 0.36, 0.57, 0.79, and 1.00. Any two adjacent boundary values ​​in the set constitute a threshold range for a degradation stage. The specific ranges corresponding to each stage and the degree of degradation are as follows: Stage 1, initial laying period, corresponds to the range of 0 to 0.18; Stage 2, degradation induction period, corresponds to the range of 0.18 to 0.36; Stage 3, rapid degradation period, corresponds to the range of 0.36 to 0.57; Stage 4, disintegration and fragmentation period, corresponds to the range of 0.57 to 0.79; and Stage 5, complete degradation period, corresponds to the range of 0.79 to 1.00. The ordered set of stage threshold boundaries clearly defines the numerical boundaries of each stage.

[0094] Step 6.32: Based on the stage threshold boundary set, perform a numerical comparison operation between the initial mulch film degradation index and the values ​​of each upper and lower boundary to determine the target threshold interval to which the initial mulch film degradation index belongs, and obtain the interval assignment result. Specifically, this includes: using the stage threshold boundary set as a unified comparison benchmark, performing a numerical comparison operation on the initial mulch film degradation index corresponding to each pixel. The comparison process employs a binary search method to improve computational efficiency. The initial mulch film degradation index of a single pixel is compared sequentially with each boundary value within the boundary set to locate the numerical interval of the index. Interval assignment is determined using a left-closed, right-open constraint rule. For an interval formed by the (k-1)th boundary value and the kth boundary value, if the initial mulch film degradation index is greater than or equal to the (k-1)th boundary value and less than the kth boundary value, then the initial mulch film degradation index is determined to belong to the threshold interval corresponding to the kth degradation stage. For the last complete degradation period interval, a bilaterally closed interval rule is used. An index greater than or equal to 0.79 and less than or equal to 1.00 is considered part of the complete degradation period. For example, if the initial mulch film degradation index of a pixel is 0.25, and the comparison shows it is greater than or equal to 0.18 and less than 0.36, it is initially determined to belong to the threshold interval corresponding to the second stage degradation induction period. Similarly, if the initial mulch film degradation index of another pixel is 0.62, and the comparison shows it is greater than or equal to 0.57 and less than 0.79, it is initially determined to belong to the threshold interval corresponding to the fourth stage disintegration and breakage period. After comparing and determining the initial mulch film degradation index of all pixels in the entire image one by one, the preliminary interval assignment conclusions corresponding to each index are obtained, and the results are summarized to form the interval assignment determination results.

[0095] Step 6.33: Based on the interval attribution determination results, perform a smooth transition mapping process based on fuzzy membership families on the initial mulch film degradation index located at the boundary of adjacent stage threshold intervals to eliminate identification abrupt changes caused by stage jumps, and obtain the corrected interval attribution determination results. Specifically, this includes: identifying all initial mulch film degradation indices falling at the boundary of adjacent stage threshold intervals based on the interval attribution determination results. A symmetrical numerical transition zone is set for each stage boundary value. The half-width of the transition zone is determined by statistically analyzing the random fluctuation amplitude of the historical field degradation index. After statistically analyzing the standard deviation of the degradation index fluctuation in multiple repeated observations of the same square, the half-width of the transition zone is finally set to 0.02. The range of the transition zone corresponding to each boundary value is from the boundary value minus 0.02 to the boundary value plus 0.02. For example, the transition zone range corresponding to the boundary value of 0.18 is 0.16 to 0.20, and the transition zone range corresponding to the boundary value of 0.36 is 0.34 to 0.38. The transition zones for other boundary values ​​are determined according to the same rules.

[0096] For an initial mulch film degradation index falling within any transition zone, a smooth transition mapping process based on fuzzy membership families is performed. The fuzzy membership degree of the index with respect to two adjacent stages is calculated, and the membership degree value is negatively correlated with the distance from the index to the central region of the corresponding stage. Taking the adjacent k-th stage and k+1-th stage as an example, the boundary value between them is denoted as... The half-width of the transition zone is denoted as For the initial mulch film degradation index falling within this transition zone The formula for calculating its membership degree in the k-th stage is: ; The formula for calculating the membership degree of the (k+1)th stage is: ; In the formula, represents the membership degree of the initial mulch film degradation index to stage k. The initial mulch film degradation index is the membership degree of the (k+1)th stage. The sum of the two membership degree values ​​is always equal to 1. Taking a boundary value of 0.36 as an example, if the initial mulch film degradation index of a certain pixel is 0.35 and falls within the transition zone corresponding to the boundary, substituting it into the formula, the membership degree for the degradation induction period is 0.75, and the membership degree for the rapid degradation period is 0.25. After calculating the membership degrees for the two stages, the stage with the larger membership degree value is selected as the corrected assignment interval for the index; if the two membership degree values ​​are exactly equal, it is preferentially assigned to the stage with a lower degree of degradation. For initial mulch film degradation indices that do not fall within any transition zone, the original interval assignment result remains unchanged. After smoothing and correcting the indices at all boundary positions, the corrected interval assignment result is obtained.

[0097] Step 6.34: Based on the corrected interval assignment results, determine the target threshold interval to which the initial mulch film degradation index belongs, and obtain the mulch film degradation index interval matching result. Specifically, this includes: determining a unique target threshold interval for each initial mulch film degradation index based on the corrected interval assignment results; performing a summary verification on the determination results of all pixels in the entire image to ensure that there are no interval assignment conflicts or boundary omissions for any index; and ensuring that each initial mulch film degradation index is accurately matched to a preset degradation stage threshold interval, ultimately forming a complete mulch film degradation index interval matching result. The result includes the target threshold interval identifier and corresponding degradation stage attribute for each pixel, providing a direct basis for the subsequent mapping output of degradation stage labels.

[0098] In this embodiment of the invention, a set of stage threshold boundaries is constructed by extracting the upper and lower boundary values ​​of the threshold intervals of each stage. The initial mulch film degradation index is compared with the boundary values ​​to determine the interval affiliation. For the index at the junction of adjacent intervals, a smooth transition mapping based on a fuzzy membership family is performed to eliminate stage jumps. Finally, the target threshold interval is determined based on the corrected interval affiliation determination result. Therefore, this technical means overcomes the technical problems of traditional hard threshold interval matching, which is prone to step-like abrupt changes near the critical value, and the technical problems of drastic jumps in the mulch film degradation stage identification results caused by small fluctuations in the field environment, lack of continuity and robustness. Thus, it effectively eliminates the identification abrupt changes caused by stage jumps, realizes smooth transition and flexible matching of degradation stage determination in the critical area, and improves the stability of intelligent identification of biodegradable mulch film degradation stages in complex field environments.

[0099] like Figure 2 As shown, embodiments of the present invention also provide an intelligent identification system for the degradation stage of biodegradable mulch film based on multispectral image fusion, comprising: The acquisition module is used to acquire multispectral images of biodegradable mulch films in the field. The multispectral images contain reflectance data of three bands, visible light, near infrared and shortwave infrared, which are acquired simultaneously. A multispectral band observation matrix is ​​constructed based on the multispectral images. The calculation module is used to calculate the local spectral variation coefficient of each pixel neighborhood in the multispectral image, use the local spectral variation coefficient as the criterion for spatial region division, and introduce a preset variation threshold determined by statistical analysis of historical field multispectral data. The determination module is used to determine that the neighborhood of the current pixel is a region with significant mulch film degradation characteristics when the local spectral coefficient of variation is greater than a preset variation threshold. For the multispectral band observation matrix of the region with significant mulch film degradation characteristics, a high-dimensional feature preservation mapping of the orthogonal decomposition family is performed to map the multispectral band observation matrix to the main degradation feature vector set. When the local spectral coefficient of variation is less than or equal to the preset variation threshold, the neighborhood of the current pixel is determined to be a background noise region. For the multispectral band observation matrix of the background noise region, a low-dimensional smooth dimensionality reduction mapping of the statistical inference family is performed to map the multispectral band observation matrix to the background suppression feature vector set. The fusion module is used to perform feature-level alignment and fusion processing on the main degradation feature vector set and the background suppression feature vector set according to the original pixel space coordinates to obtain the regional differential fusion feature tensor. The processing module is used to calculate the mulch film degradation index based on the regionally differentiated fusion feature tensor, and to match and determine the corresponding biodegradable mulch film degradation stage label based on the mulch film degradation index and multiple preset stage threshold intervals.

[0100] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0101] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A smart identification method for the degradation stage of biodegradable mulch film based on multispectral image fusion, characterized in that, The method includes: Multispectral images of biodegradable mulch film in the field were acquired. The multispectral images contained reflectance data of three bands, visible light, near infrared and shortwave infrared, which were acquired simultaneously. A multispectral band observation matrix was constructed based on the multispectral images. By calculating the local spectral variation coefficient of each pixel neighborhood in the multispectral image, the local spectral variation coefficient is used as the criterion for spatial region division, and a preset variation threshold determined based on historical field multispectral data is introduced. When the local spectral variation coefficient is greater than the preset variation threshold, the neighborhood of the current pixel is determined to be a region with significant mulch film degradation characteristics. For the multispectral band observation matrix of the region with significant mulch film degradation characteristics, a high-dimensional feature preservation mapping of the orthogonal decomposition family is performed to map the multispectral band observation matrix to the main degradation feature vector set. When the local spectral variation coefficient is less than or equal to the preset variation threshold, the neighborhood of the current pixel is determined to be a background noise region. For the multispectral band observation matrix of the background noise region, a low-dimensional smooth dimensionality reduction mapping of the statistical inference family is performed to map the multispectral band observation matrix to the background suppression feature vector set. The main degradation feature vector set and the background suppression feature vector set are aligned and fused at the feature level according to the original pixel spatial coordinates to obtain the regional differential fused feature tensor. The degradation index of the mulch film is calculated based on the regionally differentiated fusion feature tensor. The degradation index of the mulch film is matched and determined with multiple preset stage threshold intervals to obtain the corresponding biodegradable mulch film degradation stage label.

2. The intelligent identification method for the degradation stage of biodegradable mulch film based on multispectral image fusion according to claim 1, characterized in that, Multispectral images of biodegradable mulch films in the field were acquired. These images included reflectance data from simultaneously acquired visible, near-infrared, and short-wave infrared bands. A multispectral band observation matrix was constructed based on these images, including: Multispectral images of biodegradable mulch films in the field were acquired, and raw reflectance data of the visible light band, near-infrared band, and short-wave infrared band were extracted from the multispectral images respectively. Spatial coordinate alignment and radiometric normalization were performed on the raw reflectance data of the three bands to eliminate the dimensional differences and spatial pixel offsets generated when the sensors in different bands were collected, and the aligned multi-band reflectance data were obtained. Traverse the aligned multi-band reflectivity data, and perform vectorization stacking operation on the visible light reflectivity data, near-infrared reflectivity data and short-wave infrared reflectivity data in the same pixel spatial coordinates according to the preset band arrangement order to obtain the result of the vectorization stacking operation. Based on the results of the vectorized stacking operation, a multispectral band observation matrix is ​​constructed with pixel spatial coordinates as row indices and reflectance data of three bands as column elements. The multispectral band observation matrix is ​​then used as input data for subsequent calculation of local spectral variation coefficients.

3. The intelligent identification method for the degradation stage of biodegradable mulch film based on multispectral image fusion according to claim 2, characterized in that, By calculating the local spectral variation coefficient of each pixel neighborhood in a multispectral image, the local spectral variation coefficient is used as the criterion for spatial region division. A preset variation threshold, determined statistically based on historical field multispectral data, is introduced, including: Using the current pixel in the multispectral band observation matrix as the center, extract the multispectral band observation matrix data within a neighborhood window of a preset size; The ratio of the standard deviation to the mean of the reflectance data for the visible light, near-infrared and short-wave infrared bands within the neighborhood window is calculated respectively to obtain the local spectral variation coefficient of each band, and the comprehensive local spectral variation coefficient of the current pixel neighborhood is determined based on the local spectral variation coefficient of each band. A preset variation threshold is introduced, which is determined by statistical analysis of historical field multispectral data. The preset variation threshold characterizes the statistical boundary that distinguishes the degradation characteristics of plastic film from background noise.

4. The intelligent identification method for the degradation stage of biodegradable mulch film based on multispectral image fusion according to claim 3, characterized in that, When the local spectral variation coefficient exceeds a preset variation threshold, the neighborhood of the current pixel is determined to be a region with significant mulch film degradation characteristics. For the multispectral band observation matrix of this region, a high-dimensional feature-preserving mapping using an orthogonal decomposition family is performed, mapping the multispectral band observation matrix to the main degradation feature vector set, including: The multispectral band observation matrix of the region identified as having significant characteristics of plastic film degradation was extracted and used as input data for the high-dimensional feature preservation mapping. The input data is processed by feature mapping based on orthogonal decomposition family, which decomposes the multispectral band observation matrix into multiple orthogonal feature components. The eigenvalues ​​characterize the retention intensity of each orthogonal feature component for the spectral differences of mulch film degradation, and their values ​​are constrained by the covariance structure of the multispectral band observation matrix. Multiple orthogonal feature components are sorted according to the magnitude of their eigenvalues, and orthogonal feature components with eigenvalues ​​greater than a preset contribution threshold are selected as the main degradation feature components. The selected main degradation feature components and their corresponding original pixel spatial coordinates are recombined into vectors to obtain the main degradation feature vector set.

5. The intelligent identification method for the degradation stage of biodegradable mulch film based on multispectral image fusion according to claim 4, characterized in that, When the local spectral variation coefficient is less than or equal to a preset variation threshold, the neighborhood of the current pixel is determined to be a background noise region. For the multispectral band observation matrix of the background noise region, a low-dimensional smoothing dimensionality reduction mapping using a statistical inference family is performed, mapping the multispectral band observation matrix to a background suppression feature vector set, including: The multispectral band observation matrix of the region identified as background noise is extracted and used as input data for low-dimensional smooth dimensionality reduction mapping. The input data is processed by a family of statistical inferences to achieve smooth dimensionality reduction. The statistical distribution characteristics of the pixel reflectance data in the neighborhood are calculated. The smoothing weight represents the suppression strength of background noise, and its value is constrained by the variance of the reflectance data in the neighborhood window. Based on statistical distribution characteristics and smoothing weights, the multispectral band observation matrix of the background noise region is weighted, smoothed and aggregated to eliminate high-frequency spectral fluctuations and obtain low-dimensional smoothed feature components. The low-dimensional smooth feature components and their corresponding original pixel spatial coordinates are reduced in dimension and recombined to obtain a background suppression feature vector set.

6. The intelligent identification method for the degradation stage of biodegradable mulch film based on multispectral image fusion according to claim 5, characterized in that, The main degradation feature vector set and the background suppression feature vector set are aligned and fused at the feature level according to the original pixel space coordinates to obtain the region-differentiated fused feature tensor, including: Based on the pixel position identifiers carried in the main degradation feature vector set and the background suppression feature vector set, the corresponding original pixel spatial coordinate indices are extracted to obtain the spatial coordinate mapping set; Based on the spatial coordinate mapping set, the feature elements of the main degradation feature vector set and the background suppression feature vector set are reordered by gridding to eliminate the spatial discrepancy caused by the region division and obtain the coordinate aligned feature matrix. Based on the coordinate-aligned feature matrix, a feature space recombination process based on tensor algebra family is performed to map the coordinate-aligned feature matrix along the channel dimension to obtain the initial fused feature tensor. The recombination weight represents the aggregation strength of spatial continuity, and its value is constrained by the original number of channels of the multispectral band observation matrix. Based on the initial fusion feature tensor, topological interpolation smoothing aggregation based on manifold embedding families is performed on the feature vectors at the boundaries of different regions within the tensor to eliminate spectral faults caused by feature-level splicing, thus obtaining a regionally differentiated fusion feature tensor.

7. The intelligent identification method for the degradation stage of biodegradable mulch film based on multispectral image fusion according to claim 6, characterized in that, Based on the initial fused feature tensor, topological interpolation smoothing aggregation based on manifold embedding families is performed on the feature vectors at the boundaries of different regions within the tensor to eliminate spectral faults caused by feature-level splicing, resulting in a region-differentiated fused feature tensor, including: Based on the initial fusion feature tensor, feature vectors at the boundaries of different regions within the tensor are extracted to obtain a set of boundary feature vectors; Based on the set of boundary feature vectors, a dimensionality reduction mapping process based on manifold embedding family is performed to map the high-dimensional boundary feature vectors to the low-dimensional manifold space, resulting in a set of coordinate points in the manifold space. The manifold curvature represents the degree of continuous change of the spectral features of the boundary region, and its value is constrained by the local gradient of the initial fused feature tensor. Based on the set of coordinate points in the manifold space, the geodesic distance between adjacent feature points is calculated, and topological interpolation is performed on the feature vectors between adjacent regions based on the geodesic distance to fill the spectral faults generated by feature-level splicing, resulting in an interpolated smooth feature vector set. Based on the interpolated smooth feature vector set, the interpolation result is mapped back to the original high-dimensional feature space, and tensor recombination is performed with the non-boundary region features in the initial fused feature tensor to eliminate spatial discretization bias and obtain the regionally differentiated fused feature tensor.

8. The intelligent identification method for the degradation stage of biodegradable mulch film based on multispectral image fusion according to claim 7, characterized in that, The degradation index of the mulch film is calculated based on the regionally differentiated fusion feature tensor. The degradation index is then matched with multiple preset stage threshold ranges to determine the corresponding biodegradable mulch film degradation stage labels, including: Based on the regional differential fusion feature tensor, the multispectral band feature components contained in the regional differential fusion feature tensor are extracted to obtain the multispectral feature component set. Based on the set of multi-band feature components, a feature weighted aggregation process based on a family of variational optimizations is performed to map the set of multi-band feature components to a one-dimensional scalar space to obtain the initial mulch film degradation index. The aggregation weight characterizes the sensitivity of each band to the degradation state, and its value is constrained by the channel variance of the set of multi-band feature components. Based on the initial mulch film degradation index, multiple preset stage threshold intervals determined by statistical analysis of historical field multispectral data are introduced, and a numerical interval matching operation is performed to obtain the mulch film degradation index interval matching result. Based on the interval matching results of the mulch film degradation index, the mulch film degradation index falling into a specific interval is mapped to the corresponding status identifier to obtain the corresponding biodegradable mulch film degradation stage label.

9. The intelligent identification method for the degradation stage of biodegradable mulch film based on multispectral image fusion according to claim 8, characterized in that, Based on the initial plastic film degradation index, multiple preset stage threshold intervals determined by statistical analysis of historical field multispectral data are introduced. A numerical interval matching operation is then performed to obtain the plastic film degradation index interval matching results, including: Based on the initial mulch film degradation index and multiple preset stage threshold intervals, the upper and lower boundary values ​​of each stage threshold interval are extracted to obtain the stage threshold boundary set. Based on the set of stage threshold boundaries, the initial mulch film degradation index is compared with the values ​​of each upper and lower boundary to determine the target threshold interval to which the initial mulch film degradation index belongs, and the interval assignment result is obtained. Based on the interval attribution determination results, the initial mulch film degradation index at the boundary of adjacent stage threshold intervals is subjected to smooth transition mapping processing based on fuzzy membership families to eliminate the identification mutation caused by stage jumps and obtain the corrected interval attribution determination results. Based on the revised interval assignment results, the target threshold interval to which the initial plastic film degradation index belongs is determined, and the plastic film degradation index interval matching results are obtained.

10. A smart identification system for the degradation stage of biodegradable mulch film based on multispectral image fusion, wherein the system implements the method as described in any one of claims 1 to 9, characterized in that, include: The acquisition module is used to acquire multispectral images of biodegradable mulch films in the field. The multispectral images contain reflectance data of three bands, visible light, near infrared and shortwave infrared, which are acquired simultaneously. A multispectral band observation matrix is ​​constructed based on the multispectral images. The calculation module is used to calculate the local spectral variation coefficient of each pixel neighborhood in the multispectral image, use the local spectral variation coefficient as the criterion for spatial region division, and introduce a preset variation threshold determined by statistical analysis of historical field multispectral data. The determination module is used to determine that the neighborhood of the current pixel is a region with significant mulch film degradation characteristics when the local spectral variation coefficient is greater than the preset variation threshold. For the multispectral band observation matrix of the region with significant mulch film degradation characteristics, a high-dimensional feature preservation mapping of the orthogonal decomposition family is performed to map the multispectral band observation matrix to the main degradation feature vector set. When the local spectral variation coefficient is less than or equal to the preset variation threshold, the neighborhood of the current pixel is determined to be a background noise region. For the multispectral band observation matrix of the background noise region, a low-dimensional smooth dimensionality reduction mapping of the statistical inference family is performed to map the multispectral band observation matrix to the background suppression feature vector set. The fusion module is used to perform feature-level alignment and fusion processing on the main degradation feature vector set and the background suppression feature vector set according to the original pixel space coordinates to obtain the regional differential fusion feature tensor. The processing module is used to calculate the mulch film degradation index based on the regionally differentiated fusion feature tensor, and to match and determine the corresponding biodegradable mulch film degradation stage label based on the mulch film degradation index and multiple preset stage threshold intervals.

Images