A method and system for evaluating and managing the quality grade of cultivated land resources

Abstract

The present invention relates to the field of computer technology. The present invention discloses a method and system for evaluating and managing the quality grade of cultivated land resources. The method includes performing spectral residual normalization processing on multi-temporal remote sensing images to generate a sequence of corrected images, effectively suppressing the reflectance differences caused by the wet-dry alternation of water bodies, using deep segmentation and progressive edge detection to extract the first boundary result and the second boundary result respectively, enhancing the adaptability and robustness of boundary extraction, further generating a difference map based on the complex phase difference matrix and correcting the second boundary to obtain a coherent boundary result, significantly improving the boundary temporal consistency, and finally constructing cultivated land plot units based on the coherent boundary and inputting them into a random forest model to complete scoring and grade division, ensuring the stability and traceability of the cultivated land quality evaluation results under multi-temporal conditions, thereby achieving the technical goal of stably extracting boundaries and accurately completing grade division in a dynamically changing scenario.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and more specifically, to a method and system for evaluating and managing the quality grade of cultivated land resources. Background Art

[0002] In the prior art, the remote sensing evaluation method of cultivated land quality mainly relies on extracting the cultivated land boundary from single-temporal remote sensing images and constructing an evaluation model based on indicators such as vegetation index, soil moisture, and terrain factors. Typical solutions use classification models such as convolutional neural networks or support vector machines to perform pixel-level or region-level classification on remote sensing images, thereby determining the cultivated land category, and calculating the cultivated land quality score based on multi-source indicators within the plot. Some studies also introduce temporal remote sensing data for farmland dynamic analysis, but mostly focus on statistical averaging or maximum value composition of time series, ignoring the drastic changes in cultivated land status between the waterlogging period and the non-waterlogging period, and lacking targeted modeling means for maintaining boundary coherence and controlling quality fluctuations under seasonal hydrological fluctuations.

[0003] However, in cultivated land scenarios with frequent switching between the waterlogging period and the non-waterlogging period, traditional methods are easily interfered by factors such as the alternation of water body wet and dry and the change of soil reflectivity during boundary extraction, resulting in the fracture or offset of the plot contour, affecting the stability of subsequent plot unit construction and quality evaluation. Existing evaluation methods mostly rely on static or average boundaries, which are difficult to reflect the actual impact of boundary dynamic changes on the trend of cultivated land quality, and are prone to large fluctuations in the grading results between multiple temporal phases. Therefore, how to stably extract the boundary and accurately complete the grading of cultivated land quality in the dynamic change scenario between the waterlogging period and the non-waterlogging period has become an urgent technical problem to be solved.

[0004] In view of this, the present invention proposes a method and system for evaluating and managing the quality grade of cultivated land resources to solve the above problems. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art, the present invention provides a method and system for evaluating and managing the quality grade of cultivated land resources.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] In a first aspect, a method for evaluating and managing the quality grade of cultivated land resources is provided, including:

[0008] Obtaining multi-temporal remote sensing images of a target area, performing spectral residual normalization processing on the multi-temporal remote sensing images to generate a sequence of corrected images, and based on the sequence of corrected images, performing continuity marking processing to generate a set of remote sensing images in the waterlogging period and a set of remote sensing images in the non-waterlogging period;

[0009] Perform deep segmentation on the remote sensing images in the water accumulation period to obtain the first boundary result. After performing regional texture adaptive filtering on the remote sensing images in the non-water accumulation period, perform edge detection progressively at the first scale and the second scale to obtain the second boundary result, where the first scale is greater than the second scale;

[0010] Reproject the first boundary result into the coordinate system of the second boundary result to obtain the complex phase difference matrix of the corresponding pixels in the water accumulation period and the non-water accumulation period. Calculate the pixel phase deviation based on the complex phase difference matrix, generate a difference map, and correct the second boundary result in the difference map according to the boundary continuity criterion to obtain the coherent boundary result;

[0011] Construct cultivated land plot units based on the coherent boundary results of different time phases. Input the cultivated land plot units into the trained random forest model to obtain the cultivated land quality score, and divide the cultivated land quality grade according to the cultivated land quality score and the preset quality threshold.

[0012] In some embodiments, the method for performing spectral residual normalization processing on multi-temporal remote sensing images to generate a sequence of corrected images includes:

[0013] Based on each temporal remote sensing image, select the first reference pixel set and the second reference pixel set, and calculate the spectral residual vectors of each band with respect to the two types of reference pixel sets. The multi-temporal remote sensing images include Q adjacent temporal remote sensing images;

[0014] Based on the spectral residual vectors, determine the normalization adjustment parameters of each band, and combine the normalization adjustment parameters to obtain a correction parameter matrix;

[0015] Perform radiometric normalization on the correction parameter matrix to obtain a normalized image;

[0016] Using the correction parameter matrix as the input, calculate the spectral residual difference between adjacent temporal remote sensing images in chronological order to obtain a temporal difference matrix;

[0017] Perform interpolation diffusion operation based on the temporal difference matrix to generate a drift compensation field, and superimpose the drift compensation field on the normalized image according to the preset smoothing constraint to obtain a sequence of corrected images.

[0018] In some embodiments, the method for performing continuity marking processing based on the sequence of corrected images to generate a set of remote sensing images in the water accumulation period and a set of remote sensing images in the non-water accumulation period includes:

[0019] In the sequence of corrected images, based on the corresponding pixels of adjacent temporal remote sensing images, calculate their improved water body indices in complex form respectively, and generate a sequence of phase state label maps according to the improved water body indices;

[0020] Perform a sliding window voting operation on the sequence of phase state label maps along the time dimension to obtain a continuity marking matrix;

[0021] Based on the continuous label matrix, count the number of persistent water pixels in each temporal remote sensing image, and calculate the pixel ratio of the number of persistent water pixels to the total number of pixels in the image;

[0022] When the pixel ratio is not less than the preset persistent ratio threshold, classify the temporal remote sensing image into the set of remote sensing images in the waterlogging period. Otherwise, classify the temporal remote sensing image into the set of remote sensing images in the non-waterlogging period.

[0023] In some embodiments, the method of selecting the first reference pixel set and the second reference pixel set based on each temporal remote sensing image includes:

[0024] Calculate the temporal coefficient of variation of the multi-band brightness for each pixel of Q adjacent temporal remote sensing images, and generate a spectral stability grid;

[0025] In the spectral stability grid, select the pixels with a temporal coefficient of variation not higher than the preset variation threshold to construct a stable candidate set;

[0026] Perform hierarchical sampling on the stable candidate set, specifically including:

[0027] Sample from the pixels where the brightness in the red band and the short-wave infrared band are both below the preset low brightness percentile threshold, and combine morphological opening operation to remove isolated pixels to obtain the first reference pixel set;

[0028] Sample from the pixels where the brightness in the green band and the near-infrared band are both above the preset high brightness percentile threshold to obtain the second reference pixel candidate area;

[0029] In the second reference pixel candidate area, adopt a spatial grid uniform sampling strategy to extract pixels, and perform spectral homogeneity test on the 3×3 neighborhood centered on the grid center of each extracted pixel. The pixels passing the test construct the second reference pixel set.

[0030] In some embodiments, the method of generating a phase label map sequence according to the improved water index includes:

[0031] Based on each temporal remote sensing image, extract the multi-band normalized brightness value for water body identification, and calculate the improved water index according to the preset weighted combination coefficient and the normalized brightness value;

[0032] For the improved water index of each pair of adjacent temporal remote sensing images, extract the index values of the corresponding pixels respectively. Use the index value of the previous temporal remote sensing image as the real part of the complex number, and the index value of the latter temporal remote sensing image as the imaginary part of the complex number to generate a complex number expression matrix;

[0033] For each complex number in the complex expression matrix, calculate the complex phase angle, which is obtained by calculating the arctangent value of the real part and the imaginary part of the complex number;

[0034] Arrange the complex phase angles of all pixels in the original spatial position to generate the pixel complex phase difference;

[0035] When the absolute value of the pixel complex phase difference is less than the preset phase change threshold, it is marked as a non-flooded state, otherwise it is marked as a flooded state;

[0036] Arrange the phase state labels of all pixels in the original spatial position to form a phase state label map consistent with the size of the original image;

[0037] Process each pair of adjacent temporal remote sensing images in sequence to generate a sequence of phase state label maps.

[0038] In some embodiments, the method for performing depth segmentation on the flooded period remote sensing image to obtain the first boundary result includes:

[0039] Perform superpixel segmentation on the flooded period remote sensing image according to a preset segmentation scale to generate an initial region segmentation map, where the initial region segmentation map includes K superpixel regions;

[0040] Obtain the normalized brightness mean, local texture directionality index, and shape compactness feature of each superpixel region, and splice the normalized brightness mean, local texture directionality index, and shape compactness feature in a preset order to construct a region-level feature vector;

[0041] Take the region-level feature vector as the input, perform region-level classification through a preset convolutional neural network model, and output the water body prediction map of each superpixel region;

[0042] Perform connected region analysis on the water body prediction map, screen out the largest connected water body region, and perform region edge clarity enhancement processing on the connected water body region to obtain a preliminary boundary result;

[0043] Perform edge correction processing on the preliminary boundary result to obtain the first boundary result.

[0044] In some embodiments, divide the non-flooded period remote sensing image into H sliding window regions of a fixed size, and perform region texture adaptive filtering processing within each sliding window region to obtain a filtered image;

[0045] Based on the first scale, set the first edge detection operator, and perform the first round of edge detection on the filtered image according to the first edge detection operator to extract the main edge contour of the coarse grain and generate the first round of edge map;

[0046] Based on the positions in the first-round edge map where the edge response values are non-zero, construct a local mask region, set a second-edge detection operator according to the second scale, perform a second-round edge detection on the local mask region according to the second-edge detection operator, extract fine-grained supplementary edges, and generate a second-round edge map;

[0047] Perform edge merging processing on the first-round edge map and the second-round edge map, preferentially retain the main contour of the first-round edge, and introduce the second-round edge supplement at the positions where the edge is broken to form an edge fusion map;

[0048] Perform connectivity screening processing on the edge fusion map to generate a second boundary result.

[0049] In some embodiments, the method for obtaining the complex phase difference matrix of corresponding pixels in the waterlogging period and the non-waterlogging period by reprojection of the first boundary result into the coordinate system of the second boundary result includes:

[0050] Perform resampling processing on the first boundary result according to the spatial grid of the second boundary result so that the pixel row and column coordinates of the first boundary result are consistent with those of the second boundary result, where the resampling processing is based on the nearest neighbor interpolation principle;

[0051] Based on the resampled first boundary result and the second boundary result, establish the corresponding pixel relationship between the waterlogging period and the non-waterlogging period according to the pixel row and column indices, and screen out invalid pixel pairs to obtain valid corresponding pixels, where the invalid pixel pairs include pixels outside the boundary and abnormally missing pixels;

[0052] Based on each pair of valid corresponding pixels, extract their corresponding complex phase difference values and write them into the complex phase difference matrix in spatial order.

[0053] In some embodiments, the method for calculating the pixel phase deviation based on the complex phase difference matrix and generating a difference map includes:

[0054] Traverse the complex phase difference matrix pixel by pixel to obtain the complex phase difference values of each valid corresponding pixel;

[0055] Based on each pair of valid corresponding pixels, calculate the absolute value of the complex phase difference between the pixel in the waterlogging period and the pixel in the non-waterlogging period to obtain the pixel phase deviation, where the pixel phase deviation is used to characterize the change amplitude of the corresponding pixels in the waterlogging period and the non-waterlogging period in the complex phase angle;

[0056] Arrange the pixel phase deviations of all valid corresponding pixels according to the original row and column indices to generate a difference map with the same spatial size as the second boundary result.

[0057] In a second aspect, a cultivated land resource quality grade evaluation and management system is provided, which is used to implement the above-mentioned cultivated land resource quality grade evaluation and management method, including:

[0058] The first processing module: is used to obtain multi-temporal remote sensing images of the target area, perform spectral residual normalization processing on the multi-temporal remote sensing images to generate a corrected image sequence, and based on the corrected image sequence, perform continuity marking processing to generate a set of remote sensing images during the water accumulation period and a set of remote sensing images during the non-water accumulation period;

[0059] The second processing module: is used to perform depth segmentation on the remote sensing images during the water accumulation period to obtain a first boundary result, perform regional texture adaptive filtering on the remote sensing images during the non-water accumulation period, and then perform edge detection progressively according to the first scale and the second scale to obtain a second boundary result, where the first scale is greater than the second scale;

[0060] The correction module: is used to reproject the first boundary result into the coordinate system of the second boundary result to obtain a complex phase difference matrix of corresponding pixels during the water accumulation period and the non-water accumulation period, calculate the pixel phase deviation based on the complex phase difference matrix to generate a difference map, and correct the second boundary result in the difference map according to the boundary continuity criterion to obtain a coherent boundary result;

[0061] The evaluation module: is used to construct cultivated land plot units based on the coherent boundary results of different time phases, input the cultivated land plot units into a trained random forest model to obtain cultivated land quality scores, and divide the cultivated land quality grades according to the cultivated land quality scores and a preset quality threshold.

[0062] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0063] By performing spectral residual normalization processing on multi-temporal remote sensing images to generate a corrected image sequence, the present invention effectively suppresses the reflectance difference caused by the water body's wet-dry alternation. The first boundary result and the second boundary result are extracted respectively by using depth segmentation and progressive edge detection, which enhances the adaptability and robustness of boundary extraction. Further, a difference map is generated based on the complex phase difference matrix and the second boundary is corrected to obtain a coherent boundary result, significantly improving the boundary temporal consistency. Finally, cultivated land plot units are constructed based on the coherent boundary and input into the random forest model to complete scoring and grade division, ensuring the stability and traceability of the cultivated land quality assessment results under multi-temporal conditions, thereby achieving the technical goal of stably extracting boundaries and accurately completing grade division in a dynamically changing scenario. Description of the Drawings

[0064] Figure 1 It is a schematic flow chart of a method for evaluating and managing the quality grade of cultivated land resources in the present invention;

[0065] Figure 2 It is a schematic diagram of the change trend of the water accumulation state label map in the present invention;

[0066] Figure 3 It is a schematic diagram for comparing the first boundary result, the second boundary result and the coherent boundary result in the present invention;

[0067] Figure 4 It is the pixel phase deviation difference map in the present invention;

[0068] Figure 5 It is the schematic diagram for constructing cultivated land plot units based on coherent boundaries in the present invention;

[0069] Figure 6 It is the schematic structural diagram of a cultivated land resource quality grade evaluation and management system in the present invention. Specific implementation manners

[0070] To make the objectives, technical solutions, and advantages of the present invention more clear and understandable, the following further elaborates on the present invention in detail in combination with specific embodiments and with reference to the accompanying drawings. In the following detailed description, many specific details are set forth to provide a thorough understanding of the described exemplary embodiments. However, it is obvious to those skilled in the art that some or all of these specific details may not be required to practice the described embodiments. In other exemplary embodiments, well-known structures are not described in detail to avoid unnecessarily obscuring the concepts of the present disclosure. It should be understood that the specific embodiments described herein are merely used to explain the present invention and are not used to limit the present invention. At the same time, various aspects described in the embodiments can be arbitrarily combined without conflict.

[0071] Embodiment 1

[0072] Please refer to Figure 1 As shown, this embodiment discloses and provides a cultivated land resource quality grade evaluation and management method, including:

[0073] S10: Obtain multi-temporal remote sensing images of the target area, perform spectral residual normalization processing on the multi-temporal remote sensing images to generate a sequence of corrected images, and based on the sequence of corrected images, perform continuity marking processing to generate a set of remote sensing images during the waterlogging period and a set of remote sensing images during the non-waterlogging period;

[0074] In this embodiment, the target area can be any spatial unit where the cultivated land to be evaluated is located, such as a township, county, city, or basin area, and can also be refined to a farmland renovation project area, a land consolidation block, or a national land "third survey" map patch, provided that the area can be completely covered by the same or mosaicked remote sensing images throughout the time series. The multi-temporal remote sensing images refer to a set of images of the same area obtained by the same-source or multi-source satellite / aerial / UAV sensors at different observation times (such as different months, years, or key crop growth stages). Typical sources include Sentinel-2, Landsat-8 / 9, GF-6, Planet Scope, etc.; its temporal resolution can be set as dekad, month, or season according to research requirements, aiming to characterize the dynamic changes in the spectral states of paddy fields during the waterlogging period and the non-waterlogging period.

[0075] A method for performing spectral residual normalization on multi-temporal remote sensing images to generate a sequence of corrected images includes:

[0076] Based on each multi-temporal remote sensing image, a first reference pixel set and a second reference pixel set are selected, and spectral residual vectors for each band with respect to the two types of reference pixel sets are calculated. The multi-temporal remote sensing images include Q adjacent multi-temporal remote sensing images;

[0077] Based on the spectral residual vectors, normalization adjustment parameters for each band are determined, and the normalization adjustment parameters are combined to obtain a correction parameter matrix;

[0078] Radiometric normalization is performed on the correction parameter matrix to obtain a normalized image;

[0079] Taking the correction parameter matrix as input, spectral residual differences between adjacent multi-temporal remote sensing images are calculated in chronological order to obtain a time series difference matrix;

[0080] Based on the time series difference matrix, interpolation diffusion operations are performed to generate a drift compensation field, and the drift compensation field is superimposed on the normalized image according to a preset smoothing constraint to obtain a sequence of corrected images.

[0081] A method for performing continuity marking processing on the sequence of corrected images to generate a set of remote sensing images during the water accumulation period and a set of remote sensing images during the non-water accumulation period includes:

[0082] In the sequence of corrected images, based on corresponding pixels of adjacent multi-temporal remote sensing images, an improved water index in complex form is calculated for each of them, and a sequence of phase state label maps is generated according to the improved water index;

[0083] A sliding window voting operation is performed on the sequence of phase state label maps along the time dimension to obtain a continuity marking matrix;

[0084] Based on the continuity marking matrix, the number of water body continuous pixels in each multi-temporal remote sensing image is counted, and the pixel ratio of the number of water body continuous pixels to the total number of pixels in the image is calculated;

[0085] When the pixel ratio is not less than a preset continuous ratio threshold, the multi-temporal remote sensing image is classified into the set of remote sensing images during the water accumulation period; otherwise, it is classified into the set of remote sensing images during the non-water accumulation period.

[0086] It should be noted that the sliding window step size can be set to 1 frame to ensure that each multi-temporal image is covered by multiple votes. In each multi-temporal remote sensing image, first, a first reference pixel set and a second reference pixel set are determined, and then the pixel brightness of each band of the entire image is subtracted by the average brightness of the corresponding reference pixel set one by one to directly obtain the spectral residual vector. Then, the spectral residual vector is scaled by the brightness standard deviation of the same reference pixel set to generate a normalized image with the same size as the original image and a unified spectral scale.

[0087] In chronological order, the spectral residual vectors of two adjacent normalized images are subtracted pixel by pixel to form a complete temporal difference matrix. This matrix is regarded as the drift field, and through interpolation-diffusion, a continuous drift compensation field is obtained by smoothing and filling in a local range. Then, the drift compensation field is added to the normalized image pixel by pixel to generate a corrected image sequence with a unified brightness baseline, laying a consistent spectral benchmark for subsequent phase analysis.

[0088] On the phase label map corresponding to the corrected image sequence, a sliding window with a fixed length (such as 5 frames) is set along the time axis. If the number of times the same pixel is marked as water accumulation within the frames covered by the window reaches more than half of the window length, the pixel is set to 1 at the center frame of the window; otherwise, it is set to 0. As the window slides frame by frame and writes the results of the center frame, a continuous marking matrix with the same size as the original image sequence and pixel values only of 0 / 1 can be obtained, where 1 indicates that the pixel is identified as continuous water body at this phase.

[0089] As Figure 2 shown, it shows the change of the water accumulation state label of a typical pixel in Example 1 under 20 consecutive time steps. The horizontal axis is the time step number, and the vertical axis is the corresponding water accumulation state of the pixel in each phase. 1 indicates water accumulation, and 0 indicates non-water accumulation. This figure is based on the corrected image sequence. By sliding a window with a length of 5 along the time axis on the corresponding phase label map and performing a majority voting operation on each pixel, the marking results at the corresponding positions in the time continuity matrix are obtained. From the trend in the figure, it can be seen that this pixel maintains the water accumulation state in multiple consecutive phases, showing strong temporal consistency, proving that the sliding voting mechanism can effectively eliminate the interference of abnormal labels in individual phases and improve the robustness of water accumulation state recognition.

[0090] For each image in the continuous marking matrix, first count the number of pixels marked as 1 and calculate the ratio to the total number of pixels in the image to obtain the continuous ratio of this image. When the continuous ratio is not lower than the preset continuous ratio threshold (such as 25%), this image is classified into the set of remote sensing images of the water accumulation period; otherwise, it is classified into the set of remote sensing images of the non-water accumulation period. The set of remote sensing images of the water accumulation period and the set of remote sensing images of the non-water accumulation period formed in this way provide a unified and stable time segmentation basis for subsequent boundary detection and cultivated land resource quality grade evaluation.

[0091] The methods for selecting the first reference pixel set and the second reference pixel set based on each temporal remote sensing image include:

[0092] Calculate the temporal coefficient of variation of the multi-band brightness pixel by pixel for Q adjacent temporal remote sensing images and generate a spectral stability grid;

[0093] In the spectral stability grid, select the pixels with a temporal coefficient of variation not higher than the preset variation threshold to construct a stable candidate set;

[0094] Perform hierarchical sampling on the stable candidate set, specifically including:

[0095] Sample from the pixels where the brightness in both the red band and the short-wave infrared band is below the preset low-brightness percentile threshold, and combine morphological opening operation to remove isolated pixels to obtain the first reference pixel set;

[0096] Sample from the pixels where the brightness in both the green band and the near-infrared band is above the preset high-brightness percentile threshold to obtain the second reference pixel candidate area;

[0097] In the second reference pixel candidate area, adopt a spatial grid uniform sampling strategy to extract pixels, and perform spectral homogeneity test on the 3×3 neighborhood centered on the grid center of each extracted pixel. The pixels that pass the test are used to construct the second reference pixel set.

[0098] It can be understood that the coefficient of variation of time refers to a dimensionless ratio obtained by taking the standard deviation of a multi-band brightness sequence of the same pixel in Q adjacent temporal remote sensing images as the input and then dividing it by the sequence mean, which is used to measure the spectral fluctuation degree of the pixel in the time dimension. Generally, it is calculated separately for each band and then the multi-band average or weighted average is taken.

[0099] The spectral stability raster refers to a raster layer with the coefficient of variation of time as the pixel value; the smaller the pixel value, the more stable the pixel is in the time series. This raster is consistent with the original image in spatial resolution and row-column size, which is convenient for subsequent pixel screening.

[0100] The spatial grid uniform sampling strategy can be to divide the candidate area into regular grids with equal spacing (such as 50m×50m or in units of n×n pixels), and extract 1 pixel at the center or a random position of each grid; if the grid center falls in an invalid area, it is replaced by the nearest valid pixel. This strategy can avoid over-concentration of samples in a local area. The spectral homogeneity test refers to performing local statistics on the brightness of each band in the 3×3 neighborhood of the extracted pixel. When the maximum–minimum difference or variance of the pixel brightness in the neighborhood does not exceed the preset threshold, it is judged as spectrally homogeneous; otherwise, the central pixel is discarded and resampled in the corresponding grid.

[0101] In this embodiment, only the pixels that pass the test are retained as the second reference pixel set. Compared with the conventional "single-period fixed threshold" or "random sampling" in the prior art, this process superimposes time stability constraints, dual-band joint discrimination, and spatial uniform distribution control in the reference pixel selection link, significantly reducing the error amplification caused by seasonal brightness drift, local abnormal features, or sample aggregation.

[0102] By using the above first reference pixel set and second reference pixel set, a long-term stable and fully dynamic-range radiation reference can be provided in a paddy field scenario with frequent switching between the water accumulation period and the non-water accumulation period: the dark end (both red and short-wave infrared are "low") and the bright end (both green and near-infrared are "high") are simultaneously locked, so that the spectral residual vector and the normalized image maintain a unified brightness scale among multiple time phases; at the same time, the spatial grid uniform sampling strategy combined with the spectral homogeneity test ensures that the two types of reference pixels are evenly distributed and locally consistent, laying a reliable foundation for the accurate construction of subsequent temporal difference matrices, drift compensation fields, and corrected image sequences.

[0103] In this embodiment, performing interpolation diffusion operation based on the temporal difference matrix can be to first apply inverse distance weighted interpolation to fill in the missing values of the effective drift pixels in the temporal difference matrix, and then perform one iteration of diffusion on the same grid using a 3×3 mean convolution kernel to gradually smooth the drift compensation field in a way of decreasing spatial weight; when the local gradient is greater than the preset threshold, the diffusion can be stopped to retain the edge details. The preset smoothing constraint value rule can be based on the root mean square error of the local brightness noise in the corrected image sequence, and set the smoothing constraint coefficient λ = k×RMSE, where k takes the empirical range of 0.5–1.0, and RMSE is the root mean square error. When the overall variance of the temporal difference matrix is observed to increase, appropriately increase λ to obtain a stronger smoothing effect, and vice versa, decrease λ to avoid excessive blurring.

[0104] The method for generating a phase state label map sequence according to the improved water body index includes:

[0105] Based on each temporal remote sensing image, extract the multi-band normalized brightness values for water body identification, and calculate the improved water body index according to the preset weighted combination coefficient and the normalized brightness values;

[0106] For the improved water body index of each pair of adjacent temporal remote sensing images, respectively extract the index values of the corresponding pixels, use the index value of the previous temporal remote sensing image as the real part of the complex number, and the index value of the latter temporal remote sensing image as the imaginary part of the complex number to generate a complex number expression matrix;

[0107] For each complex number in the complex number expression matrix, calculate the complex number phase angle, and the complex number phase angle is obtained by calculating the arctangent value of the real part and the imaginary part of the complex number;

[0108] Arrange the complex number phase angles of all pixels according to the original spatial position to generate the pixel complex phase difference;

[0109] When the absolute value of the pixel complex phase difference is less than the preset phase change threshold, it is marked as the non-water accumulation state, and vice versa, it is marked as the water accumulation state;

[0110] Arrange the phase state labels of all pixels according to the original spatial position to form a phase state label map with the same size as the original image;

[0111] Process each pair of adjacent temporal remote sensing images in sequence to generate a sequence of phase state label maps.

[0112] In this embodiment, the four-quadrant arctangent function (such as atan2) is used to calculate the complex phase angle to ensure that the angle range covers -π to +π. The weighted combination coefficients refer to a set of coefficients assigned according to the importance of multiple preset bands for water body identification in the normalized image for differentiating water bodies. After normalization, the weighted combination coefficients are fixed and input as fixed parameters. Usually, the normalized brightness values of the visible light band (such as the green band), the near-infrared band, and the short-wave infrared band are selected, and a linear combination is performed according to the weighted combination coefficients. For example, for each pixel:

[0113] Improved water index = a × normalized green band brightness + b × normalized near-infrared band brightness + c × normalized short-wave infrared band brightness;

[0114] Where a, b, and c are preset weighted combination coefficients. The above weighted combination formula is executed after image normalization, avoiding the absolute brightness deviation introduced by different observation conditions of the original image.

[0115] In this embodiment, arranging in the original spatial position means that after calculating the complex phase angle for each pixel in the complex expression matrix, the row and column coordinates of the pixel in the remote sensing image grid data remain unchanged. Each complex phase angle value is directly filled into the corresponding spatial position to form a new two-dimensional matrix (i.e., the pixel complex phase difference image). During the entire arrangement process, the pixels are not moved, rotated, resampled, or spatially interpolated, and only direct assignment is performed based on the original spatial grid order, ensuring that the output image of the complex phase angle is strictly consistent with the original normalized image and the improved water index image in terms of spatial size and resolution.

[0116] The pixel complex phase difference is an angular quantity that comprehensively characterizes the change trend of the water index of the corresponding ground object in adjacent temporal remote sensing images in a complex space manner. By taking the improved water index of the previous temporal phase as the real part of the complex number and the improved water index of the subsequent temporal phase as the imaginary part of the complex number, a complex expression matrix is formed, and the arctangent phase angle of the complex number is calculated, which can quantitatively describe the relative change direction of the water accumulation state in time variation.

[0117] In this embodiment, a processing chain is introduced in the processing logic, which includes generating an improved water body index based on spectral residual vector normalization and multi-band weighted combination, and calculating the complex phase difference of pixels based on the complex expression matrix. By normalizing the images, the radiation scales of different time phases are unified. By improving the water body index, the spectral separation between the water body and the background ground objects is enhanced. Then, by correlating the changes in the water body index of the front and back time phases in complex form and extracting the complex phase difference, a dynamic quantitative expression of the waterlogging state change is achieved. Compared with the prior art, this solution avoids the problem that the fixed threshold discrimination is prone to failure under different waterlogging conditions, and can continuously output discriminant results with coherent structure and accurate change trend in areas with drastic water body changes or frequent wet-dry alternations.

[0118] Therefore, through the dual improvement of technical means and processing logic, this embodiment realizes the continuous improvement of the extraction of water body changes in multi-temporal remote sensing images and the enhancement of robustness without relying on a single spectral intensity or setting static classification rules, reduces the grade division fluctuations caused by the switching of the waterlogging state, and ensures the consistency and reliability of the cultivated land resource quality grade assessment on the time scale.

[0119] S20: Perform deep segmentation on the remote sensing images during the waterlogging period to obtain the first boundary result. After performing regional texture adaptive filtering on the remote sensing images during the non-waterlogging period, perform edge detection progressively according to the first scale and the second scale to obtain the second boundary result, where the first scale is greater than the second scale;

[0120] The method for performing deep segmentation on the remote sensing images during the waterlogging period to obtain the first boundary result includes:

[0121] Perform superpixel segmentation on the remote sensing images during the waterlogging period according to a preset segmentation scale to generate an initial regional segmentation map, where the initial regional segmentation map includes K superpixel regions;

[0122] Obtain the normalized brightness mean, local texture directionality index, and shape compactness feature of each superpixel region, splice the normalized brightness mean, local texture directionality index, and shape compactness feature in a preset order to construct a regional-level feature vector;

[0123] Use the regional-level feature vector as the input, perform regional-level classification through a preset convolutional neural network model, and output the water body prediction map of each superpixel region;

[0124] Perform connected region analysis on the water body prediction map, screen out the largest connected water body region, and perform regional edge sharpness enhancement processing on the connected water body region to obtain a preliminary boundary result;

[0125] Perform edge correction processing on the preliminary boundary result to obtain the first boundary result.

[0126] In this embodiment, the convolutional neural network model adopts a three-layer convolutional structure, with the size of each convolutional kernel being 3×3, ReLU activation is used, and the last layer outputs class probabilities for region classification. The segmentation scale refers to the parameter of the spatial aggregation degree set when performing superpixel segmentation, which is used to control the average size of the superpixel region. The larger the segmentation scale value, the larger the generated superpixel region. Generally, the value range is 3 to 10 times the spatial resolution of the image. For example, when the spatial resolution is 10 meters, the segmentation scale can be taken as 30 meters to 100 meters. The superpixel region represents a regular region unit formed by aggregating adjacent pixels through spectral similarity and spatial connectivity in the remote sensing image during the water accumulation period. The pixels within each superpixel region have high spectral consistency and spatial continuity. The local texture directionality index represents the degree of consistency of the pixel brightness distribution within the superpixel region in the spatial direction.

[0127] The shape compactness feature refers to an index used to measure the geometric shape regularity and boundary complexity of the superpixel region, usually defined as a function of the ratio of the region area to the area of its minimum bounding rectangle or the ratio of the perimeter squared to the area. The water body prediction chart represents the classification result layer output after performing region-level classification on each superpixel region based on the convolutional neural network model, where each superpixel region is assigned a water body class or non-water body class label. The class labels in the water body prediction chart are used to guide the subsequent connected region extraction and boundary correction to ensure that the first boundary result preferentially covers the water accumulation area.

[0128] The method for constructing a region-level feature vector based on the normalized brightness mean, local texture directionality index, and shape compactness feature includes:

[0129] Concatenate the normalized brightness mean, local texture directionality index, and shape compactness feature in a preset arrangement order to form a feature vector with a fixed dimension as the region-level feature vector of this superpixel region.

[0130] It should be added that the extraction of the water body region boundary usually adopts the method of directly extracting the contour after pixel-level classification, or performing global edge detection after simple threshold segmentation. Such methods are easily affected by noise pixels, fragmented small regions, and boundary blur problems, resulting in problems such as breaks, pseudo-connections, and edge burrs in the extracted water body boundary. Especially in the remote sensing image during the water accumulation period, due to the similar spectra of water bodies and surrounding moist soil and vegetation, traditional methods are difficult to ensure that the extracted boundary has spatial coherence and clear contours.

[0131] In this embodiment, by performing screening on the maximum connected sub-regions of the water body prediction map, the connected water body region with the largest area is preferentially extracted, and small-area interference regions are removed, ensuring that subsequent processing only operates on the main water body region, significantly reducing the interference of the noise region on the boundary quality. On this basis, further perform enhanced processing on the connected water body region for the clarity of the region edge. Through edge gradient reconstruction and local smooth reconstruction operations, the smoothness and integrity of the boundary of the connected region are improved, forming a preliminary boundary result with coherent structure and clear contour. Compared with the method of directly extracting the contour, this solution effectively reduces the number of pseudo-boundaries and improves the continuity and spatial accuracy of the main water body contour.

[0132] Through the above two-stage processing, this embodiment can stably extract the true boundary of the target water body under the complex spectral background during the waterlogging period, ensuring that a coherent and regular preliminary boundary is used as the basis for subsequent boundary correction processing, significantly reducing the boundary fracture correction amount and misjudgment rate, and finally obtaining a more accurate and coherent first boundary result, providing a high-quality plot boundary basis for the subsequent evaluation of the quality grade of cultivated land resources.

[0133] The method for obtaining the second boundary result by performing region texture adaptive filtering on the remote sensing image in the non-waterlogging period and then performing edge detection progressively according to the first scale and the second scale includes:

[0134] Divide the remote sensing image in the non-waterlogging period into H sliding window regions of a fixed size. In each sliding window region, perform region texture adaptive filtering processing to obtain a filtered image;

[0135] Set a first edge detection operator based on the first scale, and perform the first round of edge detection on the filtered image according to the first edge detection operator to extract the main edge contour with coarse granularity and generate the first-round edge map;

[0136] Based on the positions where the edge response values in the first-round edge map are not zero, construct a local mask region. Set a second edge detection operator based on the second scale, and perform the second round of edge detection on the local mask region according to the second edge detection operator to extract the supplementary edges with fine granularity and generate the second-round edge map;

[0137] Perform edge merging processing on the first-round edge map and the second-round edge map, preferentially retain the main contour of the first-round edge, and introduce the second-round edge supplement at the positions where the edge is broken to form an edge fusion map;

[0138] Perform connectivity screening processing on the edge fusion map to generate the second boundary result.

[0139] In this embodiment, according to the preset first scale value, the convolution kernel size, gradient calculation neighborhood range and edge response suppression parameter of the edge detection operator are determined, so that the edge detection operator can preferentially extract coarse-grained and main-structured edge features in the filtered image, wherein the convolution kernel size is positively correlated with the first scale. The larger the first scale, the larger the convolution kernel size; the edge detection neighborhood range is set to cover the pixel distance corresponding to the first scale; the edge response suppression parameter is set to adapt to large-scale texture changes to ensure that the response amplitude of the main edge is highlighted during the edge detection process, while the small texture disturbance is weakened. The second edge detection operator is set based on the second scale in the same way, and this embodiment will not go into details.

[0140] Based on the position where the edge response value in the first round edge map is not zero, the method of constructing the local mask area includes:

[0141] Traverse the first round of edge graph pixel by pixel and detect the edge response value of each pixel;

[0142] Mark the pixel positions with edge response values ​​greater than zero as candidate area pixels;

[0143] Taking each candidate area pixel as the center, a local window is formed by expanding it according to the preset local window size;

[0144] Merge the local windows to generate a local mask area.

[0145] In this embodiment, the local window size can be a fixed pixel size set according to the difference between the first scale and the second scale, and is usually set to a range of 5×5, 7×7 or 9×9 pixels, where the larger the local window size, the wider the coverage of the local mask area. The edge response value is a value obtained by performing a gradient operation or a brightness change rate calculation on each pixel, which is used to quantitatively characterize the intensity of the grayscale change in the local area of ​​the pixel. The edge response value is proportional to the brightness change amplitude of the corresponding position. The larger the edge response value, the more drastic the local grayscale change, and the more likely it is to correspond to a real edge position; conversely, when the edge response value is close to zero, it indicates that the grayscale change inside the area is gentle and the edge characteristics are not significant.

[0146] The method of performing connectivity screening on the edge fusion graph to generate a second boundary result includes:

[0147] Perform connected region marking processing on the edge fusion graph, and classify pixels with edge response values ​​greater than zero and spatially adjacent as the same connected region;

[0148] Calculate the area of ​​each connected region, where the area is the number of internal pixels in the connected region;

[0149] Remove the connected regions with an area smaller than the preset area threshold, and retain the connected regions with an area not smaller than the preset area threshold;

[0150] Mark the retained connected regions on the edge fusion map, extract the corresponding edge pixels, and generate the second boundary result.

[0151] It can be understood that the pixels with an edge response value greater than zero and spatially adjacent in the edge fusion map are aggregated into connected regions to avoid interference from single-pixel noise on boundary extraction; subsequently, calculate the area of each connected region, filter out isolated small-area noise regions based on the area size, and only retain the effective connected regions with an area not smaller than the preset area threshold; finally, extract the corresponding edge pixels according to the retained connected regions marked on the edge fusion map to ensure that the second boundary result only contains real boundaries with good continuity and reasonable spatial coverage.

[0152] Compared with the common method of directly extracting edges based on single-pixel response in the prior art, the existing method is prone to misidentifying noise pixels, local texture perturbations, or small pseudo-edges as effective boundaries, resulting in serious boundary fragmentation and breakage, which affects the accuracy of subsequent plot division and quality assessment.

[0153] S30: Reproject the first boundary result into the coordinate system of the second boundary result to obtain the complex phase difference matrix of the corresponding pixels in the waterlogging period and the non-waterlogging period, calculate the pixel phase deviation based on the complex phase difference matrix, generate a difference map, and correct the second boundary result in the difference map according to the boundary continuity criterion to obtain a coherent boundary result;

[0154] The method of reprojecting the first boundary result into the coordinate system of the second boundary result to obtain the complex phase difference matrix of the corresponding pixels in the waterlogging period and the non-waterlogging period includes:

[0155] Resample the first boundary result according to the spatial grid of the second boundary result to make the row and column coordinates of the pixels in the first boundary result consistent with those of the second boundary result, where the resampling process is based on the nearest neighbor interpolation principle;

[0156] Based on the resampled first boundary result and the second boundary result, establish the corresponding pixel relationship between the waterlogging period and the non-waterlogging period according to the pixel row and column indices, and filter out invalid pixel pairs to obtain effective corresponding pixels, where the invalid pixel pairs include pixels outside the boundary and abnormally missing pixels;

[0157] Based on each pair of effective corresponding pixels, extract their corresponding complex phase difference values and write them into the complex phase difference matrix in spatial order.

[0158] The method of calculating the pixel phase deviation based on the complex phase difference matrix and generating a difference map includes:

[0159] Traverse each pixel of the complex phase difference matrix to obtain the complex phase difference value of each valid corresponding pixel;

[0160] Based on each pair of valid corresponding pixels, calculate the absolute value of the complex phase difference between the waterlogged period pixel and the non-waterlogged period pixel to obtain the pixel phase deviation, where the pixel phase deviation is used to characterize the change amplitude of the corresponding pixels in the waterlogged period and the non-waterlogged period in the complex phase angle;

[0161] Arrange the pixel phase deviations of all valid corresponding pixels according to the original row and column indices to generate a difference map with the same spatial size as the second boundary result.

[0162] As Figure 3 shown, it shows the spatial comparison relationship between the first boundary result, the second boundary result and the corrected boundary result obtained in Embodiment 1 in the same coordinate space. The red line in the figure represents the first boundary result generated based on the waterlogged period remote sensing image, the blue line is the original second boundary result extracted based on the non-waterlogged period remote sensing image, and the green line is the corrected second boundary result according to the pixel phase deviation difference map. It can be observed that there is an obvious spatial offset between the original second boundary and the first boundary, and the boundary consistency is poor; while the corrected boundary basically fits the first boundary, realizing enhanced spatial consistency. This figure verifies the effectiveness of the boundary correction of the present invention through the complex phase difference matrix and lays a foundation for the subsequent construction of plot units.

[0163] It should be noted that in this embodiment, the pixel complex phase difference and the pixel phase deviation are not the same concept. The pixel complex phase difference refers to the complex phase angle calculated by the arctangent function with the improved water index values of two adjacent remote sensing images as the real part and the imaginary part of the complex number respectively, which is used to characterize the relative change direction and trend of the waterlogged state in the time change process, while the pixel phase deviation is a scalar value obtained by calculating the absolute difference amplitude of the complex phase difference values of the corresponding pixels in the waterlogged period and the non-waterlogged period based on the obtained pixel complex phase difference matrix, which is used to quantify the severity of the water body state change at the corresponding position. Therefore, the pixel complex phase difference emphasizes the angular information of the change direction and belongs to data with vector properties, while the pixel phase deviation emphasizes the magnitude information of the change amplitude and belongs to a non-directional scalar value. Those skilled in the art can understand that the two are respectively used to describe different characteristics of waterlogged changes and play different functional roles in the subsequent generation of the difference map and boundary correction.

[0164] As Figure 4The figure shows a difference map composed of pixel phase deviations calculated based on images of the waterlogging period and the non-waterlogging period. The spatial distribution of the image is consistent with the original second boundary result. The pixel color in the figure represents the difference amplitude of the complex phase angle of the pixel as it changes with time. The brighter the color, the more drastic the change. The difference map is calculated by the complex phase difference matrix and can be used to characterize the significant difference in phase angle between the corresponding pixels of the waterlogging period and the non-waterlogging period. It is an important reference for the subsequent second boundary correction. The map can effectively identify the boundary drift area and the strong deformation area, and assist in completing the boundary coherence correction.

[0165] In this embodiment, when the first boundary result is resampled according to the spatial grid of the second boundary result, the nearest neighbor interpolation principle is adopted. This is based on the consideration that the boundary result itself has discrete category characteristics. The nearest neighbor interpolation can directly retain the category attributes of the original boundary pixels without introducing continuous changes in pixel values, avoiding category ambiguity caused by bilinear interpolation or cubic convolution interpolation, thereby ensuring the clarity of the boundary position after resampling and the integrity of the boundary morphology. Those skilled in the art know that for discrete label data (such as boundary masks, segmentation label maps), nearest neighbor interpolation is a conventional and preferred resampling method.

[0166] When writing the complex phase difference values ​​of valid corresponding pixels into the complex phase difference matrix, writing in spatial order means arranging them in the original row and column index order of the remote sensing image, keeping the complex phase difference matrix completely consistent with the original image in spatial distribution. This processing can ensure that when subsequent spatial analysis is performed based on the complex phase difference matrix (such as differential map generation, boundary correction), there is no need for additional position mapping or spatial relocation operations, which simplifies the processing flow and ensures the traceability and correctness of the data in spatial consistency.

[0167] In the process of generating the differential image, the pixel phase deviations of all valid corresponding pixels are arranged according to the original row and column indexes to ensure that the spatial grid of the differential image strictly corresponds to the second boundary result, so that each pixel position in the differential image maintains a one-to-one correspondence with the corresponding ground feature area in the second boundary result. In this way, the second boundary result can be located, screened and corrected directly on the differential image without the need for additional spatial alignment or pixel position mapping, thereby ensuring the efficiency and consistency of the processing flow. Those skilled in the art can understand that maintaining the original row and column index arrangement is a necessary and routine operational requirement when performing pixel-level difference analysis and consistency correction.

[0168] In this embodiment, the method for correcting the second boundary result in the difference map according to the boundary continuity criterion may be as follows: First, traverse the difference map pixel by pixel, and mark the pixels with pixel phase deviation values less than the preset coherence determination threshold as the coherent candidate pixel set. Here, the coherence determination threshold is a fixed parameter used to screen areas with relatively small boundary change amplitudes. Then, extract the broken boundary segments in the second boundary result that are inside the coherent candidate pixel set, and connect and complete the adjacent coherent candidate pixels to form a continuous boundary path. Preferentially complete it according to the shortest path or local connection principle to avoid large-area expansion. Finally, retain the corrected connected main boundary region, eliminate the fragment regions with areas smaller than the preset area threshold, and output the coherent boundary result as the cultivated land plot unit in the subsequent evaluation of the quality grade of cultivated land resources.

[0169] S40: Construct a cultivated land plot unit based on the coherent boundary results of different time phases, input the cultivated land plot unit into the trained random forest model to obtain a cultivated land quality score, and divide the cultivated land quality grade according to the cultivated land quality score and the preset quality threshold.

[0170] It should be noted that in this embodiment, the method for constructing a cultivated land plot unit based on the coherent boundary results of different time phases may be to superimpose the coherent boundary results of each time phase according to the spatial coordinates. For the overlapping regions, perform a union operation on the boundary lines to form a unified boundary set. In the unified boundary set, for any pair of spatially adjacent closed polygons, calculate the length of their common boundary and calculate the ratio with the total side length of the smaller polygon among the two to obtain the shared side length ratio. When the shared side length ratio is not less than the preset threshold (for example, 50%), determine that the two polygons belong to the same cultivated land plot and perform a topological merging operation; the rest remain independent; subsequently, assign a unique plot number to each merged closed polygon and record the corresponding pixel range of the number in each time phase, and finally form a cultivated land plot unit with spatial consistency and temporal traceability.

[0171] As Figure 5 shown, the upper row of images shows the situation where the uncorrected second boundary result cannot form a closed patch, and the lower row of images is a comparison diagram of successfully constructing a cultivated land plot unit using the coherent boundary result. It can be seen that there are many broken segments in the second boundary before correction, resulting in the inability to form a complete closed polygon and unable to participate in the subsequent scoring as an evaluation unit; while the corrected boundary has improved structural closure and successfully formed a connected patch. This figure verifies the feasibility of patch merging and plot numbering based on the coherent boundary result, meets the construction requirements of the "cultivated land plot unit with spatial consistency and temporal traceability" proposed by the present invention, and is an important basis for subsequent scoring and grade division.

[0172] The training method of the random forest model can be as follows: using the historical cultivated land plot units as the input data of the random forest model, and the historical cultivated land quality scores as the output data of the random forest model. The bootstrap sampling method is used to randomly extract feature subsets and sample subsets to construct several decision trees. The Gini coefficient minimization criterion is used for node division. After each decision tree is generated, the out-of-bag error evaluation is performed on the remaining samples that did not participate in the training of this tree to optimize the number of decision trees and the maximum depth. Finally, the random forest model with the minimum out-of-bag error is saved and used for subsequent cultivated land quality score prediction.

[0173] In the random forest model, the input features of each cultivated land plot unit are composed of three types of indicators: the vegetation index, soil moisture index, and terrain factor calculated for the plot at each time phase. When a plot unit maintains a relatively high vegetation index, a soil moisture index in the appropriate range, and a terrain factor indicating a gentle slope for most time phases, the decision trees integrated in the model will vote multiple times in favor of the high-quality label, and then output a higher cultivated land quality score. Conversely, when the vegetation index is consistently low or fluctuates significantly, the soil moisture index is too dry or too wet, and the terrain factor indicates a large slope, the model will give a lower cultivated land quality score after comprehensively splitting the node statistics. Subsequently, the system will then divide the high-score plots into high-quality grades and the low-score plots into low-quality grades according to the preset quality threshold.

[0174] In this embodiment, by performing spectral residual normalization processing on multi-temporal remote sensing images to generate a sequence of corrected images, the reflectance differences caused by the wet-dry alternation of water bodies are effectively suppressed; the first boundary result and the second boundary result are respectively extracted by using depth segmentation and progressive edge detection, enhancing the adaptability and robustness of boundary extraction; further, a difference map is generated based on the complex phase difference matrix and the second boundary is corrected to obtain a coherent boundary result, significantly improving the boundary temporal consistency; finally, cultivated land plot units are constructed based on the coherent boundary and input into the random forest model to complete scoring and grading, ensuring the stability and traceability of the cultivated land quality assessment results in multi-temporal scenarios, thus achieving the technical goal of stably extracting boundaries and accurately completing grading in dynamic change scenarios.

[0175] Embodiment 2

[0176] Please refer to Figure 6 As shown, based on the same inventive concept, this embodiment discloses and provides a cultivated land resource quality grade evaluation and management system. For the details not described in this embodiment, please refer to the relevant parts in Embodiment 1. The system includes:

[0177] The first processing module: used to obtain multi-temporal remote sensing images of the target area, perform spectral residual normalization processing on the multi-temporal remote sensing images to generate a sequence of corrected images, and based on the sequence of corrected images, perform continuous marking processing to generate a set of remote sensing images during the water accumulation period and a set of remote sensing images during the non-water accumulation period;

[0178] The second processing module: is used to perform depth segmentation on the remote sensing images during the waterlogging period to obtain the first boundary result, perform regional texture adaptive filtering on the remote sensing images during the non-waterlogging period, and then perform edge detection progressively according to the first scale and the second scale to obtain the second boundary result, where the first scale is greater than the second scale;

[0179] The correction module: is used to reproject the first boundary result into the coordinate system of the second boundary result to obtain the complex phase difference matrix of the corresponding pixels during the waterlogging period and the non-waterlogging period, calculate the pixel phase deviation based on the complex phase difference matrix, generate a difference map, and correct the second boundary result in the difference map according to the boundary continuity criterion to obtain the coherent boundary result;

[0180] The evaluation module: is used to construct cultivated land plot units based on the coherent boundary results at different times, input the cultivated land plot units into the trained random forest model to obtain the cultivated land quality score, and divide the cultivated land quality grades according to the cultivated land quality score and the preset quality threshold.

[0181] The detailed description set forth above in conjunction with the accompanying drawings describes examples and does not represent all examples that can be implemented or fall within the scope of the claims. The terms "example" and "exemplary" when used in this specification mean "serving as an example, instance, or illustration" and do not mean "superior to or better than other examples".

[0182] The phrase "one embodiment" or "an embodiment" recited throughout this specification means that a particular feature, structure, or characteristic described in connection with that embodiment is included in at least one embodiment of the present invention. Therefore, the use of these phrases may refer to more than just one embodiment. In addition, the described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0183] It should also be noted that these embodiments may be described as processes depicted as flowcharts, structure diagrams, or block diagrams. Although a flowchart may describe the operations as a sequential process, many of these operations can be performed in parallel or concurrently. In addition, the order of these operations may be rearranged.

Claims

1. A method for evaluating and managing the quality grade of cultivated land resources, characterized in that, Including: Obtain multi-temporal remote sensing images of the target area, perform spectral residual normalization processing on the multi-temporal remote sensing images to generate a sequence of corrected images, and based on the sequence of corrected images, perform continuity marking processing to generate a set of remote sensing images during the water accumulation period and a set of remote sensing images during the non-water accumulation period; Perform depth segmentation on the remote sensing images during the water accumulation period to obtain a first boundary result. After performing regional texture adaptive filtering on the remote sensing images during the non-water accumulation period, perform edge detection progressively at a first scale and a second scale to obtain a second boundary result, where the first scale is greater than the second scale; Reproject the first boundary result into the coordinate system of the second boundary result to obtain a complex phase difference matrix of corresponding pixels during the water accumulation period and the non-water accumulation period. Calculate the pixel phase deviation based on the complex phase difference matrix to generate a difference map, and correct the second boundary result in the difference map according to the boundary continuity criterion to obtain a coherent boundary result; Construct cultivated land plot units based on the coherent boundary results of different time phases, input the cultivated land plot units into a trained random forest model to obtain cultivated land quality scores, and divide the cultivated land quality grades according to the cultivated land quality scores and a preset quality threshold.

2. The quality grade evaluation and management method of cultivated land resources according to claim 1, wherein The method for performing spectral residual normalization processing on the multi-temporal remote sensing images to generate a sequence of corrected images includes: Based on each multi-temporal remote sensing image, select a first reference pixel set and a second reference pixel set, and calculate the spectral residual vectors of each band relative to the two types of reference pixel sets. The multi-temporal remote sensing images include Q adjacent multi-temporal remote sensing images; Based on the spectral residual vectors, determine the normalization adjustment parameters of each band, and combine the normalization adjustment parameters to obtain a correction parameter matrix; Perform radiometric normalization on the correction parameter matrix to obtain a normalized image; Taking the correction parameter matrix as input, calculate the spectral residual differences of adjacent multi-temporal remote sensing images in chronological order to obtain a temporal difference matrix; Perform interpolation diffusion operation based on the temporal difference matrix to generate a drift compensation field, and superimpose the drift compensation field on the normalized image according to a preset smoothing constraint to obtain a sequence of corrected images.

3. The method for evaluating and managing the quality grade of cultivated land resources according to claim 2, wherein The method for performing continuity marking processing based on the sequence of corrected images to generate a set of remote sensing images during the water accumulation period and a set of remote sensing images during the non-water accumulation period includes: In the sequence of corrected images, based on the corresponding pixels of adjacent multi-temporal remote sensing images, calculate their complex improved water body index respectively, and generate a sequence of phase state label maps according to the improved water body index; Perform a sliding window voting operation on the sequence of phase state label maps along the time dimension to obtain a continuity marking matrix; Based on the continuity marking matrix, count the number of water body continuous pixels in each multi-temporal remote sensing image, and calculate the pixel ratio of the number of water body continuous pixels to the total number of pixels in the image; When the pixel ratio is not less than a preset continuous ratio threshold, classify the multi-temporal remote sensing image into the set of remote sensing images during the water accumulation period, otherwise, classify the multi-temporal remote sensing image into the set of remote sensing images during the non-water accumulation period.

4. The cultivated land resource quality grade evaluation and management method according to claim 2, characterized in that The method for selecting a first reference pixel set and a second reference pixel set based on each multi-temporal remote sensing image includes: Calculate the temporal variation coefficient of the multi-band brightness pixel by pixel for Q adjacent multi-temporal remote sensing images and generate a spectral stability grid; In the spectral stability grid, pixels with a coefficient of variation of time not higher than a preset variation threshold are selected to construct a stable candidate set; Perform hierarchical sampling on the stable candidate set, specifically including: Sample from pixels where the brightness in both the red band and the short-wave infrared band is below the preset low brightness percentile threshold, and combine morphological opening operation to remove isolated pixels to obtain a first reference pixel set; Sample from pixels where the brightness in both the green band and the near-infrared band is above the preset high brightness percentile threshold to obtain a second reference pixel candidate area; In the second reference pixel candidate area, adopt a spatial grid uniform sampling strategy to extract pixels, and perform spectral homogeneity test on the 3×3 neighborhood centered on the grid center of each extracted pixel. The pixels passing the test are used to construct a second reference pixel set.

5. The cultivated land resource quality grade evaluation and management method according to claim 4, characterized in that The method for generating a sequence of phase state label maps based on the improved water index includes: Based on each temporal remote sensing image, extract the multi-band normalized brightness value for water body identification, and calculate the improved water index according to the preset weighted combination coefficient and the normalized brightness value; For the improved water indices of each pair of adjacent temporal remote sensing images, extract the index values of the corresponding pixels respectively. Use the index value of the previous temporal remote sensing image as the real part of the complex number, and the index value of the latter temporal remote sensing image as the imaginary part of the complex number to generate a complex number expression matrix; For each complex number in the complex number expression matrix, calculate the complex number phase angle, which is obtained by calculating the arctangent value of the real part and the imaginary part of the complex number; Arrange the complex number phase angles of all pixels in the original spatial position to generate a pixel complex phase difference; When the absolute value of the pixel complex phase difference is less than the preset phase change threshold, it is marked as a non-ponding state, otherwise it is marked as a ponding state; Arrange the phase state labels of all pixels in the original spatial position to form a phase state label map with the same size as the original image; Process each pair of adjacent temporal remote sensing images in sequence to generate a sequence of phase state label maps.

6. The method for evaluating and managing the quality grade of cultivated land resources according to claim 3, wherein The method for performing depth segmentation on the ponding period remote sensing image to obtain a first boundary result includes: Perform superpixel segmentation on the ponding period remote sensing image according to a preset segmentation scale to generate an initial region segmentation map, where the initial region segmentation map includes K superpixel regions; Obtain the normalized brightness mean, local texture directionality index, and shape compactness feature of each superpixel region, and splice the normalized brightness mean, local texture directionality index, and shape compactness feature in a preset order to construct a region-level feature vector; Take the region-level feature vector as the input, perform region-level classification through a preset convolutional neural network model, and output the water body prediction map of each superpixel region; Perform connected region analysis on the water body prediction map, screen out the largest connected water body region, and perform region edge sharpness enhancement processing on the connected water body region to obtain a preliminary boundary result; Perform edge correction processing on the preliminary boundary result to obtain a first boundary result.

7. The method for evaluating and managing the quality grade of cultivated land resources according to claim 6, characterized in that, Divide the non-ponding period remote sensing image into H sliding window regions of a fixed size, and perform region texture adaptive filtering processing within each sliding window region to obtain a filtered image; Set the first edge detection operator based on the first scale, perform the first round of edge detection on the filtered image according to the first edge detection operator, extract the main edge contours with coarse granularity, and generate the first-round edge map; Based on the positions in the first-round edge map where the edge response values are not zero, construct a local mask region, set the second edge detection operator based on the second scale, perform the second round of edge detection on the local mask region according to the second edge detection operator, extract the supplementary edges with fine granularity, and generate the second-round edge map; Perform edge merging processing on the first-round edge map and the second-round edge map, preferentially retain the main contours of the first-round edges, and introduce the second-round edge supplements at the positions where the edges are broken to form an edge fusion map; Perform connectivity screening processing on the edge fusion map to generate the second boundary result.

8. The cultivated land resource quality grade evaluation and management method according to claim 7, characterized in that The method of reprojection of the first boundary result into the coordinate system of the second boundary result to obtain the complex phase difference matrix of the corresponding pixels in the water accumulation period and the non-water accumulation period includes: Resample the first boundary result according to the spatial grid of the second boundary result to make the pixel row and column coordinates of the first boundary result consistent with those of the second boundary result, where the resampling process is based on the nearest neighbor interpolation principle; Based on the resampled first boundary result and the second boundary result, establish the corresponding pixel relationship between the water accumulation period and the non-water accumulation period according to the pixel row and column indices, and screen out the invalid pixel pairs to obtain the valid corresponding pixels, where the invalid pixel pairs include the pixels outside the boundary and the abnormally missing pixels; Based on each pair of valid corresponding pixels, extract their corresponding complex phase difference values and write them into the complex phase difference matrix in spatial order.

9. The arable land resource quality grade evaluation and management method according to claim 8, characterized in that The method of calculating the pixel phase deviation based on the complex phase difference matrix to generate the difference map includes: Traverse the complex phase difference matrix pixel by pixel to obtain the complex phase difference values of each valid corresponding pixel; Based on each pair of valid corresponding pixels, calculate the absolute value of the complex phase difference between the pixels in the water accumulation period and the non-water accumulation period to obtain the pixel phase deviation, where the pixel phase deviation is used to characterize the change amplitude of the corresponding pixels in the water accumulation period and the non-water accumulation period in the complex phase angle; Arrange the pixel phase deviations of all valid corresponding pixels according to the original row and column indices to generate a difference map with the same spatial size as the second boundary result.

10. A cultivated land resource quality grade evaluation and management system, which is used to implement the cultivated land resource quality grade evaluation and management method described in any one of claims 1-9, and is characterized in that, Includes: The first processing module: used to obtain the multi-temporal remote sensing images of the target area, perform spectral residual normalization processing on the multi-temporal remote sensing images to generate a sequence of corrected images, and based on the sequence of corrected images, perform continuity marking processing to generate a set of remote sensing images in the water accumulation period and a set of remote sensing images in the non-water accumulation period; The second processing module: used to perform depth segmentation on the remote sensing images in the water accumulation period to obtain the first boundary result, perform regional texture adaptive filtering on the remote sensing images in the non-water accumulation period, and then perform edge detection progressively according to the first scale and the second scale to obtain the second boundary result, where the first scale is greater than the second scale; The correction module: used to reproject the first boundary result into the coordinate system of the second boundary result to obtain the complex phase difference matrix of the corresponding pixels in the water accumulation period and the non-water accumulation period, calculate the pixel phase deviation based on the complex phase difference matrix to generate a difference map, and correct the second boundary result in the difference map according to the boundary continuity criterion to obtain a coherent boundary result; Evaluation module: It is used to construct cultivated land plot units based on the coherent boundary results of different time phases, input the cultivated land plot units into a trained random forest model to obtain cultivated land quality scores, and divide the cultivated land quality grades according to the cultivated land quality scores and a preset quality threshold.

Images