Construction method of forestry vegetation and non-vegetation identification index based on near-infrared spectrum

Abstract

The application discloses a forestry vegetation and non-vegetation identification index construction method based on near-infrared spectroscopy, and relates to the technical field of forestry resource monitoring. The method comprises the following steps: S1, collecting near-infrared hyperspectral images of a forestry monitoring area, and splitting the near-infrared hyperspectral images into a plurality of characteristic wavelength images; S2, pre-processing the plurality of characteristic wavelength images to obtain a plurality of weighted wavelength images; S3, splicing the weighted wavelength images, and generating a plurality of texture feature maps according to a feature matrix of the spliced double-band images; and S4, inputting the plurality of weighted wavelength images and the plurality of texture feature maps into a vegetation identification model to obtain an identification result. The application performs double-channel and multi-scale convolution branch processing on an image set input into a vegetation identification network, covers large-scale tree crown clusters, medium-scale single trees and small-scale edge mixed pixels, is suitable for complex forestry scenes, and significantly improves the identification precision and effectively reduces the misjudgment rate.

Description

Technical Field

[0001] This invention relates to the field of forestry resource monitoring technology, specifically to a method for constructing forestry vegetation and non-vegetation identification indices based on near-infrared spectroscopy. Background Technology

[0002] Accurate identification of forest vegetation and non-vegetation is fundamental to forest resource surveys, ecosystem monitoring, and pest and disease early warning. Traditional identification methods rely on manual field surveys, which are inefficient, costly, and have limited coverage, making them unsuitable for large-scale forestry monitoring.

[0003] Spectral recognition technology, with its advantages of speed, non-contact operation, and wide-range detection, has been widely applied in the field of vegetation identification. The near-infrared spectral region contains spectral characteristics unique to vegetation, while non-vegetation (such as soil, rocks, dead wood, construction debris, etc.) has no spectral characteristics, and its near-infrared spectral reflectance often shows monotonic changes or irregular fluctuations with increasing wavelength.

[0004] Existing vegetation identification indices based on near-infrared spectroscopy (such as Normalized Difference Vegetation Index (NDVI) and Enhanced Vegetation Index (EVI)) mostly rely on combinations of red and near-infrared single bands, failing to fully utilize the full-band information of the near-infrared spectrum. They construct indices solely based on the difference / ratio between a single near-infrared band and the red band, making it difficult to distinguish non-vegetation species with similar spectral characteristics (such as deadwood and weakened vegetation). Furthermore, they have poor adaptability to complex forestry scenarios (such as understory shadows, mixed pixels, and different forest stand types), making them prone to misjudgment.

[0005] Therefore, how to fully explore the multi-band characteristics of the near-infrared spectrum and construct a forest vegetation and non-vegetation identification index method with strong adaptability, high identification accuracy and strong anti-interference ability has become a technical problem that needs to be solved in the field of forestry spectral monitoring. Summary of the Invention

[0006] To address the above problems, this invention proposes a method for constructing identification indices for forest vegetation and non-vegetation based on near-infrared spectroscopy.

[0007] The technical solution of this invention is: a method for constructing forest vegetation and non-vegetation identification indices based on near-infrared spectroscopy, comprising the following steps:

[0008] S1. Collect near-infrared hyperspectral images of the forestry monitoring area and break them down into several characteristic wavelength images;

[0009] S2. Preprocess several characteristic wavelength images to obtain several weighted wavelength images;

[0010] S3. Stitch together the images of each weighted wavelength, and generate several texture feature maps based on the feature matrix of the stitched dual-band images.

[0011] S4. Input several weighted wavelength images and several texture feature maps into the vegetation recognition model to obtain the recognition results.

[0012] Furthermore, in S1, near-infrared hyperspectral images of the forestry monitoring area are collected, atmospheric correction is performed on the near-infrared hyperspectral images, and single-wavelength images corresponding to the bands closest to the wavelengths of 850nm, 1200nm and 1900nm are extracted and used as the first characteristic wavelength image, the second characteristic wavelength image and the third characteristic wavelength image, respectively.

[0013] The data format of near-infrared hyperspectral images is height × width × number of bands, for example, 300 (height) × 200 (width) × 150 (number of bands). The near-infrared hyperspectral image is broken down into 150 independent two-dimensional images (each a 300×200 pixel matrix), meaning each image is a single-wavelength image corresponding to a band index. These 150 bands correspond to 150 near-infrared wavelengths (e.g., from 780nm to 2500nm, with each band representing 11.4nm). Then, the single-wavelength images closest to 850nm, 1200nm, and 1900nm are extracted.

[0014] The characteristic wavelength image is actually three single-band images corresponding to the near-infrared short-wave peak, the near-infrared mid-wave absorption valley, and the near-infrared long-wave absorption valley:

[0015] Channel 1: 850nm (near-infrared short-wave strong reflection peak). Due to internal scattering of leaves, vegetation has extremely high reflectivity, while non-vegetation (soil / dead wood / rock) has low and stable reflectivity, which is used to distinguish between vegetation (high reflectivity) and non-vegetation (low reflectivity).

[0016] Channel 2: 1200nm (near-infrared mid-wave moisture absorption valley). Due to moisture absorption, the reflectivity of vegetation drops sharply, while the reflectivity of non-vegetation (soil / dead wood / rock) remains stable without decreasing. It is used to distinguish between living vegetation (with absorption valley) and dead wood (without absorption valley).

[0017] Channel 3: 1900nm (near-infrared long-wave strong water absorption valley), strong water absorption, vegetation has extremely low reflectivity, non-vegetation (soil / dead wood / rock) has no obvious valley in reflectivity, used to enhance differentiation and exclude special non-vegetation with high reflectivity.

[0018] If the collected near-infrared hyperspectral data does not cover 1900nm (for example, only up to 1600nm), it can be replaced with 1600nm (another moisture absorption valley in the near-infrared long wavelength).

[0019] Furthermore, S2 includes the following sub-steps:

[0020] S21. Use a Gaussian filter to denoise the first characteristic wavelength image, the second characteristic wavelength image and the third characteristic wavelength image respectively to obtain the first smoothed wavelength image, the second smoothed wavelength image and the third smoothed wavelength image.

[0021] S22. The first smoothed wavelength image, the second smoothed wavelength image, and the third smoothed wavelength image are processed using spatial attention, and the processing results are multiplied element-wise with the corresponding smoothed wavelength images to obtain the first weighted wavelength image, the second weighted wavelength image, and the third weighted wavelength image.

[0022] The beneficial effects of the above-mentioned further solutions are: In this invention, spatial attention weighting is applied to the Gaussian smoothed image to improve the contrast of vegetation areas in the weighted wavelength image and weaken the interference of non-vegetation areas.

[0023] Furthermore, S3 includes the following sub-steps:

[0024] S31. The first weighted wavelength image and the second weighted wavelength image are stitched together by channel to obtain the first dual-band image;

[0025] S32. The second weighted wavelength image and the third weighted wavelength image are stitched together by channel to obtain the second dual-band image;

[0026] S33. Generate three texture feature maps for the first dual-band image and the second dual-band image respectively.

[0027] The beneficial effects of the above-mentioned further solutions are as follows: In this invention, images of 850nm (strong reflection from vegetation) and 1200nm (absorption valley from vegetation) are stitched together as a basis, and images of 1200nm (weak absorption valley) and 1900nm (strong absorption valley) are stitched together as a basis, preserving the spatial topology of the original images and allowing each pixel to carry spectral information of two bands.

[0028] Furthermore, S33 includes the following sub-steps:

[0029] S331. Select a local window for the pixels of the dual-band image, and extract the gray-level co-occurrence matrix corresponding to the first channel and the gray-level co-occurrence matrix corresponding to the second channel of the pixels within the local window.

[0030] S332. Calculate the sum of the rank of the gray-level co-occurrence matrix corresponding to the first channel and the rank of the gray-level co-occurrence matrix corresponding to the second channel. Use the proportion of the rank of the gray-level co-occurrence matrix corresponding to the first channel to the total as its weight, and use the proportion of the rank of the gray-level co-occurrence matrix corresponding to the second channel to the total as its weight. Perform weighted fusion of the two gray-level co-occurrence matrices to obtain the fused gray-level co-occurrence matrix of the pixel.

[0031] S333: Assign the contrast, entropy and correlation values ​​corresponding to the fused gray-level co-occurrence matrix to the pixels respectively;

[0032] S334. Repeat S331-S333 until all pixels have been traversed, and obtain the three texture feature maps corresponding to the dual-band image.

[0033] The beneficial effects of the above-mentioned further scheme are as follows: In this invention, the first weighted wavelength image and the second weighted wavelength image are stitched together by channel to obtain a complete dual-band image, where each pixel contains values ​​from two channels. Vegetation has a uniform texture, resulting in a concentrated distribution of GLCM elements and a lower rank; non-vegetation has a chaotic texture, leading to a dispersed distribution of GLCM elements and a higher rank. Using the proportion of the rank of a single-channel GLCM to the total rank as the weight, the GLCM rank is more stable and thus carries a higher weight in the fusion process.

[0034] For vegetated areas, the reflectance of the two bands (high at 850, low at 1200) follows a stable pattern, resulting in lower texture features (more uniform texture) in dual-band GLCM. For non-vegetated areas, the reflectance of the two bands does not follow a fixed pattern (e.g., soil has moderate reflectance at both 850 and 1200), leading to more chaotic grayscale combinations within local windows. This results in higher texture features and more disordered texture in dual-band GLCM. Therefore, this symbiotic relationship of dual-band grayscale combinations better reflects the essential differences between vegetation and non-vegetated areas than single-band GLCM.

[0035] Centered on the pixel in row 3, column 3, select the surrounding pixels to form a 5×5 region. Based on the grayscale values ​​of these 25 pixels, calculate the GLCM matrix of this window. Based on this GLCM matrix, calculate the contrast a1, entropy a2, and correlation a3, and assign these three values ​​to the pixel at the center of the window in row 3, column 3. Next, move the window 1 pixel to the right, with the new window centered on the pixel in row 3, column 4. Calculate the GLCM and three feature values ​​for this new window and assign them to the pixel in row 3, column 4. Continue moving the window (right, down) until all pixels in the image have been treated as window centers. Finally, each pixel has corresponding contrast, entropy, and correlation values.

[0036] Furthermore, in S4, the vegetation recognition model includes an input layer, an initial convolutional block, dual attention channels, a first convolutional layer, a second convolutional layer, a third convolutional layer, a splicing layer, an aggregation layer, and a classification layer;

[0037] The input layer, initial convolutional block, and dual attention channels are connected sequentially, and then connected to the first, second, and third convolutional layers respectively; the first, second, and third convolutional layers are all connected to the splicing layer; the splicing layer, aggregation layer, and classification layer are connected sequentially.

[0038] The beneficial effects of the above-mentioned further scheme are as follows: In this invention, the initial convolutional block fuses spectral and texture information from 9 channels and eliminates dimensional differences. Dual attention channels perform average pooling and convolution operations on the output of the initial convolutional block to generate a weight mask, and then element-wise weight the weight mask with the output of the initial convolutional block.

[0039] The first convolutional layer uses convolutional kernels to extract large-scale global features from the output of the dual attention channels, suitable for the spatial distribution of forest canopy clusters. Then, an upsampling layer uses bilinear interpolation to restore the large-scale global features to the original image size. The second convolutional layer uses convolutional kernels to extract mesoscale global features from the output of the dual attention channels, suitable for the canopy structure of individual trees in forestry. The third convolutional layer uses convolutional kernels to extract small-scale global features from the output of the dual attention channels, suitable for areas at the edge of the tree canopy or where vegetation and soil are mixed. The stitching layer stitches the large-scale, mesoscale, and small-scale global features along the channel dimension to obtain the output of the stitching layer. , This represents the output after passing through the first convolutional layer and the upsampling layer. This represents the output of the second convolutional layer. This represents the output of the third convolutional layer. After the aggregation layer reduces the dimensionality and removes redundancy from the output of the splicing layer, it enters the classification layer. The classification layer uses the sigmoid function to determine the vegetation cover rate, outputting a value between 0 and 1, which is then converted into a percentage.

[0040] Furthermore, the expression for the initial convolutional block is:

[0041] ;

[0042] in, This represents the output of the initial convolutional block. This indicates a batch normalization operation. This indicates an activation operation. This represents the convolution operation. This represents the output of the input layer.

[0043] Furthermore, the expression for the dual attention channel is:

[0044] ;

[0045] ;

[0046] ;

[0047] in, This represents the output of the initial convolutional block. This represents the Hada code product operation. Indicates the first weight. Indicates the second weight. Indicates global average pooling. Indicates average pooling. This represents the activation function. This represents the output of the dual attention channel. This indicates a fully connected layer.

[0048] The beneficial effects of this invention are:

[0049] (1) The present invention preprocesses the images of the bands with the most significant near-infrared spectral response of vegetation, preserves the spectral differences between vegetation and non-vegetation, and performs channel splicing and gray-level co-occurrence matrix fusion processing to preserve the spatial topology of forest land cover and the uniform texture of vegetation.

[0050] (2) The present invention performs dual-channel and multi-scale convolutional branching on the image set of the input vegetation recognition network, covering large-scale tree canopy clusters, medium-scale single trees and small-scale edge mixed pixels, which is suitable for complex forestry scenarios, significantly improving recognition accuracy and effectively reducing the misjudgment rate. Attached Figure Description

[0051] Figure 1 This is a flowchart of a method for constructing forest vegetation and non-vegetation identification indices based on near-infrared spectroscopy.

[0052] Figure 2 This is a schematic diagram of the vegetation identification model. Detailed Implementation

[0053] The embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0054] like Figure 1 As shown, this invention provides a method for constructing forest vegetation and non-vegetation identification indices based on near-infrared spectroscopy, comprising the following steps:

[0055] S1. Collect near-infrared hyperspectral images of the forestry monitoring area and break them down into several characteristic wavelength images;

[0056] S2. Preprocess several characteristic wavelength images to obtain several weighted wavelength images;

[0057] S3. Stitch together the images of each weighted wavelength, and generate several texture feature maps based on the feature matrix of the stitched dual-band images.

[0058] S4. Input several weighted wavelength images and several texture feature maps into the vegetation recognition model to obtain the recognition results.

[0059] In this embodiment of the invention, in S1, near-infrared hyperspectral images of the forestry monitoring area are collected, atmospheric correction is performed on the near-infrared hyperspectral images, and single-wavelength images corresponding to the bands closest to the wavelengths of 850nm, 1200nm and 1900nm are extracted and used as the first characteristic wavelength image, the second characteristic wavelength image and the third characteristic wavelength image, respectively.

[0060] In this invention, the data format of the near-infrared hyperspectral image is height × width × number of bands, for example, 300 (height) × 200 (width) × 150 (number of bands). The near-infrared hyperspectral image is split into 150 independent two-dimensional images (each a 300 × 200 pixel matrix), that is, each image is a single-wavelength image corresponding to a band index. The 150 bands correspond to 150 near-infrared wavelengths (for example, from 780nm to 2500nm, one band every 11.4nm), and then the single-wavelength images closest to 850nm, 1200nm, and 1900nm are extracted.

[0061] The characteristic wavelength image is actually three single-band images corresponding to the near-infrared short-wave peak, the near-infrared mid-wave absorption valley, and the near-infrared long-wave absorption valley:

[0062] Channel 1: 850nm (near-infrared short-wave strong reflection peak). Due to internal scattering of leaves, vegetation has extremely high reflectivity, while non-vegetation (soil / dead wood / rock) has low and stable reflectivity, which is used to distinguish between vegetation (high reflectivity) and non-vegetation (low reflectivity).

[0063] Channel 2: 1200nm (near-infrared mid-wave moisture absorption valley). Due to moisture absorption, the reflectivity of vegetation drops sharply, while the reflectivity of non-vegetation (soil / dead wood / rock) remains stable without decreasing. It is used to distinguish between living vegetation (with absorption valley) and dead wood (without absorption valley).

[0064] Channel 3: 1900nm (near-infrared long-wave strong water absorption valley), strong water absorption, vegetation has extremely low reflectivity, non-vegetation (soil / dead wood / rock) has no obvious valley in reflectivity, used to enhance differentiation and exclude special non-vegetation with high reflectivity.

[0065] If the collected near-infrared hyperspectral data does not cover 1900nm (for example, only up to 1600nm), it can be replaced with 1600nm (another moisture absorption valley in the near-infrared long wavelength).

[0066] In this embodiment of the invention, S2 includes the following sub-steps:

[0067] S21. Use a Gaussian filter to denoise the first characteristic wavelength image, the second characteristic wavelength image and the third characteristic wavelength image respectively to obtain the first smoothed wavelength image, the second smoothed wavelength image and the third smoothed wavelength image.

[0068] S22. The first smoothed wavelength image, the second smoothed wavelength image, and the third smoothed wavelength image are processed using spatial attention, and the processing results are multiplied element-wise with the corresponding smoothed wavelength images to obtain the first weighted wavelength image, the second weighted wavelength image, and the third weighted wavelength image.

[0069] In this invention, spatial attention weighting is applied to the Gaussian-smoothed image to improve the contrast of vegetation areas in the weighted wavelength image and weaken the interference from non-vegetation areas.

[0070] In this embodiment of the invention, S3 includes the following sub-steps:

[0071] S31. The first weighted wavelength image and the second weighted wavelength image are stitched together by channel to obtain the first dual-band image;

[0072] S32. The second weighted wavelength image and the third weighted wavelength image are stitched together by channel to obtain the second dual-band image;

[0073] S33. Generate three texture feature maps for the first dual-band image and the second dual-band image respectively.

[0074] In this invention, images at 850nm (strong reflection from vegetation) and 1200nm (absorption valley from vegetation) are stitched together, and images at 1200nm (weak absorption valley) and 1900nm (strong absorption valley) are stitched together, preserving the spatial topology of the original images and allowing each pixel to carry spectral information of two bands.

[0075] In this embodiment of the invention, S33 includes the following sub-steps:

[0076] S331. Select a local window for the pixels of the dual-band image, and extract the gray-level co-occurrence matrix corresponding to the first channel and the gray-level co-occurrence matrix corresponding to the second channel of the pixels within the local window.

[0077] S332. Calculate the sum of the rank of the gray-level co-occurrence matrix corresponding to the first channel and the rank of the gray-level co-occurrence matrix corresponding to the second channel. Use the proportion of the rank of the gray-level co-occurrence matrix corresponding to the first channel to the total as its weight, and use the proportion of the rank of the gray-level co-occurrence matrix corresponding to the second channel to the total as its weight. Perform weighted fusion of the two gray-level co-occurrence matrices to obtain the fused gray-level co-occurrence matrix of the pixel.

[0078] S333: Assign the contrast, entropy and correlation values ​​corresponding to the fused gray-level co-occurrence matrix to the pixels respectively;

[0079] S334. Repeat S331-S333 until all pixels have been traversed, and obtain the three texture feature maps corresponding to the dual-band image.

[0080] In this invention, the first weighted wavelength image and the second weighted wavelength image are stitched together by channel to obtain a complete dual-band image, where each pixel contains values ​​from two channels. Vegetation has a uniform texture, resulting in a concentrated distribution of GLCM elements and a lower rank; non-vegetation has a chaotic texture, leading to a dispersed distribution of GLCM elements and a higher rank. The weight is determined by the proportion of the rank of a single-channel GLCM to the total rank; GLCM rank is more stable and therefore carries a higher weight in the fusion process.

[0081] For vegetated areas, the reflectance of the two bands (high at 850, low at 1200) follows a stable pattern, resulting in lower texture features (more uniform texture) in dual-band GLCM. For non-vegetated areas, the reflectance of the two bands does not follow a fixed pattern (e.g., soil has moderate reflectance at both 850 and 1200), leading to more chaotic grayscale combinations within local windows. This results in higher texture features and more disordered texture in dual-band GLCM. Therefore, this symbiotic relationship of dual-band grayscale combinations better reflects the essential differences between vegetation and non-vegetated areas than single-band GLCM.

[0082] Centered on the pixel in row 3, column 3, select the surrounding pixels to form a 5×5 region. Based on the grayscale values ​​of these 25 pixels, calculate the GLCM matrix of this window. Based on this GLCM matrix, calculate the contrast a1, entropy a2, and correlation a3, and assign these three values ​​to the pixel at the center of the window in row 3, column 3. Next, move the window 1 pixel to the right, with the new window centered on the pixel in row 3, column 4. Calculate the GLCM and three feature values ​​for this new window and assign them to the pixel in row 3, column 4. Continue moving the window (right, down) until all pixels in the image have been treated as window centers. Finally, each pixel has corresponding contrast, entropy, and correlation values.

[0083] In this embodiment of the invention, in S4, the vegetation recognition model includes an input layer, an initial convolutional block, dual attention channels, a first convolutional layer, a second convolutional layer, a third convolutional layer, a splicing layer, an aggregation layer, and a classification layer.

[0084] The input layer, initial convolutional block, and dual attention channels are connected sequentially, and then connected to the first, second, and third convolutional layers respectively; the first, second, and third convolutional layers are all connected to the splicing layer; the splicing layer, aggregation layer, and classification layer are connected sequentially.

[0085] In this invention, the input layer is used to input three weighted images. and 6 texture feature maps Its expression is: , , , ,in, Indicates the height of the image. Indicates the width of the image.

[0086] The three weighted images (850nm, 1200nm, and 1900nm) are single-wavelength images extracted from near-infrared hyperspectral images of the same forestry monitoring area, so the H×W of each single-wavelength image is exactly the same.

[0087] The six texture feature maps (GLCM contrast, entropy, and correlation) are calculated using a pixel-by-pixel local window based on the band combination of the three weighted images. The calculation of the local window involves traversing point by point in an H×W space of the same size as the original weighted image. Therefore, the H×W of the generated texture feature map is consistent with that of the original weighted image.

[0088] The initial convolutional block fuses spectral and texture information from 9 channels and eliminates dimensional differences. Dual attention channels perform average pooling and convolution operations on the output of the initial convolutional block to generate a weight mask, which is then element-wise weighted with the output of the initial convolutional block.

[0089] The first convolutional layer uses convolutional kernels to extract large-scale global features from the output of the dual attention channels, suitable for the spatial distribution of forest canopy clusters. Then, an upsampling layer uses bilinear interpolation to restore the large-scale global features to the original image size. The second convolutional layer uses convolutional kernels to extract mesoscale global features from the output of the dual attention channels, suitable for the canopy structure of individual trees in forestry. The third convolutional layer uses convolutional kernels to extract small-scale global features from the output of the dual attention channels, suitable for areas at the edge of the tree canopy or where vegetation and soil are mixed. The stitching layer stitches the large-scale, mesoscale, and small-scale global features along the channel dimension to obtain the output of the stitching layer. , This represents the output after passing through the first convolutional layer and the upsampling layer. This represents the output of the second convolutional layer. This represents the output of the third convolutional layer. After the aggregation layer reduces the dimensionality and removes redundancy from the output of the splicing layer, it enters the classification layer. The classification layer uses the sigmoid function to determine the vegetation cover rate, outputting a value between 0 and 1, which is then converted into a percentage.

[0090] In this embodiment of the invention, the expression for the initial convolutional block is:

[0091] ;

[0092] in, This represents the output of the initial convolutional block. This indicates a batch normalization operation. This indicates an activation operation. This represents the convolution operation. This represents the output of the input layer.

[0093] In this embodiment of the invention, the expression for the dual attention channel is:

[0094] ;

[0095] ;

[0096] ;

[0097] in, This represents the output of the initial convolutional block. This represents the Hada code product operation. Indicates the first weight. Indicates the second weight. Indicates global average pooling. Indicates average pooling. This represents the activation function. This represents the output of the dual attention channel. This indicates a fully connected layer.

[0098] In this embodiment of the invention, a near-infrared hyperspectral image is acquired using a hyperspectral camera. The spectral range is 780-2500 nm, with 150 bands, a spectral resolution of 11.4 nm, and a spatial resolution of 0.5 m. The hyperspectral image of the monitored area is acquired and cropped to a specified pixel size. If the hyperspectral data only extends to 1600 nm, it can be replaced with the 1600 nm band.

[0099] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A method for constructing forest vegetation and non-vegetation identification indices based on near-infrared spectroscopy, characterized in that, Includes the following steps: S1. Collect near-infrared hyperspectral images of the forestry monitoring area and split them into three characteristic wavelength images; S2. Preprocess the three characteristic wavelength images to obtain three weighted wavelength images; S3. Stitch together the images of each weighted wavelength, and generate three texture feature maps based on the feature matrix of the stitched dual-band images; S4. Input the three weighted wavelength images and three texture feature maps into the vegetation recognition model to obtain the recognition result; the recognition result is the vegetation coverage rate, outputting a value of 0-1, and converting it into a percentage form; S3 includes the following sub-steps: S31. The first weighted wavelength image and the second weighted wavelength image are stitched together by channel to obtain the first dual-band image; S32. The second weighted wavelength image and the third weighted wavelength image are stitched together by channel to obtain the second dual-band image; S33. Generate three texture feature maps for the first dual-band image and the second dual-band image respectively; S33 includes the following sub-steps: S331. Select a local window for the pixels of the dual-band image, and extract the gray-level co-occurrence matrix corresponding to the first channel and the gray-level co-occurrence matrix corresponding to the second channel of the pixels within the local window. S332. Calculate the sum of the rank of the gray-level co-occurrence matrix corresponding to the first channel and the rank of the gray-level co-occurrence matrix corresponding to the second channel. Use the proportion of the rank of the gray-level co-occurrence matrix corresponding to the first channel to the total as its weight, and use the proportion of the rank of the gray-level co-occurrence matrix corresponding to the second channel to the total as its weight. Perform weighted fusion of the two gray-level co-occurrence matrices to obtain the fused gray-level co-occurrence matrix of the pixel. S333: Assign the contrast, entropy and correlation values ​​corresponding to the fused gray-level co-occurrence matrix to the pixels respectively; S334. Repeat S331-S333 until all pixels have been traversed, and obtain the three texture feature maps corresponding to the dual-band image.

2. The method for constructing forestry vegetation and non-vegetation identification indices based on near-infrared spectroscopy according to claim 1, characterized in that, In step S1, near-infrared hyperspectral images of the forestry monitoring area are collected, atmospheric correction is performed on the near-infrared hyperspectral images, and single-wavelength images corresponding to the bands closest to the wavelengths of 850nm, 1200nm and 1900nm are extracted and used as the first characteristic wavelength image, the second characteristic wavelength image and the third characteristic wavelength image, respectively.

3. The method for constructing forestry vegetation and non-vegetation identification indices based on near-infrared spectroscopy according to claim 1, characterized in that, S2 includes the following sub-steps: S21. Use a Gaussian filter to denoise the first characteristic wavelength image, the second characteristic wavelength image and the third characteristic wavelength image respectively to obtain the first smoothed wavelength image, the second smoothed wavelength image and the third smoothed wavelength image. S22. The first smoothed wavelength image, the second smoothed wavelength image, and the third smoothed wavelength image are processed using spatial attention, and the processing results are multiplied element-wise with the corresponding smoothed wavelength images to obtain the first weighted wavelength image, the second weighted wavelength image, and the third weighted wavelength image.

4. The method for constructing forestry vegetation and non-vegetation identification indices based on near-infrared spectroscopy according to claim 1, characterized in that, In S4, the vegetation recognition model includes an input layer, an initial convolutional block, dual attention channels, a first convolutional layer, a second convolutional layer, a third convolutional layer, a splicing layer, an aggregation layer, and a classification layer. The input layer, initial convolutional block, and dual attention channels are connected sequentially, and then connected to the first convolutional layer, the second convolutional layer, and the third convolutional layer, respectively. The first convolutional layer, the second convolutional layer, and the third convolutional layer are all connected to the splicing layer. The splicing layer, the aggregation layer, and the classification layer are connected sequentially.

5. The method for constructing forest vegetation and non-vegetation identification indices based on near-infrared spectroscopy according to claim 4, characterized in that, The expression for the initial convolutional block is: ； in, This represents the output of the initial convolutional block. This indicates a batch normalization operation. This indicates an activation operation. This represents the convolution operation. This represents the output of the input layer.

6. The method for constructing forest vegetation and non-vegetation identification indices based on near-infrared spectroscopy according to claim 4, characterized in that, The expression for the dual attention channel is: ； ； ； in, This represents the output of the initial convolutional block. This represents the Hadamard product operation. Indicates the first weight. Indicates the second weight. Indicates global average pooling. Indicates average pooling. This represents the activation function. This represents the output of the dual attention channel. This indicates a fully connected layer.

Images