An erosion channel and slope farmland remote sensing image data processing method, system and related device

Abstract

The application provides an erosion channel and slope farmland remote sensing image data processing method, system and related device, including: obtaining multi-source remote sensing data of a research area, obtaining a multi-band composite image and a digital elevation model based on the multi-source remote sensing data; obtaining a catchment network model based on the digital elevation model; inputting the multi-band composite image into a pre-constructed deep learning model to output corresponding geometric parameters; determining the vectorization boundary and development stage information of the erosion channel based on the output results of the catchment network model and the deep learning model. The application realizes the analysis and geometric parameter measurement of the internal structure of the erosion channel through the deep learning model, and corrects the deep fusion of the catchment network model and the remote sensing recognition result, thereby improving the spatial precision and integrity of the erosion channel boundary extraction, and realizing the automatic determination of the development stage of the erosion channel based on multiple indexes such as the head shape, the gully wall topography and the catchment dynamics.

Description

Technical Field

[0001] This application relates to the field of remote sensing image processing, and in particular to a method, system and related apparatus for processing remote sensing image data of eroded gullies and sloping farmland. Background Technology

[0002] Gullies are gully-like landforms formed by concentrated surface runoff eroding surface soil under the influence of external forces such as water, gravity, and freeze-thaw cycles. Their development is characterized by upstream advancement of the gully head, gully wall collapse and expansion, and deepening of the gully bottom, marking a significant stage in the development of soil erosion. In areas with concentrated sloping farmland, such as the Loess Plateau in my country, gullies are numerous and actively developing, leading to increased farmland fragmentation, thinning of the soil layer year by year, continuous loss of soil organic matter, and difficulty for large agricultural machinery to operate, severely restricting the sustainable development of regional agriculture and the capacity for food security. Therefore, accurate and efficient identification and monitoring of gullies, obtaining information on their spatial distribution, geometric scale, and development status, is fundamental to carrying out soil and water conservation, assessing the effectiveness of conservation efforts, and formulating prevention and control plans.

[0003] Traditional methods for surveying gullies primarily rely on a combination of manual field measurements and visual interpretation of remote sensing images. While manual field measurements can obtain highly accurate gully morphology data, they are labor-intensive, time-consuming, and costly. Furthermore, they are limited by factors such as terrain conditions and accessibility, making large-scale, high-frequency dynamic monitoring difficult. Visual interpretation of remote sensing images improves efficiency to some extent, but this method is highly dependent on the professional experience and interpretation skills of the interpreters. Consistency between different interpreters is difficult to guarantee, and the accuracy and reliability of the interpretation results are significantly affected by subjective factors. Moreover, for shallow and narrow gullies in their early stages of development, their spectral and textural features on remote sensing images are relatively weak, leading to a relatively high rate of missed and false positives in visual interpretation. Summary of the Invention

[0004] This application provides a method, system, and related apparatus for processing remote sensing image data of eroded gullies and sloping farmland to improve the above-mentioned problems.

[0005] To achieve the above objectives, this application adopts the following technical solution: In a first aspect, embodiments of this application propose a method for processing remote sensing image data of erosion gullies and sloping farmland, the method comprising: Acquire multi-source remote sensing data of the study area, and obtain multi-band synthetic images and digital elevation models based on the multi-source remote sensing data; Obtaining a catchment network model based on a digital elevation model; The multi-band synthetic image is input into a pre-built deep learning model, which is used to output the erosion channel and its corresponding boundary, head, wall and bottom, as well as the geometric parameters of the erosion channel based on the input multi-band synthetic image. Based on the output results of the catchment network model and the deep learning model, the vectorized boundary and development stage information of the erosion channel are determined.

[0006] In conjunction with the first aspect, optionally, multi-source remote sensing data of the study area is acquired, and multi-band synthetic images and digital elevation models are obtained based on the multi-source remote sensing data, including: Acquire multispectral and panchromatic images covering the study area; Multispectral and panchromatic images are fused to generate a fused image. The corresponding images of the near-infrared band, red-edge band, and short-wave infrared band in the fused image are assigned to the red, green, and blue channels respectively for false-color synthesis to obtain the composite image; Obtain the gray value of each pixel in the synthesized image, and determine the gray value distribution data based on the gray value; The low-end truncation threshold and the high-end truncation threshold are determined based on the grayscale distribution data. The low-end truncation threshold corresponds to the first preset percentage of the grayscale value distribution in the grayscale distribution data, and the high-end truncation threshold corresponds to the second preset percentage of the grayscale value distribution in the grayscale distribution data. Pixels with gray values ​​below the low-end truncation threshold are set to the minimum value of the preset gray level range, and pixels with gray values ​​above the high-end truncation threshold are set to the maximum value of the preset gray level range. Pixels with gray values ​​between the low-end and high-end truncation thresholds are linearly mapped to the preset gray level range, and multi-band composite images are obtained.

[0007] In conjunction with the first aspect, optionally, the multispectral image and the panchromatic image are fused to generate a fused image, including: Based on different bands in the multispectral image, a panchromatic simulated band is generated. The panchromatic simulated band is a single-band raster image generated by linear combination of different bands of the multispectral image. Acquire the difference image between panchromatic image and panchromatic analog band based on different pixel gray values; The gray values ​​of each pixel in the difference image are superimposed onto each band of the multispectral image according to a preset weighting coefficient to generate a fused image.

[0008] In conjunction with the first aspect, optionally, the gray value of each pixel in the difference image is superimposed onto each band of the multispectral image according to a preset weighting coefficient to generate a fused image, including: Divide the panchromatic image into multiple non-overlapping image sub-blocks; Obtain the local variance of the gray values ​​of all pixels in each image sub-block, and generate a local variance matrix corresponding to the panchromatic image; Normalize each variance value in the local variance matrix to obtain a normalized variance value. The normalization process involves linearly mapping the variance value to a preset weight range. The normalized variance values ​​are used as the initial weight coefficients at the corresponding spatial locations to generate the initial weight coefficient matrix. Determine whether each band in the multispectral image is a near-infrared band or a short-wave infrared band; If so, the initial weighting coefficient is multiplied by the preset spectral protection factor to obtain the corrected weighting coefficient. Then, the pixel gray value at the corresponding spatial location in the difference image is multiplied by the corrected weighting coefficient and superimposed onto the band. If not, the pixel gray value at the corresponding spatial location in the difference image is multiplied by the initial weighting coefficient and then directly superimposed onto that band.

[0009] In conjunction with the first aspect, optionally, the multi-band synthetic image is input into a pre-constructed deep learning model, wherein the deep learning model is used to output, based on the input multi-band synthetic image, erosion channels and their corresponding boundaries, heads, walls, and bottoms, as well as the geometric parameters corresponding to the erosion channels, including: Multi-band synthetic images are input into an encoder network, and multi-scale feature maps are extracted through multiple cascaded convolutional layers and downsampling layers. The multi-scale feature maps include feature tensors at multiple levels with different spatial resolutions. The highest-level feature tensor is input into the global context attention module to generate a channel-weighted vector. The channel-weighted vector is then multiplied by the highest-level feature tensor to obtain the enhanced high-level semantic features. The feature tensors of each level except the highest level in the multi-scale feature map are fused with the upsampled feature tensors in the corresponding decoder level through skip connections, and combined with the enhanced high-level semantic features to generate a semantic segmentation probability map with the same spatial size as the multi-band synthetic image. Based on the semantic segmentation probability map, a classification label map containing the boundary, head region, wall region and bottom region of the erosion channel is generated through pixel-by-pixel classification. Extract the binary mask data corresponding to the head region, wall region, and bottom region of the trench from the classification label image; Based on binary mask data, at least one geometric parameter among the length, width, area, and depth of the erosion channel is determined.

[0010] In conjunction with the first aspect, optionally, based on binary mask data, at least one geometric parameter among the length, width, area, and depth of the erosion channel is determined, including: Extract the skeleton line of the binary mask data corresponding to the bottom region of the trench, and use the length of the skeleton line as the length of the erosion trench. In the binary mask data corresponding to the trench wall region, the distance between the two trench wall boundaries is calculated along the normal direction of the skeleton line, and the average value is taken to determine the width of the erosion channel; Based on the digital elevation model, the elevation values ​​of the bottom area and the top edge of the gully wall are obtained, and the difference in elevation is used as the depth of the erosion channel.

[0011] In conjunction with the first aspect, optionally, based on the output results of the catchment network model and the deep learning model, the vectorized boundary and development stage information of the erosion channel are determined, including: The boundary of the erosion channel output by the deep learning model is obtained as the initial channel boundary, and the mask data of the channel head region, channel wall region and channel bottom region are determined. Obtain raster data of runoff accumulation and water flow direction from the catchment network model; The initial channel boundary is spatially overlaid with the runoff accumulation raster data to identify pixels within the initial channel boundary whose runoff accumulation exceeds a preset accumulation threshold, and these pixels are then used as candidate pixels for the channel centerline. Based on the water flow direction raster data, determine the channel centerline vector with a continuous single pixel width; Based on the centerline vector of the channel, extend to both sides until the extension boundary touches the boundary of the channel wall area to generate the vectorized boundary of the erosion channel; Based on the mask data of the gully head region, the ratio of the gully head head erosion length to the gully head width is determined as an indicator of gully head activity. Based on the mask data and digital elevation model of the trench wall area, the average slope of the trench wall area is determined as an indicator of trench wall stability. The catchment area of ​​the erosion channel is determined based on the water catchment network model and used as a water catchment dynamic indicator. The gully head activity index, gully wall stability index, and water collection dynamic index are weighted and combined to obtain the gully development stage index. Based on the comparison between the gully development stage index and the preset classification threshold, the development stage information of the erosion gully is output. The development stage information includes active period, semi-stable period, or stable period.

[0012] Secondly, this application proposes a remote sensing image data processing system for erosion gullies and sloping farmland, the system comprising: The first acquisition module is used to acquire multi-source remote sensing data of the study area, and to acquire multi-band synthetic images and digital elevation models based on the multi-source remote sensing data. The second acquisition module is used to acquire the catchment network model based on the digital elevation model. The first output module is used to input multi-band synthetic images into a pre-built deep learning model. The deep learning model is used to output erosion channels and their corresponding boundaries, head, walls and bottom, as well as the geometric parameters corresponding to the erosion channels, based on the input multi-band synthetic images. The second output module is used to determine the vectorized boundary and development stage information of erosion channels based on the output results of the water catchment network model and the deep learning model.

[0013] A third aspect of this invention provides an electronic device, which includes: At least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the method proposed in the first aspect of the present invention.

[0014] A fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method as described in the first aspect of the present invention.

[0015] In summary, the above method and apparatus have the following technical effects: This invention proposes a method for processing remote sensing image data of erosion gullies and sloping farmland, comprising: acquiring multi-source remote sensing data of the study area; acquiring multi-band synthetic images and a digital elevation model (DEM) based on the multi-source remote sensing data; acquiring a catchment network model based on the DEM; inputting the multi-band synthetic images into a pre-constructed deep learning model, which outputs erosion gullies and their boundaries, gully heads, gully walls, and gully bottoms, as well as their corresponding geometric parameters; and determining the vectorized boundaries and developmental stages of erosion gullies based on the outputs of the catchment network model and the deep learning model. This invention achieves the analysis of the internal structure and measurement of geometric parameters of erosion gullies through a deep learning model, and improves the spatial accuracy and completeness of erosion gully boundary extraction by utilizing the deep fusion and correction of the catchment network model and remote sensing identification results. Simultaneously, it achieves automatic determination of the developmental stage of erosion gullies based on multiple indicators such as gully head morphology, gully wall topography, and catchment dynamics. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a method for processing remote sensing image data of erosion gullies and sloping farmland proposed in an embodiment of this application.

[0017] Figure 2 This is a schematic diagram of the structure of a remote sensing image data processing system for erosion gullies and sloping farmland proposed in an embodiment of this application. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] This application proposes a method for processing remote sensing image data of erosion gullies and sloping farmland. Please refer to [link to relevant documentation]. Figure 1 The method includes the following steps: S101: Acquire multi-source remote sensing data of the study area, and acquire multi-band synthetic images and digital elevation models based on the multi-source remote sensing data.

[0020] As is understandable, multi-source remote sensing data refers to one or more remote sensing observation data covering the same study area acquired through different remote sensing platforms and sensors. In this step, multi-source remote sensing data includes at least high spatial resolution multispectral imagery, panchromatic imagery, and radar interferometric data or stereo image pairs used to extract topographic information. Multispectral imagery records the reflection or radiation information of ground features in multiple discrete spectral bands, such as the visible light band, near-infrared band, red-edge band, and short-wave infrared band; panchromatic imagery covers a wider spectral range and typically has higher spatial resolution than multispectral imagery; radar interferometric data or stereo image pairs can be used to reconstruct the three-dimensional morphology of the Earth's surface.

[0021] Multiband composite image refers to the combination of preprocessed and enhanced image data from multiple spectral bands into a single multichannel remote sensing image. This image can comprehensively reflect the response characteristics of ground features in different spectral ranges, and is beneficial for highlighting the spectral differences and texture features between erosion channels and sloping farmland.

[0022] Specifically, in this embodiment, step S101 may include the following steps: S1011: Acquire multispectral and panchromatic images covering the study area.

[0023] Understandably, multispectral imagery contains multiple relatively narrow spectral bands, each recording the reflectance or radiation intensity of ground features within a specific wavelength range, such as the blue, green, and red bands in the visible light range, as well as non-visible light bands such as near-infrared, red-edge, and short-wave infrared. Panchromatic imagery, on the other hand, is a single-band grayscale image. Its spectral response typically covers a wider range from visible to near-infrared, offering higher spatial resolution than multispectral imagery and revealing clearer ground feature textures and edge details.

[0024] S1012: Perform fusion processing on multispectral and panchromatic images to generate a fused image.

[0025] Understandably, the purpose of image fusion is to combine the high spatial resolution detail of panchromatic imagery with the rich spectral information of multispectral imagery, thereby obtaining a fused image that possesses both high spatial resolution and retains the spectral characteristics of multiple spectral bands. Through fusion processing, the geometric boundaries of ground features in the fused image are clearer, while the values ​​of each spectral band still reflect the spectral response characteristics of ground features at different wavelengths. This is beneficial for distinguishing ground feature types with different spectral characteristics in subsequent processing, such as bare soil areas and moist soil areas.

[0026] Furthermore, in this embodiment, the specific generation of the fused image can be achieved through the following steps: S202: Generate panchromatic simulated bands based on different bands in multispectral images. The panchromatic simulated bands are single-band raster images generated by linear combination of different bands in multispectral images.

[0027] Understandably, a panchromatic simulation band is a single-band raster image where the gray value of each pixel is calculated by linearly combining the gray values ​​of pixels in multiple bands of the multispectral image. By simulating the spectral response characteristics of the panchromatic image—that is, by using a weighted summation of the multispectral bands in the visible to near-infrared range—it is possible to approximate the wide-band radiation intensity recorded by the panchromatic sensor. For example, if the multispectral image contains four bands—blue, green, red, and near-infrared—the gray value of a pixel at a certain spatial location in the panchromatic simulation band can be expressed as the sum of the gray values ​​of the blue, green, red, and near-infrared bands at that location, multiplied by their respective weighting coefficients. The weighting coefficients can be determined based on the spectral response functions of both the multispectral and panchromatic sensors, ensuring that the spectral coverage of the panchromatic simulation band is as consistent as possible with the real panchromatic image.

[0028] S203: Acquire the difference image between the panchromatic image and the panchromatic analog band based on different pixel gray values.

[0029] Understandably, a difference image is a raster image with the same spatial size and resolution as a panchromatic image. The gray value of a pixel at each spatial location is equal to the gray value of the pixel in the panchromatic image at that location minus the gray value of the corresponding pixel in the panchromatic analog band. Since the panchromatic analog band is generated based on multispectral imagery and has a relatively low spatial resolution, while the panchromatic imagery has a higher spatial resolution, the difference imagery mainly retains high-frequency spatial details in the panchromatic imagery that exceed the spatial resolution of the multispectral imagery, such as the edges, textures, and fine structures of ground features, while most of the shared low-frequency spectral information is canceled out during the subtraction process.

[0030] S204: The gray value of each pixel in the difference image is superimposed onto each band of the multispectral image according to the preset weighting coefficient to generate a fused image.

[0031] Specifically, for each band in the multispectral image, the gray value of the pixel at each spatial location in the original image of that band is added to the product of the gray value of the corresponding pixel at the spatial location in the difference image and a preset weighting coefficient for that band, resulting in the fused pixel gray value for that band. The weighting coefficient controls the intensity of spatial details injected into each band and can be set according to the spectral characteristics of each band and the degree of overlap with the spectral response range of the panchromatic band. For example, for bands that overlap significantly with the panchromatic band spectrum, such as the green and red bands, a larger weighting coefficient can be set to fully inject spatial details; for bands that overlap less or almost not with the panchromatic band spectrum, a smaller weighting coefficient can be set to avoid excessive distortion of spectral information. By weighting and superimposing the high-frequency detail information in the difference image onto each band of the multispectral image, the final fused image has a spatial resolution close to that of the panchromatic image, while the spectral characteristics of each band are well preserved.

[0032] Furthermore, in this embodiment, in order to dynamically adjust the injection intensity of spatial details according to the texture complexity of different regions of the image during the fusion process, and to implement additional spectral protection for specific spectral bands, step S204 may include the following steps: S2041: Divide the panchromatic image into multiple non-overlapping image sub-blocks.

[0033] The size of the image sub-blocks can be set according to the image spatial resolution and the scale characteristics of the ground texture. For example, the image can be divided into several square blocks with a side length of a certain number of pixels. Each image sub-block contains a certain number of adjacent pixels, and there are no overlapping areas between the sub-blocks, which together cover the entire panchromatic image range.

[0034] S2042: Obtain the local variance of the gray values ​​of all pixels in each image sub-block and generate a local variance matrix corresponding to the panchromatic image.

[0035] As is understandable, local variance is a statistical measure of the dispersion or fluctuation of pixel grayscale values ​​within a sub-block. It is calculated as the average of the squares of the differences between the grayscale values ​​of each pixel within the sub-block and its mean. A larger local variance indicates more dramatic grayscale variations and richer texture details within the sub-block; a smaller local variance indicates a more uniform grayscale distribution and gentler spatial variations within the sub-block. For example, in textured areas such as the head and edges of eroded gullies, local variance is typically high; while within flat sloping farmland, local variance is relatively low. The spatial resolution of the local variance matrix is ​​lower than that of the original panchromatic image, but it can be restored to a spatial resolution consistent with the panchromatic image through nearest-neighbor interpolation or bilinear interpolation.

[0036] S2043: Normalize each variance value in the local variance matrix to obtain a normalized variance value, wherein the normalization process is to linearly map the variance value to a preset weight range.

[0037] Understandably, normalization refers to transforming the original variance values ​​to a preset weight range using a linear mapping. For example, the preset weight range could be a closed interval from the minimum to the maximum value. The linear mapping rule is: map the minimum variance value in the local variance matrix to the minimum value of the weight range, map the maximum variance value to the maximum value of the weight range, and linearly scale the remaining variance values ​​proportionally to this interval. After normalization, the variance differences between different image sub-blocks are uniformly transformed to the same dimension, making them easier to use as weighting coefficients.

[0038] S2044: Use the normalized variance value as the initial weight coefficient at the corresponding spatial location to generate the initial weight coefficient matrix.

[0039] Understandably, each element in the initial weighting coefficient matrix corresponds to a weight value for the corresponding spatial location in the panchromatic image. In areas with high local variance, such as gully heads and gully edges, the initial weighting coefficients are higher; in areas with low local variance, such as flat slopes, the initial weighting coefficients are lower. These initial weighting coefficients reflect the basic strength of spatial detail injection at each spatial location.

[0040] S2045: Determine whether each band in the multispectral image is a near-infrared band or a short-wave infrared band.

[0041] Understandably, the near-infrared band is more sensitive to distinguishing between vegetation cover and bare soil, while the short-wave infrared band responds significantly to changes in soil moisture content. During the fusion process, excessive injection of spatial details into these two bands could lead to significant shifts in their original spectral values, thereby weakening their ability to reflect the composition and moisture status of surface materials.

[0042] S2046: If so, multiply the initial weighting coefficient by the preset spectral protection factor to obtain the corrected weighting coefficient, and then multiply the pixel gray value at the corresponding spatial location in the difference image by the corrected weighting coefficient and superimpose it onto the band.

[0043] In this application, the spectral protection factor is a value less than one, used to suppress the intensity of spatial detail injection into the band, thereby improving spatial resolution while maintaining spectral fidelity. Subsequently, the pixel gray value at the corresponding spatial location in the difference image is multiplied by the corrected weighting coefficient, and the product is then superimposed onto the corresponding spatial location in the original image of the band to complete the fusion of the band.

[0044] The spectral protection factor is a pre-defined numerical parameter whose value is mainly based on the assessment of the spectral importance of the near-infrared band and the short-wave infrared band in the task of identifying erosion trenches, as well as the consideration of the balance between spectral fidelity and spatial resolution improvement during the fusion process.

[0045] Specifically, the near-infrared band is highly sensitive to differences in vegetation cover and bare soil, while the short-wave infrared band shows a significant response to changes in soil moisture content. These two bands are key spectral information carriers for distinguishing between the bare soil areas on the walls of gullies and the moist soil areas at the bottom. Injecting excessive spatial detail into these two bands during the fusion process can lead to significant changes in their pixel grayscale values, thereby weakening or distorting the land cover attribute information carried by the original spectral data and affecting the subsequent identification results based on spectral features.

[0046] The spectral protection factor is typically less than one, and its specific value can be determined by analyzing the overlap between the spectral response range of the panchromatic image and the spectral response range of each multispectral band. For example, for bands with a small overlap between their spectral response range and the panchromatic band, a smaller spectral protection factor can be set to apply a stronger suppression effect; for bands with a large overlap, a relatively larger spectral protection factor can be set. Furthermore, by conducting comparative experiments on the spectral curves of typical ground cover samples before and after fusion, a factor value that keeps the spectral deviation of key bands within an acceptable range can be selected as a preset value.

[0047] S2047: If not, the pixel gray value at the corresponding spatial location in the difference image is multiplied by the initial weighting coefficient and then directly superimposed onto the band.

[0048] For bands determined to be neither near-infrared nor short-wave infrared, such as the blue, green, and red bands in the visible light range, no additional spectral protection measures are applied. In this case, the pixel gray value at the corresponding spatial location in the difference image is directly multiplied by the initial weighting coefficient and then superimposed onto the corresponding spatial location in the original image of that band to complete the fusion process.

[0049] S1013: Assign the corresponding images of the near-infrared band, red-edge band and short-wave infrared band in the fused image to the red, green and blue channels respectively for false color synthesis to obtain a composite image.

[0050] Understandably, false-color composite is a display and enhancement method that maps multiple single-band images onto the red, green, and blue color channels respectively. The colors of ground features in the resulting composite image are not their true natural colors, but are determined by the combination of spectral intensities of the selected bands. False-color composite is performed by assigning the near-infrared band to the red channel, the red-edge band to the green channel, and the short-wave infrared band to the blue channel. In the image synthesized using this combination, the gully wall area and the gully bottom area will exhibit a significant tonal difference, thereby enhancing the visual contrast between different structural units within the gully and facilitating the differentiation of different components within the gully.

[0051] S1014: Obtain the gray value of each pixel in the synthesized image, and determine the gray distribution data based on the gray value.

[0052] Understandably, a composite image consists of three color channels, which can be processed separately or converted into a single-band brightness image before processing. Grayscale values ​​refer to the numerical value corresponding to a pixel, reflecting the brightness level of that location across each color channel. By statistically analyzing the grayscale values ​​of all pixels in the composite image, grayscale distribution data reflecting the frequency or cumulative distribution of grayscale values ​​can be generated, such as a grayscale cumulative distribution histogram. This grayscale distribution data characterizes the overall brightness distribution features and dynamic range of the composite image.

[0053] S1015: Determine the low-end truncation threshold and the high-end truncation threshold based on the grayscale distribution data, wherein the low-end truncation threshold corresponds to the first preset percentage of the grayscale value distribution in the grayscale distribution data, and the high-end truncation threshold corresponds to the second preset percentage of the grayscale value distribution in the grayscale distribution data.

[0054] For example, if the first preset percentage is set to 2%, the lower cutoff threshold is the gray value corresponding to the cumulative percentage reaching 2% in the gray-scale cumulative distribution curve; if the second preset percentage is set to 98%, the higher cutoff threshold is the gray value corresponding to the cumulative percentage reaching 98%. Determining the cutoff threshold based on the gray-scale distribution percentage allows the threshold selection to adapt to the gray-scale statistical characteristics of different images, rather than relying on a fixed absolute value.

[0055] S1016: Set pixels with gray values ​​below the low-end truncation threshold to the minimum value of the preset gray level range, set pixels with gray values ​​above the high-end truncation threshold to the maximum value of the preset gray level range, and linearly map pixels with gray values ​​between the low-end and high-end truncation thresholds to the preset gray level range, and acquire multi-band composite images.

[0056] Understandably, the preset grayscale range refers to the range of grayscale values ​​that can be represented when storing image data. For example, for an 8-bit image, the preset grayscale range is 0 to 255. Through this linear stretching process, the effective information in the synthesized image, which was originally concentrated in a narrow grayscale range, is stretched to the entire usable grayscale range. Meanwhile, a very small number of extreme pixels located at both ends of the distribution are truncated to the boundary values, effectively suppressing the influence of noise and outliers. At the same time, it enhances the overall contrast of the image, making the boundaries and internal structure of the erosion channels clearer and more discernible.

[0057] S102: Obtaining the catchment network model based on the digital elevation model.

[0058] Understandably, a digital elevation model (DEM) is a rasterized representation of the surface elevation information of a study area. Each raster cell stores the ground elevation value at that spatial location, reflecting the undulating shape of the terrain. A catchment network model, on the other hand, is a spatial data model based on the DEM, extracted through hydrological analysis methods. It describes the convergence path and catchment relationships of surface runoff, reflecting the flow direction, convergence process, and distribution pattern of gully systems under gravity.

[0059] Specifically, the process of obtaining a catchment network model based on a digital elevation model (DEM) includes the following hydrological analysis steps. First, depression filling is performed on the DEM to identify and fill local depressions. Local depressions refer to areas where the elevation value is lower than that of all surrounding adjacent raster cells. These areas create "pseudo-catchment areas" in natural surface runoff simulations, preventing water flow. Depression filling raises the elevation value of the depressions to the elevation of the lowest outflow point around them, eliminating discontinuities in the topographic data and ensuring that water can flow continuously to the boundary of the study area.

[0060] Secondly, based on the digital elevation model after filling the depressions, the flow direction of each grid cell is calculated. The flow direction determines which of the eight neighboring grid cells the surface runoff flows from the current grid cell to. A commonly used calculation method is the eight-flow-direction algorithm, which compares the elevation gradient between the central grid cell and the eight neighboring grid cells and directs the flow direction in the direction of the steepest elevation drop.

[0061] Next, based on the flow direction data, the runoff accumulation for each grid cell is calculated. Runoff accumulation refers to the sum of the number of grid cells in the upstream catchment area that ultimately flow into that cell, reflecting the magnitude of the surface runoff flowing through that cell. During the calculation, starting from the grid cell at the watershed, the count is accumulated unit by unit downstream along the flow direction until the watershed outlet. A grid cell with a larger runoff accumulation indicates its location further downstream in the catchment network, carrying a larger volume of upstream water, corresponding to the location of channels or rivers in the actual terrain.

[0062] Finally, based on runoff accumulation data, a raster drainage network was extracted by setting a runoff accumulation threshold. Raster cells with runoff accumulation exceeding the preset threshold were identified as components of the drainage system, resulting in a binarized raster drainage distribution map. This raster drainage network serves as the basic representation of the drainage network model, reflecting the gully drainage framework formed by surface runoff erosion within the study area.

[0063] S103: Input the multi-band synthetic image into the pre-built deep learning model, wherein the deep learning model is used to output the erosion channel and its corresponding boundary, head, wall and bottom, as well as the geometric parameters of the erosion channel, based on the input multi-band synthetic image.

[0064] Understandably, the multi-band composite image obtained in the aforementioned steps is input into a pre-constructed deep learning model, which then outputs identification results related to erosion channels based on the content of the input image. This deep learning model is a pre-trained neural network model whose internal parameters have been determined through learning from a large number of labeled samples, and it has the ability to automatically extract specific ground features and their structural characteristics from remote sensing images.

[0065] Synthetic band imagery, used as input data for the deep learning model, provides multi-dimensional spatial and spectral information on land cover type, land cover texture, and spectral reflectance characteristics within the study area. After receiving the imagery, the model analyzes and determines each spatial location through multi-layered nonlinear transformations and feature extraction operations, ultimately outputting the following identification results: First, the overall spatial extent and boundary location of the erosion channels, i.e., the dividing line between pixels belonging to the erosion channels and pixels belonging to the background or other land cover; Second, the structural partitions within the erosion channels, specifically including the head region, wall region, and bottom region, corresponding to different geomorphic parts in the development process of the erosion channels; Third, the geometric parameters corresponding to the erosion channels. These geometric parameters are quantitative indicators extracted from the morphology and spatial relationships of the identified structural regions, and may include one or more of the following: length, width, area, and depth. For example, length reflects the extension distance of the erosion channel along the main channel direction, width reflects the degree of lateral expansion of the channel, area reflects the planar distribution scale of the erosion channel, and depth reflects the intensity of erosion incision. Through the processing of this deep learning model, the erosion channel information contained in the multi-band synthetic image is transformed into structured recognition results and quantitative geometric descriptions.

[0066] As a specific example, in this step, step S103 may include the following steps: S1031: Input the multi-band synthetic image into the encoder network, and extract multi-scale feature maps through multiple cascaded convolutional layers and downsampling layers. The multi-scale feature maps include feature tensors of multiple levels with different spatial resolutions.

[0067] The encoder network consists of cascaded convolutional and downsampling layers, which perform layer-by-layer abstraction and feature extraction on the input image. Convolutional layers use learnable convolutional kernels to slide across the image, extracting texture, edge, and shape features within a local spatial range. Downsampling layers progressively reduce the spatial resolution of the feature maps while expanding the receptive field of subsequent convolutional layers. After processing through multiple convolutional and downsampling layers, the encoder network outputs multi-scale feature maps. These feature maps form a set of feature tensors, with the spatial resolution decreasing and the number of channels increasing sequentially across different levels. Lower-level feature tensors have higher spatial resolution, preserving rich information about ground feature boundaries and texture details; higher-level feature tensors have lower spatial resolution but contain more abstract semantic information, reflecting the category and overall structural features of ground features.

[0068] S1032: Input the highest-level feature tensor into the global context attention module to generate a channel-weighted vector, and multiply the channel-weighted vector with the highest-level feature tensor to obtain the enhanced high-level semantic features.

[0069] Understandably, the highest-level feature tensor has the smallest spatial size and the most channels, containing highly condensed information that reflects the global semantic content of the input image. The global context attention module first performs global average pooling on this feature tensor, compressing the spatial information within each channel into a scalar value, forming a channel description vector. Then, a fully connected layer performs a nonlinear transformation on this channel description vector, generating a set of weighting coefficients with the same number of channels as the original, forming a channel weighting vector. Finally, the channel weighting vector is multiplied channel-by-channel by the original highest-level feature tensor, enhancing the channel responses corresponding to important semantic features while suppressing the channel responses corresponding to minor or irrelevant features. The enhanced high-level semantic features obtained through this process, while preserving the original semantic information, strengthen the feature channels that are more critical to the erosion channel identification task.

[0070] S1033: The feature tensors of each level except the highest level in the multi-scale feature map are fused with the upsampled feature tensors in the corresponding decoder level through skip connections, and combined with the enhanced high-level semantic features to generate a semantic segmentation probability map with the same spatial size as the multi-band synthetic image.

[0071] Specifically, the decoder network is symmetrical to the encoder network, and the spatial resolution of the feature map is gradually restored through cascaded upsampling operations. At each level of the decoder, the feature map output from the previous level is first upsampled to expand its spatial size. Then, through skip connections, the feature tensor of the encoder corresponding to the spatial resolution is concatenated with the upsampled feature tensor along the channel dimension, allowing the decoder to utilize both deep high-level semantic information and shallow spatial detail information simultaneously. The enhanced high-level semantic features then serve as the initial input to the deepest layer of the decoder, passed up level by level. After feature fusion and upsampling restoration at each level of the decoder, a semantic segmentation probability map with the same spatial size as the original multi-band composite image is finally generated. Each spatial location in this probability map contains a probability vector, where the value of each element represents the predicted probability that the pixel at that location belongs to a predefined category, such as the erosion gully boundary, gully head region, gully wall region, gully bottom region, and background.

[0072] S1034: Based on the semantic segmentation probability map, a classification label map containing the boundary, head region, wall region, and bottom region of the erosion channel is generated through pixel-by-pixel classification.

[0073] Specifically, for each spatial location in the probability map, the probability values ​​corresponding to each category at that location are compared, and the category with the highest probability value is selected as the final category assignment for that pixel. The index code of that category is then assigned to that pixel. After this pixel-by-pixel determination operation, the semantic segmentation probability map is converted into a single-band classification label map. In the classification label map, each pixel is assigned an integer label value, with different label values ​​representing different land cover types. For example, label value zero can represent the background area, label value one can represent the boundary of an erosion gully, label value two can represent the gully head area, label value three can represent the gully wall area, and label value four can represent the gully bottom area.

[0074] S1035: Extract the binary mask data corresponding to the head region, wall region, and bottom region of the trench from the classification label map.

[0075] For example, for a specific target category, all pixels in the classification label image can be traversed. Pixels whose category label equals the index value corresponding to the target category are assigned a value of one, while pixels whose category label does not equal the index value are assigned a value of zero. This generates a binary image containing only the values ​​of zero and one, which is the binary mask data for the target region. The set of pixels with a value of one in the binary mask data constitutes the precise spatial extent of the target region, which can be used for subsequent geometric parameter calculations. Using the same method, binary mask data for the gully head region, gully wall region, and gully bottom region can be extracted separately.

[0076] S1036: Based on binary mask data, determine at least one geometric parameter among the length, width, area, and depth of the erosion channel.

[0077] Understandably, the geometric parameters can be determined using the target area's shape and spatial location information defined by binary mask data. For example, the skeleton lines of the binary mask data corresponding to the trench bottom region can be extracted, and the length of the skeleton lines can be used as the length of the erosion channel. In the binary mask data corresponding to the trench wall region, the distance between the two trench wall boundaries is calculated along the normal direction of the skeleton lines, and the average value is taken to determine the width of the erosion channel. Based on a digital elevation model, the elevation values ​​of the trench bottom region and the top edge of the trench wall are obtained, and the difference in elevation is used as the depth of the erosion channel.

[0078] The above is one specific implementation method related to the deep learning model in this application. It should be understood that this implementation method is only used to illustrate the working principle and implementation process of the deep learning model in the erosion channel identification task, and does not constitute a limitation on the scope of protection of this application. In other embodiments, the specific architecture, training strategy, and operation mode of the deep learning model can be flexibly adjusted according to the actual application scenario and needs.

[0079] For example, in addition to constructing encoder networks using cascaded convolutional layers and downsampling layers, other backbone network architectures already disclosed in the field of image semantic segmentation can be adopted, including but not limited to residual networks and their variants, densely connected networks, efficient networks, or transform networks based on visual self-attention mechanisms. Decoder networks can employ upsampling methods such as transposed convolutions, bilinear interpolation combined with convolutional layers, pixel recombination, or other learnable upsampling operators. Feature fusion methods can include attention gating mechanisms, feature pyramid networks, or progressive feature aggregation structures, in addition to skip connections.

[0080] For example, the global context attention module can be replaced with a spatial attention module, a hybrid channel and spatial attention module, or a multi-scale context aggregation module can be used to enhance the perception of channel structures at different scales. In addition to pixel-by-pixel classification, the generation of semantic segmentation probability maps can also be combined with post-processing methods such as conditional random fields to finely adjust the segmentation boundaries, or a boundary-aware loss function can be used to enhance the model's attention to the edges of erosion channels during training.

[0081] For example, in addition to the multi-band synthetic imagery of this application, the training data sources for the deep learning model can also include topographic factors derived from the digital elevation model, such as slope, aspect, and topographic roughness, as auxiliary input channels to enhance the model's ability to perceive the micro-topography of gullies. During training, a transfer learning strategy can be employed, using model parameters pre-trained on large natural image or remote sensing image datasets as initial weights, and fine-tuning them on labeled samples of eroded gullies to accelerate convergence and improve generalization performance with a limited number of samples.

[0082] Furthermore, during the model inference stage, sliding window or block prediction methods can be used to process large-format remote sensing images. Weighted fusion of overlapping areas or voting mechanisms can be used to eliminate prediction inconsistencies at stitching boundaries. A multi-scale inference strategy can also be employed, inputting images of the same area at different spatial resolutions into the model and fusing the output multi-scale segmentation results to balance the continuity of large-scale channel trends with the precision of local channel heads and wall edges.

[0083] Therefore, this application does not strictly limit the specific implementation form of the deep learning model. Any neural network model architecture and operation mode that can extract erosion channels and their internal structure information from multi-band synthetic images and output corresponding geometric parameters can be used as a component or equivalent alternative to the technical solution of this application.

[0084] S104: Based on the output results of the catchment network model and the deep learning model, determine the vectorized boundary and development stage information of the erosion channel.

[0085] The catchment network model reflects the convergence path and drainage pattern of surface runoff within the study area. The accumulated runoff information it provides indicates locations with high concentrations of surface water flow, which typically correspond spatially to the bottom line of erosion channels. The output of the deep learning model includes the boundary extent of the erosion channel and identification information of internal structural regions such as the gully head, walls, and bottom, reflecting the spatial morphology of the erosion channel from the perspective of spectral and textural features of remote sensing images. In this step, determining the vectorized boundary of the erosion channel involves fusing the above two types of information to generate the spatial boundary line of the erosion channel expressed in vector data format. Vector boundary data consists of a series of ordered points with spatial coordinates, accurately describing the contour morphology of the channel and facilitating measurement, statistical analysis, and overlay analysis in a geographic information system. Compared to raster-based identification results, the development process of erosion channels can usually be divided into several stages with different morphological and activity characteristics, such as active, semi-stable, and stable periods. During the active phase, gullies are characterized by active headward erosion, significant gully wall collapse and expansion, and continuous downcutting of the gully floor. In the semi-stable phase, the erosion rate slows down, and the gully morphology becomes more fixed. In the stable phase, erosion essentially ceases, and vegetation on both sides of the gully gradually recovers. The output of developmental stage information can provide direct evidence for formulating soil and water conservation measures and prioritizing treatment efforts.

[0086] Specifically, step S104 may include the following steps: S1041: Obtain the boundary of the erosion channel output by the deep learning model as the initial channel boundary, and determine the mask data of the channel head region, channel wall region and channel bottom region.

[0087] Understandably, the initial channel boundary is represented by a raster of the erosion channel's outer contour, its location derived from the deep learning model's analysis of spectral and textural features in the multi-band composite image. The mask data for the channel head, wall, and bottom regions are binary raster data, where a pixel with a value of one indicates its spatial location within the corresponding structural region, and a pixel with a value of zero indicates it does not belong to that region.

[0088] S1042: Obtain raster data of runoff accumulation and flow direction from the runoff network model.

[0089] Each pixel value in the runoff accumulation raster data represents the number of raster cells accumulated in the upstream catchment area at that location. The larger the value, the higher the degree of surface runoff accumulation in the area, which usually has a high degree of consistency with the bottom location of the channel in the actual terrain.

[0090] S1043: Spatially overlay the initial channel boundary with the accumulator raster data to determine the pixels within the initial channel boundary whose accumulator volume is greater than a preset accumulator threshold, and use them as candidate pixels for the channel centerline.

[0091] For example, the spatial overlay operation is performed pixel-by-pixel under the same geographic coordinate system and spatial resolution. For each pixel within the initial channel boundary, the runoff accumulation value at its corresponding spatial location is read. If the value exceeds a preset accumulation threshold, the pixel is marked as a candidate pixel for the channel centerline. The preset accumulation threshold can be determined based on the topographic relief of the study area and the spatial resolution of the digital elevation model. Its purpose is to screen out runoff convergence areas with significant water collection effects sufficient to form stable channels.

[0092] S1044: Determine the channel centerline vector with a continuous single-pixel width based on the water flow direction raster data.

[0093] Understandably, starting with candidate pixels for the channel centerline, the connection paths between adjacent candidate pixels upstream and downstream are traced based on the flow direction relationship indicated by the raster data, connecting the discretely distributed candidate pixels into an initial centerline according to the continuity of the flow direction. Subsequently, a morphological thinning algorithm can be used to process the initial centerline, reducing the line width to the width of a single pixel while maintaining the topological connectivity of the centerline. Finally, the thinned raster centerline is converted into a vector format consisting of a series of ordered coordinate points, forming the channel centerline vector.

[0094] S1045: Based on the centerline vector of the channel, extend to both sides until the extension boundary touches the boundary of the channel wall area to generate the vectorized boundary of the erosion channel.

[0095] Specifically, at each point along the centerline vector of the channel, the tangent direction of the centerline at that point is calculated, and the normal direction perpendicular to the tangent direction is determined. Starting from the centerline position, the expansion proceeds pixel by pixel to the left and right along the normal direction. At each step, it is determined whether the current expansion position has reached the boundary of a pixel with a value of one in the mask data of the channel wall region. If the channel wall boundary is reached, the expansion on that side stops; if not, the expansion continues. Thus, the expansion termination positions on both sides of the centerline together define the lateral extent of the channel at that location. Connecting the expansion boundary points on both sides of all cross-sections sequentially generates the vectorized boundary of the erosion channel, expressed as a vector polygon or vector line.

[0096] S1046: Based on the mask data of the gully head region, determine the ratio of the gully head source erosion length to the gully head width as an indicator of gully head activity.

[0097] It should be understood that the head erosion length refers to the farthest longitudinal extension of the head region along the gully channel, which can be obtained by extracting its minimum bounding rectangle from the head region mask data or by projecting its length along the main channel direction. The head width refers to the lateral width of the head region perpendicular to the direction of head erosion. The ratio of these two values ​​reflects the relative development degree of the head in both longitudinal extension and lateral widening dimensions. A larger ratio indicates that head erosion is more significant than lateral expansion, and the head is in a more active developmental state.

[0098] S1047: Based on the mask data and digital elevation model of the trench wall area, determine the average slope of the trench wall area as an indicator of trench wall stability.

[0099] Specifically, firstly, the spatial extent of the trench wall is defined using mask data of the trench wall region. Then, based on a digital elevation model, the elevation values ​​of each pixel within this range are extracted, and the local terrain slope at each pixel is calculated. The slope can be obtained using the arctangent function of the ratio of the elevation difference between adjacent pixels to the horizontal distance. Finally, the slope values ​​of all pixels within the trench wall region are statistically analyzed, and their arithmetic mean is taken as the average slope of the trench wall region. A larger average slope indicates a steeper trench wall, a higher probability of collapse, landslides, and other instability phenomena under the influence of gravity and hydraulic forces, and lower trench wall stability.

[0100] S1048: Based on the water catchment network model, determine the catchment area of ​​the water flowing into the erosion channel and use it as a water catchment dynamic indicator.

[0101] It should be understood that the catchment area refers to the total area of ​​the upstream water catchment area of ​​the erosion channel, that is, the sum of the surface areas of all surface runoff source areas that ultimately flow into the channel. Using the flow direction data in the catchment network model, all grid cells flowing upstream into the channel can be traced from the channel outlet. The number of these grid cells is counted, and multiplied by the actual area of ​​a single grid cell to obtain the catchment area. A larger catchment area indicates a larger flow rate into the channel, stronger scouring and erosion dynamics, and more abundant hydrodynamic conditions driving the channel's continued development.

[0102] S1049: The gully head activity index, gully wall stability index, and water collection dynamic index are weighted and combined to obtain the gully development stage index. Based on the comparison between the gully development stage index and the preset classification threshold, the development stage information of the erosion gully is output. The development stage information includes active period, semi-stable period, or stable period.

[0103] Specifically, the weighted aggregation method involves multiplying each normalized index value by a preset weighting coefficient and summing the results. The allocation of weighting coefficients can be determined based on the importance of each index in identifying the development stage of the erosion gully. The calculated gully development stage index is compared with preset first and second grading thresholds: if the index is higher than the first grading threshold, the erosion gully is determined to be in an active phase, characterized by strong erosion and continuous morphological changes; if the index is between the first and second grading thresholds, it is determined to be in a semi-stable phase, with a slower erosion rate; if the index is lower than the second grading threshold, it is determined to be in a stable phase, with erosion essentially ceasing. The output development stage information falls into one of these three categories, providing direct scientific basis for erosion gully management decisions.

[0104] This invention proposes a method for processing remote sensing image data of erosion gullies and sloping farmland, comprising: acquiring multi-source remote sensing data of the study area; acquiring multi-band synthetic images and a digital elevation model (DEM) based on the multi-source remote sensing data; acquiring a catchment network model based on the DEM; inputting the multi-band synthetic images into a pre-constructed deep learning model, which outputs erosion gullies and their boundaries, gully heads, gully walls, and gully bottoms, as well as their corresponding geometric parameters; and determining the vectorized boundaries and developmental stages of erosion gullies based on the outputs of the catchment network model and the deep learning model. This invention achieves the analysis of the internal structure and measurement of geometric parameters of erosion gullies through a deep learning model, and improves the spatial accuracy and completeness of erosion gully boundary extraction by utilizing the deep fusion and correction of the catchment network model and remote sensing identification results. Simultaneously, it achieves automatic determination of the developmental stage of erosion gullies based on multiple indicators such as gully head morphology, gully wall topography, and catchment dynamics.

[0105] Based on the same inventive concept, embodiments of this application also propose a remote sensing image data processing system for erosion gullies and sloping farmland. Please refer to [link to relevant documentation]. Figure 2 The system includes: The first acquisition module is used to acquire multi-source remote sensing data of the study area, and to acquire multi-band synthetic images and digital elevation models based on the multi-source remote sensing data. The second acquisition module is used to acquire the catchment network model based on the digital elevation model. The first output module is used to input multi-band synthetic images into a pre-built deep learning model. The deep learning model is used to output erosion channels and their corresponding boundaries, head, walls and bottom, as well as the geometric parameters corresponding to the erosion channels, based on the input multi-band synthetic images. The second output module is used to determine the vectorized boundary and development stage information of erosion channels based on the output results of the water catchment network model and the deep learning model.

[0106] Based on the same inventive concept, embodiments of this application also propose an electronic device, which includes: At least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the remote sensing image data processing method for eroded gullies and sloping farmland according to embodiments of this application.

[0107] In addition, to achieve the above objectives, embodiments of this application also propose a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the remote sensing image data processing method for erosion gullies and sloping farmland according to embodiments of this application.

[0108] The following is a detailed introduction to the various components of the electronic device: In this context, the processor is the control center of the electronic device. It can be a single processor or a collective term for multiple processing elements. For example, a processor can be one or more central processing units (CPUs), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention, such as one or more digital signal processors (DSPs), or one or more field-programmable gate arrays (FPGAs).

[0109] Alternatively, the processor can perform various functions of the electronic device by running or executing software programs stored in memory and by calling data stored in memory.

[0110] The memory is used to store the software program that executes the solution of the present invention, and the execution is controlled by the processor. The specific implementation method can be referred to the above method embodiment, which will not be repeated here.

[0111] Optionally, the memory can be read-only memory (ROM) or other types of static storage devices capable of storing static information and instructions, random access memory (RAM) or other types of dynamic storage devices capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto. The memory can be integrated with the processor or exist independently and coupled to the processor through an interface circuit of an electronic device; the embodiments of the present invention do not specifically limit this.

[0112] A transceiver is used to communicate with network devices or with terminal devices.

[0113] Optionally, the transceiver may include a receiver and a transmitter. The receiver is used to implement the receiving function, and the transmitter is used to implement the sending function.

[0114] Optionally, the transceiver can be integrated with the processor or exist independently and coupled to the processor through the router's interface circuit. This embodiment of the invention does not specifically limit this.

[0115] Furthermore, the technical effects of the electronic device can be referred to the technical effects of the data transmission method in the above method embodiments, and will not be repeated here.

[0116] It should be understood that the processor in the embodiments of the present invention can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor.

[0117] It should also be understood that the memory in the embodiments of the present invention can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of random access memory (RAM) are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate synchronous DRAM (DDRSDRAM), enhanced synchronous DRAM (ESDRAM), synchronous linked DRAM (SLDRAM), and direct rambus RAM (DRRAM).

[0118] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the flow or function according to the embodiments of the present invention is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. A computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. Available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media. Semiconductor media can be solid-state drives.

[0119] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0120] In this invention, "at least one" means one or more, and "more than one" means two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of a single item or a plurality of items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be a single item or multiple items.

[0121] It should be understood that, in various embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0122] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

Claims

1. A method for processing remote sensing image data of erosion gullies and sloping farmland, characterized in that, The method includes: Acquire multi-source remote sensing data of the study area, and acquire multi-band synthetic images and digital elevation models based on the multi-source remote sensing data; A catchment network model is obtained based on the digital elevation model; The multi-band synthetic image is input into a pre-built deep learning model, wherein the deep learning model is used to output the erosion channel and its corresponding boundary, head, wall and bottom, as well as the geometric parameters corresponding to the erosion channel, based on the input multi-band synthetic image. Based on the output of the water catchment network model and the deep learning model, the vectorized boundary and development stage information of the erosion channel are determined.

2. The method for processing remote sensing image data of erosion gullies and sloping farmland according to claim 1, characterized in that, Acquire multi-source remote sensing data of the study area, and based on the multi-source remote sensing data, acquire multi-band synthetic images and digital elevation models, including: Acquire multispectral and panchromatic images covering the study area; The multispectral image and the panchromatic image are fused to generate a fused image; The corresponding images of the near-infrared band, red-edge band, and short-wave infrared band in the fused image are assigned to the red, green, and blue channels respectively for false-color synthesis to obtain a composite image; Obtain the grayscale value of each pixel in the synthesized image, and determine the grayscale distribution data based on the grayscale value; Based on the grayscale distribution data, a low-end truncation threshold and a high-end truncation threshold are determined, wherein the low-end truncation threshold corresponds to a first preset percentage of the grayscale value distribution in the grayscale distribution data, and the high-end truncation threshold corresponds to a second preset percentage of the grayscale value distribution in the grayscale distribution data; Pixels with gray values ​​lower than the low-end truncation threshold are set to the minimum value of the preset gray level range, and pixels with gray values ​​higher than the high-end truncation threshold are set to the maximum value of the preset gray level range. Pixels with gray values ​​between the low-end truncation threshold and the high-end truncation threshold are linearly mapped to the preset gray level range, and the multi-band composite image is obtained.

3. The method for processing remote sensing image data of erosion gullies and sloping farmland according to claim 2, characterized in that, The multispectral image and the panchromatic image are fused to generate a fused image, including: Based on different bands in the multispectral image, a panchromatic simulated band is generated, wherein the panchromatic simulated band is a single-band raster image generated by linear combination of different bands of the multispectral image. Obtain the difference image between the panchromatic image and the panchromatic analog band based on different pixel gray values; The gray value of each pixel in the difference image is superimposed onto each band of the multispectral image according to a preset weighting coefficient to generate the fused image.

4. The method for processing remote sensing image data of erosion gullies and sloping farmland according to claim 3, characterized in that, The gray value of each pixel in the difference image is superimposed onto each band of the multispectral image according to a preset weighting coefficient to generate the fused image, including: The panchromatic image is divided into multiple non-overlapping image sub-blocks; Obtain the local variance of the gray values ​​of all pixels in each image sub-block, and generate a local variance matrix corresponding to the panchromatic image; Normalize each variance value in the local variance matrix to obtain a normalized variance value, wherein the normalization process is to linearly map the variance value to a preset weight value range. The normalized variance value is used as the initial weight coefficient at the corresponding spatial location to generate an initial weight coefficient matrix. Determine whether each band in the multispectral image is a near-infrared band or a short-wave infrared band; If so, the initial weighting coefficient is multiplied by a preset spectral protection factor to obtain the corrected weighting coefficient, and then the pixel gray value at the corresponding spatial location in the difference image is multiplied by the corrected weighting coefficient and superimposed onto the band. If not, the pixel gray value at the corresponding spatial location in the difference image is multiplied by the initial weighting coefficient and then directly superimposed onto the band.

5. A method for processing remote sensing image data of erosion gullies and sloping farmland according to claim 1, characterized in that, The multi-band composite image is input into a pre-built deep learning model, wherein the deep learning model is used to output, based on the input multi-band composite image, erosion channels and their corresponding boundaries, heads, walls, and bottoms, as well as the geometric parameters corresponding to the erosion channels, including: The multi-band synthetic image is input into the encoder network, and multi-scale feature maps are extracted through multiple cascaded convolutional layers and downsampling layers. The multi-scale feature maps include feature tensors of multiple levels with different spatial resolutions. The highest-level feature tensor is input into the global context attention module to generate a channel-weighted vector. The channel-weighted vector is then multiplied by the highest-level feature tensor to obtain the enhanced high-level semantic features. The feature tensors of each level in the multi-scale feature map, except for the highest level, are fused with the upsampled feature tensors in the corresponding decoder level through skip connections. Combined with the enhanced high-level semantic features, a semantic segmentation probability map with the same spatial size as the multi-band synthetic image is generated. Based on the semantic segmentation probability map, a classification label map containing the boundary, head region, wall region and bottom region of the erosion channel is generated by pixel-by-pixel classification judgment; Extract the binary mask data corresponding to the trench head region, the trench wall region, and the trench bottom region from the classification label map, respectively; Based on the binary mask data, at least one geometric parameter among the length, width, area, and depth of the erosion channel is determined.

6. The method for processing remote sensing image data of erosion gullies and sloping farmland according to claim 5, characterized in that, Based on the binary mask data, determining at least one geometric parameter among the length, width, area, and depth of the erosion channel includes: Extract the skeleton line of the binary mask data corresponding to the trench bottom region, and use the length of the skeleton line as the length of the erosion trench. In the binary mask data corresponding to the trench wall region, the distance between the two trench wall boundaries is calculated along the normal direction of the skeleton line, and the average value is taken to determine the width of the erosion channel; Based on the digital elevation model, the elevation values ​​of the bottom region and the top edge of the gully wall are obtained, and the elevation difference is used as the depth of the erosion channel.

7. The method for processing remote sensing image data of erosion gullies and sloping farmland according to claim 1, characterized in that, Based on the output of the water catchment network model and the deep learning model, the vectorized boundary and development stage information of the erosion channels are determined, including: The boundary of the erosion channel output by the deep learning model is obtained as the initial channel boundary, and the mask data of the channel head region, channel wall region and channel bottom region are determined. Obtain the cumulative runoff raster data and the flow direction raster data from the aforementioned runoff network model; The initial channel boundary is spatially superimposed with the flow accumulation raster data to determine the pixels inside the initial channel boundary whose flow accumulation is greater than a preset accumulation threshold, and these pixels are used as candidate pixels for the channel centerline. Based on the water flow direction grid data, determine the channel centerline vector with a continuous single pixel width; Based on the centerline vector of the channel, the vector is extended to both sides until the extension boundary touches the boundary of the channel wall region to generate the vectorized boundary of the erosion channel. Based on the mask data of the gully head region, the ratio of the gully head source erosion length to the gully head width is determined as an indicator of gully head activity. Based on the mask data of the trench wall region and the digital elevation model, the average slope of the trench wall region is determined as an indicator of trench wall stability. The catchment area of ​​the water-collecting zone flowing into the erosion channel is determined based on the water catchment network model and used as a water catchment dynamic indicator. The gully head activity index, the gully wall stability index, and the water collection dynamic index are weighted and combined to obtain the gully development stage index. Based on the comparison result of the gully development stage index and the preset classification threshold, the development stage information of the erosion gully is output. The development stage information includes active period, semi-stable period, or stable period.

8. A remote sensing image data processing system for eroded gullies and sloping farmland, characterized in that, The system includes: The first acquisition module is used to acquire multi-source remote sensing data of the study area, and acquire multi-band synthetic images and digital elevation models based on the multi-source remote sensing data. The second acquisition module is used to acquire the catchment network model based on the digital elevation model. The first output module is used to input the multi-band synthetic image into a pre-constructed deep learning model, wherein the deep learning model is used to output the erosion channel and its corresponding boundary, head, wall and bottom, as well as the geometric parameters corresponding to the erosion channel, based on the input multi-band synthetic image. The second output module is used to determine the vectorized boundary and development stage information of the erosion channel based on the output results of the water catchment network model and the deep learning model.

9. An electronic device, characterized in that, include: At least one processor; And, a memory communicatively connected to at least one of the processors; The memory stores instructions that can be executed by at least one of the processors, which are executed by at least one of the processors to enable the at least one of the processors to perform a remote sensing image data processing method for erosion gullies and sloping farmland as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements a method for processing remote sensing image data of erosion gullies and sloping farmland as described in any one of claims 1-7.

Images