A crop risk early warning method based on remote sensing data

Abstract

This invention discloses a crop risk early warning method based on remote sensing data, comprising the following steps: Step 1: Acquire multi-period remote sensing image sequences of crops and perform registration; Step 2: Extract the main perturbation period and period stability of pixels to generate a grayscale periodic mapping map; Step 3: Statistically analyze the perturbation entropy distribution to generate an abnormal texture region map; Step 4: Generate a structural mutation channel map; Step 5: Using the grayscale periodic mapping map, abnormal texture region map, and structural mutation channel map as input, perform spatial domain processing based on a multi-level morphological evolution convolutional fusion module with the InternImage model as the backbone, employing morphological extension and morphological compression branches to generate a fused feature map; Step 6: Generate a risk spatial diffusion prediction map; Step 7: Output a structured crop risk early warning report file. This invention integrates multi-period remote sensing time series analysis with a deep morphological evolution model to achieve accurate identification and early warning of crop risks.

Description

Technical Field

[0001] This invention relates to the field of agricultural risk early warning technology, and in particular to a crop risk early warning method based on remote sensing data. Background Technology

[0002] With the development of remote sensing technology, agricultural monitoring and risk assessment based on multi-period remote sensing data is becoming an important direction for agricultural informatization. Existing crop risk early warning methods usually rely on the spectral characteristics or statistical models of single-period images for regional-level analysis, which makes it difficult to capture potential temporal anomalies and spatial disturbance patterns during crop growth. These methods suffer from problems such as lagging risk identification and insufficient accuracy, and generally face the following technical challenges in practical applications: Remote sensing images exhibit significant temporal variations, but existing methods lack in-depth modeling of pixel-level grayscale time series, failing to effectively extract key time-frequency information on perturbation periods and stability. Texture feature extraction often employs fixed-directional gradient calculations, making it difficult to perceive the asymmetry of directional perturbation distribution, resulting in insensitivity to abnormal states such as pests, diseases, and drought. In terms of structural feature modeling, existing methods mostly use single-scale filtering, failing to fully utilize the multi-scale and multi-directional expressive power of Gabor responses, thus limiting the ability to detect changes in crop growth structure. Furthermore, the feature fusion process lacks modeling of morphological spatial structures, leading to weak expressive power in the fusion results and affecting the accuracy and spatial continuity of risk level identification.

[0003] Therefore, how to provide a crop risk early warning method based on remote sensing data is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] One objective of this invention is to propose a crop risk early warning method based on remote sensing data. This invention integrates multi-period remote sensing time series analysis, frequency domain perturbation modeling, structural mutation extraction, and deep learning feature fusion methods. It constructs a multi-level morphological evolution convolutional fusion module with the InternImage model as the backbone, and uses morphological extension and compression operations to strengthen edge and periodic features, thereby improving the ability to model spatiotemporal perturbation patterns. Then, it achieves pixel-level risk probability value prediction through a risk scoring module, and generates a risk spatial diffusion prediction map based on the K-means spatial clustering algorithm. Finally, it integrates plot boundary information to generate a field-level crop risk level map and a structured crop risk early warning report file, which has the advantages of high temporal sensitivity, strong spatial recognition accuracy, and clear risk expression.

[0005] A crop risk early warning method based on remote sensing data according to an embodiment of the present invention includes the following steps: Step 1: Obtain a multi-phase remote sensing image sequence of the crop, and register the multi-phase remote sensing image sequence to obtain a registered image sequence; Step 2: Based on the gray-level time series of each pixel in the registered image sequence, and perform frequency domain analysis, extract the main perturbation period and period stability of the pixel, and generate a gray-level period mapping map. Step 3: Calculate the directional gradient based on the registered image sequence, statistically analyze the perturbation entropy distribution, extract the asymmetric regions of the perturbation entropy distribution, and generate an abnormal texture region map; Step 4: Perform Gabor filtering on the images in the registered image sequence to construct a structural perturbation response map, and fuse it with the original registered image to generate a structural mutation channel map; Step 5: Using the grayscale periodic map, abnormal texture region map, and structural mutation channel map as input, and based on the multi-level morphological evolution convolutional fusion module with the InternImage model as the backbone, spatial domain processing is performed using morphological extension branch and morphological compression branch to generate fused feature map; Step 6: Generate a pixel-level risk probability map based on the fused feature map, and use the K-means algorithm for clustering to generate a risk spatial diffusion prediction map; Step 7: Overlay the risk spatial diffusion prediction map onto the crop plot boundary layer to generate a field-level crop risk level map and output a structured crop risk early warning report file.

[0006] Optionally, step one specifically includes: Acquire a multi-period remote sensing image sequence covering the same crop area. The multi-period remote sensing image sequence is acquired from the same satellite platform, with a spatial resolution of no more than 2 meters and an interval of 5 to 10 days between adjacent images. The earliest remote sensing image acquired in the time series was selected as the reference image, and the remaining remote sensing images were used as images to be registered. Extract a reference patch of size 21×21 pixels from the reference image, and extract candidate patches of the same size by sliding them in each image to be registered. Calculate the normalized cross-correlation coefficient for the reference patch and each candidate patch separately; The candidate patch position with the largest normalized cross-correlation coefficient is selected as the integer pixel-level matching position; The image to be registered is translated and transformed according to the integer pixel-level matching positions to generate a registered image; All registered remote sensing images are arranged in chronological order to construct a registered image sequence.

[0007] Optionally, step two specifically includes: Extract the grayscale value of each pixel in all periods from the registered image sequence to form the grayscale time series corresponding to the pixel; Perform a Fast Fourier Transform on each grayscale time series to obtain a spectrum vector, which includes the amplitude of each frequency. In the spectrum vector, the component with amplitude corresponding to zero frequency is removed, and the frequency component with the largest amplitude is selected from the remaining frequency components. The reciprocal of the corresponding frequency is taken as the main perturbation period of the pixel. The ratio of the amplitude of the frequency corresponding to the main disturbance period to the sum of the amplitudes of all non-zero frequencies is used as the periodic stability of the pixel. The main disturbance period and periodic stability are mapped to grayscale values ​​and transparency, respectively, and combined to generate a grayscale periodic mapping map. In the grayscale periodic inversion mapping map, the grayscale value of each pixel position represents the main disturbance period, and the transparency represents the periodic stability.

[0008] Optionally, step three specifically includes: Each image is selected from the registered image sequence, and the Sobel operator is used to calculate the grayscale change value of each pixel in the horizontal and vertical directions respectively. The gradient magnitude of a pixel is obtained by adding the squares of the horizontal and vertical grayscale changes of each pixel and taking the square root. Divide the vertical grayscale change value by the horizontal grayscale change value, and take the result of the arctangent function to obtain the gradient direction angle of the pixel; The gradient direction of all pixels is divided according to a preset direction range, which is from 0 degrees to 180 degrees and divided into multiple equally spaced direction segments. The gradient magnitude of the corresponding pixel is recorded in each direction segment. For each directional segment, the probability distributions containing pixel gradient magnitudes are statistically analyzed, and each probability value is multiplied by the logarithm of the probability value, then summed and negative to calculate the perturbation entropy of the corresponding directional segment. The perturbation entropy of each directional segment is combined in angular order to form a perturbation entropy distribution vector; The perturbation entropy distribution vector is paired and compared in symmetrical directions. If the difference in perturbation entropy between two directional segments exceeds a preset tolerance threshold, it is marked as an asymmetrical region of perturbation entropy distribution. Combine the pixel locations corresponding to all asymmetrical regions of perturbation entropy distribution to generate an abnormal texture region map.

[0009] Optionally, step four specifically involves: Gabor filtering is performed on each image in the registered image sequence to construct a multi-scale, multi-directional Gabor filter set; Convolution operations are performed on the Gabor filter set for each image to obtain the response values ​​of each pixel at different scales and directions, resulting in a total of 30 filter response layers; For each pixel, the response values ​​in the 30 filtered response layers are combined into a structural response vector of length 30. In the time dimension, the standard deviation of the structural response vector of the corresponding pixel in all periods is calculated to form a structural perturbation response map. The structural disturbance response map is fused with the original registration image of each period according to pixel position to generate a structural mutation layer; Combine all structural mutation layers in chronological order to form a structural mutation channel map.

[0010] Optionally, the Gabor filter set specifically includes 5 filter groups, each corresponding to a spatial frequency of 0.05, 0.1, 0.2, 0.3 and 0.4 respectively. Each filter group includes 6 directional filters with directions of 0 degrees, 30 degrees, 60 degrees, 90 degrees, 120 degrees and 150 degrees respectively. For each image in the registered image sequence, a Gabor filter set is applied sequentially for convolution processing to obtain the response results of each pixel under different combinations of directions and frequencies. The response results are in the complex form of the Gabor convolution output. The square of the real part and the square of the imaginary part are added together and the square root is taken to obtain the response value of each pixel under the current direction and spatial frequency, which is used as the structural response intensity.

[0011] Optionally, step five specifically includes: The grayscale periodic map, abnormal texture region map, and structural mutation channel map are aligned according to pixel position and then stitched together along the channel dimension to form the initial input feature map. The initial input feature map is input into the InternImage model for feature extraction. The InternImage model includes four stages: Stage 1, Stage 2, Stage 3 and Stage 4. Stage 1 receives the initial input feature map, which has 64 input channels and is constructed by stacking two InternImage Blocks sequentially. Stage 2 consists of two InternImage Blocks with 64 input channels and 128 output channels. It is connected from Stage 1 to Stage 2 through a downsampling convolution with a stride of 2. Stage 3 consists of three InternImage Blocks with 128 input channels and 256 output channels. It is connected from Stage 2 to Stage 3 through downsampling convolution with a stride of 2. Stage 4 consists of three InternImage Blocks with 256 input channels and 512 output channels. It is connected from Stage 3 to Stage 4 through a downsampling convolution with a stride of 2. The multi-stage morphological evolution convolutional fusion module is inserted between Stage 2 and Stage 3; The multi-stage morphological evolution convolutional fusion module consists of three sequentially connected morphological evolution blocks. Each morphological evolution block includes a morphological extension branch, a morphological compression branch, and a regular convolution branch. The three branches of the first morphological evolution block receive the same Stage2 output feature map as the evolution input feature map. In the morphological extension branch, three sliding windows of 3×3, 5×5 and 7×7 are used to perform a pixel-by-pixel maximum value replacement operation on the evolutionary input feature map, generating extended feature maps of three scales, and then splicing them in the channel dimension. In the morphological extrusion branch, a sliding window of the same size as the morphological extension branch is used to perform a pixel-by-pixel minimum value replacement operation on the evolutionary input feature map, generating extrusion feature maps of three scales, which are then stitched together in the channel dimension. In the regular convolution branch, a 3×3 standard convolution is performed on the evolving input feature map to generate a convolutional feature map; The outputs of the three branches are concatenated along the channel dimension, input into a 1×1 convolution kernel for channel compression, and then passed through a normalization unit and an activation unit in sequence to generate the evolution output feature map of the current morphological evolution block. The evolutionary output feature map of the current morphological evolution block is added element by element to the evolutionary input feature map to form a residual connection, which is used as the input for the next morphological evolution block. The three morphological evolution blocks are connected in sequence. The third morphological evolution block outputs a morphological feature map, which is used as the input of Stage 3. Stage 4 outputs a fused feature map.

[0012] Optionally, the step of generating a pixel-level risk probability map based on the fused feature map specifically involves: The fused feature map is input into the risk scoring module, which consists of two convolutional layers and corresponding activation functions. The first convolutional layer uses a 3×3 kernel with a stride of 1 and a padding size of 1. The number of input channels is consistent with the fused feature map, and the number of output channels is set to 128. The second convolutional layer uses a 1×1 kernel with a stride of 1, padding size of 0, and the number of output channels is set to 1. The output of the first convolutional layer is followed by a ReLU activation function to truncate all negative values; The output of the second convolutional layer is followed by a Sigmoid activation function, which maps the convolutional output value of each pixel position to between 0 and 1, generating a risk probability value for the corresponding pixel position. The risk probability values ​​for all pixel locations are combined to generate a pixel-level risk probability map.

[0013] Optionally, the step of using the K-means algorithm for clustering to generate a risk spatial diffusion prediction map specifically involves: Read the risk probability value of each pixel location from the pixel-level risk probability map, and record the row and column coordinates of the pixel in the image. Construct each pixel into a three-dimensional feature vector including row coordinates, column coordinates and risk probability value. The three-dimensional feature vectors corresponding to all pixels are used as a clustering sample set and input into the K-means clustering algorithm. The number of clusters K is set to 3, representing low risk, medium risk and high risk respectively. All cluster samples are sorted from smallest to largest according to their risk probability values. The mean of the cluster samples with risk probability values ​​in the top 10% is used as the first cluster center, the mean of the cluster samples with risk probability values ​​in the bottom 40% to 60% range is used as the second cluster center, and the mean of the cluster samples with risk probability values ​​in the bottom 10% is used as the third cluster center, thus completing the initial cluster center setting. In each round of clustering iteration, the Euclidean distance between each cluster sample and the three cluster centers is calculated sequentially, and the cluster sample is assigned to the category corresponding to the cluster center with the smallest Euclidean distance. After classifying all clustered samples, the row coordinates, column coordinates, and risk probability values ​​of all clustered samples within each class are averaged to update the cluster center position of the corresponding class. Repeat the process of dividing cluster samples and determining cluster center locations until the changes in cluster center locations in two consecutive rounds are less than a preset threshold, or the preset maximum number of iterations is reached, at which point the clustering process stops. Based on the final clustering results, the category corresponding to each pixel is mapped back to the original spatial location to generate a risk spatial diffusion prediction map with the same size as the pixel-level risk probability map.

[0014] Optionally, step seven specifically includes: The risk spatial diffusion prediction map is overlaid on the crop plot boundary layer according to pixel location. Based on the proportion of risk categories of pixels contained within each plot boundary, the risk category with the highest proportion is used as the risk level label of the plot, and a plot-level crop risk level map is generated. The information on field numbers, spatial locations, risk level labels, number of pixels and percentage of each risk level is compiled into structured entries and then summarized to generate a structured crop risk warning report file. The structured crop risk warning report file is in CSV or Excel spreadsheet format, with each field recording the corresponding field data on a separate line.

[0015] The beneficial effects of this invention are: This invention constructs a multi-level morphological evolution convolutional fusion module with the InternImage model as its backbone structure. Addressing the multimodal heterogeneity of crop growth perturbations in remote sensing images, it fuses grayscale periodic maps, anomalous texture region maps, and structural abrupt change channels. Figure 3 For spatial feature information, multi-scale sliding windows are used in the morphological extension and compression branches to replace the maximum and minimum values ​​of each pixel, enhancing the expressive power of structural response to perturbation areas at different scales. Residual connections maintain spatial semantic continuity, effectively improving the boundary sensitivity and expression accuracy of the fused feature map for high-risk areas. In the pixel-level risk probability map generation stage, a risk scoring module consisting of two convolutional layers is constructed. The first convolutional layer is followed by a ReLU activation function to suppress background interference, and the second convolutional layer is followed by a Sigmoid activation function to complete the pixel-level risk probability normalization output. In the risk spatial diffusion prediction map generation process, a three-dimensional feature vector containing spatial location and risk probability is used as the input for K-means clustering. Initial cluster centers are set based on the risk probability ranking interval, guiding the clustering process to quickly converge to spatially coherent risk areas. Finally, the risk spatial diffusion prediction map is superimposed on the crop plot boundary layer, and a plot-level crop risk level map is generated based on the pixel proportion within the plot. A structured crop risk warning report is output, comprehensively improving the spatial resolution and decision-making usability of crop risk identification driven by remote sensing data. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of a crop risk early warning method based on remote sensing data proposed in this invention; Figure 2 This is a schematic diagram of the multi-stage morphological evolution convolutional fusion module structure of a crop risk early warning method based on remote sensing data proposed in this invention, with the InternImage model as the main structure. Figure 3 This is a flowchart illustrating the K-means clustering method for generating a spatial diffusion prediction map of crop risk based on remote sensing data, as proposed in this invention. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0018] refer to Figures 1-3 A crop risk early warning method based on remote sensing data includes the following steps: Step 1: Obtain a multi-phase remote sensing image sequence of the crop, and register the multi-phase remote sensing image sequence to obtain a registered image sequence; Step 2: Based on the gray-level time series of each pixel in the registered image sequence, and perform frequency domain analysis, extract the main perturbation period and period stability of the pixel, and generate a gray-level period mapping map. Step 3: Calculate the directional gradient based on the registered image sequence, statistically analyze the perturbation entropy distribution, extract the asymmetric regions of the perturbation entropy distribution, and generate an abnormal texture region map; Step 4: Perform Gabor filtering on the images in the registered image sequence to construct a structural perturbation response map, and fuse it with the original registered image to generate a structural mutation channel map; Step 5: Using the grayscale periodic map, abnormal texture region map, and structural mutation channel map as input, and based on the multi-level morphological evolution convolutional fusion module with the InternImage model as the backbone, spatial domain processing is performed using morphological extension branch and morphological compression branch to generate fused feature map; Step 6: Generate a pixel-level risk probability map based on the fused feature map, and use the K-means algorithm for clustering to generate a risk spatial diffusion prediction map; Step 7: Overlay the risk spatial diffusion prediction map onto the crop plot boundary layer to generate a field-level crop risk level map and output a structured crop risk early warning report file.

[0019] In this embodiment, step one specifically includes: Acquire a multi-period remote sensing image sequence covering the same crop area. The multi-period remote sensing image sequence is acquired from the same satellite platform, with a spatial resolution of no more than 2 meters and an interval of 5 to 10 days between adjacent images. The earliest remote sensing image acquired in the time series was selected as the reference image, and the remaining remote sensing images were used as images to be registered. Extract a reference patch of size 21×21 pixels from the reference image, and extract candidate patches of the same size by sliding them in each image to be registered. Calculate the normalized cross-correlation coefficient for the reference tile and each candidate tile separately. Specifically: Calculate the average grayscale value of all pixels in the reference patch and the candidate patch respectively; Subtract the average value of the corresponding patch from the gray value of each pixel to obtain the zero-mean patch of the two patches; Multiply the corresponding pixel gray values ​​in the two zero-mean patches and sum them to get the numerator; Sum the squares of all pixel grayscale values ​​in the two zero-mean patches and take the square root. Multiply the two square roots and use the result as the denominator. Divide the numerator by the denominator to obtain the normalized cross-correlation coefficient. The candidate patch position with the largest normalized cross-correlation coefficient is selected as the integer pixel-level matching position; The image to be registered is translated and transformed according to the integer pixel-level matching positions to generate a registered image; All registered remote sensing images are arranged in chronological order to construct a registered image sequence; This invention extracts reference patches of fixed size from a baseline image and slides to extract candidate patches in the image to be registered. It then uses normalized cross-correlation coefficients to measure the similarity of each pair of patches and selects the position with the highest matching score to achieve integer pixel-level registration. This registration method eliminates the need for complex sub-pixel interpolation calculations, reducing computational burden and improving processing efficiency. It is suitable for rapid preprocessing of large-scale remote sensing image data. By arranging all registered remote sensing image sequences in chronological order, the consistency and accuracy of time-series feature extraction are ensured, providing a stable and reliable data foundation for periodic disturbance analysis, texture feature extraction, and risk prediction.

[0020] In this embodiment, step two specifically includes: Extract the grayscale value of each pixel in all periods from the registered image sequence to form the grayscale time series corresponding to the pixel; Perform a Fast Fourier Transform on each grayscale time series to obtain a spectrum vector, which includes the amplitude of each frequency. In the spectrum vector, the component with amplitude corresponding to zero frequency is removed, and the frequency component with the largest amplitude is selected from the remaining frequency components. The reciprocal of the corresponding frequency is taken as the main perturbation period of the pixel. The ratio of the amplitude of the frequency corresponding to the main disturbance period to the sum of the amplitudes of all non-zero frequencies is used as the periodic stability of the pixel. The main perturbation period and periodic stability are mapped to grayscale values ​​and transparency, respectively, and combined to generate a grayscale periodic mapping map. In the grayscale periodic inversion mapping map, the grayscale value of each pixel represents the main perturbation period, and the transparency represents the periodic stability. Specifically, the minimum value of the main perturbation period in the entire image is used as the scaling start point, and the maximum value of the main perturbation period is used as the scaling end point. The main perturbation period of each pixel is proportionally mapped to an integer value between 0 and 255, which serves as the pixel grayscale value of the grayscale periodic mapping map. The minimum value of the periodic stability is used as the transparency start point, and the maximum value of the periodic stability is used as the transparency end point. The periodic stability of each pixel is proportionally mapped to an integer transparency value between 0 and 255. By constructing a grayscale time series for each pixel in the registered image sequence and extracting frequency domain features using Fast Fourier Transform, this invention can efficiently obtain the main perturbation period and periodic stability of the surface reflectance of ground objects over time. The spectral vector contains a component with an amplitude corresponding to zero frequency, representing the average value of the original pixel grayscale time series. Since the zero-frequency component does not carry periodic perturbation information, the non-zero frequency with the largest amplitude is identified after removing the zero-frequency component, and its reciprocal is used as the main perturbation period. Furthermore, its amplitude proportion in the overall spectrum is used as the periodic stability, accurately reflecting the perturbation periodic characteristics caused by environmental stress, water stress, or pests and diseases during crop growth. Mapping the main perturbation period to image grayscale values ​​and the periodic stability to image transparency not only preserves the spatial differences in periodic changes but also expresses the regularity and reliability of the perturbation. The generated grayscale periodic mapping map provides a unified data representation basis for texture analysis, structural anomaly detection, and risk clustering, significantly improving the model's accuracy in identifying abnormal regions and its ability to represent temporal stability.

[0021] In this embodiment, step three specifically includes: Each image is selected from the registered image sequence, and the Sobel operator is used to calculate the grayscale change value of each pixel in the horizontal and vertical directions respectively. The gradient magnitude of a pixel is obtained by adding the squares of the horizontal and vertical grayscale changes of each pixel and taking the square root. Divide the vertical grayscale change value by the horizontal grayscale change value, and take the result of the arctangent function to obtain the gradient direction angle of the pixel; The gradient direction of all pixels is divided according to a preset direction range, which is from 0 degrees to 180 degrees and divided into multiple equally spaced direction segments. The gradient magnitude of the corresponding pixel is recorded in each direction segment. For each directional segment, the probability distributions containing pixel gradient magnitudes are statistically analyzed, and each probability value is multiplied by the logarithm of the probability value, then summed and negative to calculate the perturbation entropy of the corresponding directional segment. The perturbation entropy of each directional segment is combined in angular order to form a perturbation entropy distribution vector; The perturbation entropy distribution vector is paired and compared in symmetrical directions. If the difference in perturbation entropy between two directional segments exceeds a preset tolerance threshold, it is marked as an asymmetrical region of perturbation entropy distribution. The preset tolerance threshold is 10% to 15% of the theoretical maximum value of perturbation entropy, and the specific value range is 0.1 to 0.3. Combine the pixel locations corresponding to all regions with asymmetrical perturbation entropy distribution to generate an abnormal texture region map; In remote sensing crop monitoring, abnormal texture area maps can effectively reveal abnormal growth phenomena in crop areas caused by pests and diseases, drought, waterlogging, lodging, or human interference. These anomalies often manifest as directional changes in the local structure of the image, such as disordered texture direction and uneven edge gradient changes. Conventional spectral or time series analysis methods are difficult to accurately capture these structural disturbances. However, by performing entropy statistical analysis on the gradient amplitude of each pixel in multiple directions in the registered image sequence, the stability and complexity of texture information in each direction can be comprehensively quantified. Furthermore, by comparing the symmetry of the perturbation entropy distribution, regions with significantly uneven directional distribution can be identified, thereby determining the potential risk of texture anomalies in these regions during temporal evolution. The resulting abnormal texture area map can intuitively present risk areas with prominent texture changes in the form of a spatial distribution image, providing important structural discrimination criteria and spatial positioning references for structural feature fusion, risk probability map construction, and crop early warning.

[0022] In this embodiment, step four specifically includes: Gabor filtering is performed on each image in the registered image sequence to construct a multi-scale, multi-directional Gabor filter set; Convolution operations are performed on the Gabor filter set for each image to obtain the response values ​​of each pixel at different scales and directions, resulting in a total of 30 filter response layers; For each pixel, the response values ​​in the 30 filtered response layers are combined into a structural response vector of length 30. In the time dimension, the standard deviation of the structural response vector of the corresponding pixel in all periods is calculated to form a structural perturbation response map. The structural disturbance response map is fused with the original registration image of each period according to pixel position. The fusion method is as follows: the gray value of each pixel in the original registration image is normalized and then linearly superimposed with the structural disturbance response value at the corresponding position according to a set weighting ratio (0.6:0.4) to generate a structural mutation layer. All structural mutation layers are combined in chronological order to form a structural mutation channel map, which is used to characterize the spatial structural perturbation of crops in multi-period images. This invention mines structural perturbation features during crop growth from remote sensing images, enhancing the ability to identify spatial variations in crops. In traditional image analysis, relying solely on grayscale, color, or texture information often fails to accurately capture directional changes at multiple scales. However, by introducing a multi-scale, multi-directional Gabor filter set, the local structural responses of images at different scales and directions can be effectively extracted, constructing structural response vectors that reflect the local structural attributes of each pixel in space.

[0023] Furthermore, calculating the standard deviation of the structural response vector in the time dimension can reflect the magnitude of structural changes of pixels throughout the entire time series, forming a structural perturbation response map, which is used to characterize the stability and variability of crop regions at different stages. This map is then fused with the original image to generate a structural mutation layer, which not only preserves the original details of the image but also enhances the expressive ability of the mutation region.

[0024] The resulting structural mutation channel map has good spatial positioning ability and temporal sensitivity, and can serve as an important basis for judging crop growth abnormalities, pests and diseases, lodging and premature aging risk events, providing high-quality structural feature support for risk probability modeling and level assessment.

[0025] In this embodiment, the Gabor filter set specifically includes 5 filter groups, each corresponding to a spatial frequency of 0.05, 0.1, 0.2, 0.3 and 0.4 respectively. Each filter group includes 6 directional filters with directions of 0 degrees, 30 degrees, 60 degrees, 90 degrees, 120 degrees and 150 degrees respectively. For each image in the registered image sequence, Gabor filter set is applied sequentially for convolution processing to obtain the response result of each pixel under different combinations of directions and frequencies. The response result is the complex form of the Gabor convolution output. The square of the real part and the square of the imaginary part are added together and the square root is taken to obtain the response value of each pixel under the current direction and spatial frequency, which is used as the structural response intensity. The Gabor filter is a linear filter that simulates the receptive field of human visual cortex cells. It is widely used for edge detection and texture analysis of images. Its core is a bandpass filter kernel formed by combining a sine wave and a Gaussian function. It can simultaneously extract spatial location and frequency information and has good time-frequency localization characteristics. In this invention, the Gabor filter used adopts the two-dimensional Gabor function commonly used in the prior art. Its real part is a cosine wave modulated by a Gaussian function, and its imaginary part is a sine wave modulated by a Gaussian function. The convolution output is a complex response. The real and imaginary parts together reflect the structural features of the image at specific directions and frequencies. The response value intensity is calculated by taking the square root of the sum of the square of the real part and the square of the square of the imaginary part. It has the advantages of strong directional selectivity and high frequency resolution, which helps to capture small structural perturbations and anomalies in crop areas.

[0026] In this embodiment, step five specifically includes: The grayscale periodic map, abnormal texture region map, and structural mutation channel map are aligned according to pixel position and then stitched together along the channel dimension to form the initial input feature map. The initial input feature map is input into the InternImage model for feature extraction. The InternImage model includes four stages: Stage 1, Stage 2, Stage 3 and Stage 4. Stage 1 receives the initial input feature map, which has 64 input channels and is constructed by stacking two InternImage Blocks sequentially. Stage 2 consists of two InternImage Blocks with 64 input channels and 128 output channels. It is connected from Stage 1 to Stage 2 through a downsampling convolution with a stride of 2. Stage 3 consists of three InternImage Blocks with 128 input channels and 256 output channels. It is connected from Stage 2 to Stage 3 through downsampling convolution with a stride of 2. Stage 4 consists of three InternImage Blocks with 256 input channels and 512 output channels. It is connected from Stage 3 to Stage 4 through a downsampling convolution with a stride of 2. In this invention, the InternImage Block, as the basic building block of the InternImage model, is mainly composed of four functional units: a dynamic convolution kernel parameter generation unit, a deformable convolution operation unit, a normalization unit, and an activation unit. These units are connected sequentially to form a complete data processing path. The dynamic convolutional kernel parameter generation unit employs a lightweight attention mechanism. Specifically, the input feature map first undergoes a global average pooling operation to obtain a statistical description vector for each channel. This vector then enters a sub-network consisting of two fully connected layers. The first layer has one-quarter the number of input channels and uses ReLU activation. The second layer has the original number of channels and uses Sigmoid activation, which outputs normalized channel weights. Finally, this weight vector is used to scale the weights of each channel convolutional kernel in the deformable convolution, achieving dynamic adjustment. The deformable convolutional operation unit uses a 7×7 deformable convolutional kernel with a stride of 1 and padding of 3. Its structure consists of two parts: the first part is the prediction module, which performs a 3×3 convolution operation on the input feature map to predict the offset of each sampling point; the second part samples the input feature map according to the predicted offset position and performs point-by-point convolution operation in combination with the aforementioned channel-weighted dynamic convolutional kernel parameters to output the feature map. The normalization unit uses the Batch Normalization operation to standardize the mean and variance of the feature map in each training batch to prevent gradient vanishing and model training instability. After normalization, it is multiplied by the learnable scaling parameter and the bias parameter is added. The activation unit adopts the GELU activation function, which has continuous and differentiable nonlinear mapping capabilities and has better convergence and representation capabilities than ReLU. The InternImage Block, which combines the above four units, can realize flexible modeling and efficient expression of spatial structure and enhance the quality of feature extraction. The multi-stage morphological evolution convolutional fusion module is inserted between Stage 2 and Stage 3; The multi-stage morphological evolution convolutional fusion module consists of three sequentially connected morphological evolution blocks. Each morphological evolution block includes a morphological extension branch, a morphological compression branch, and a regular convolution branch. The three branches of the first morphological evolution block receive the same Stage2 output feature map as the evolution input feature map. In the morphological extension branch, three sliding windows of 3×3, 5×5 and 7×7 are used to perform a pixel-by-pixel maximum value replacement operation on the evolutionary input feature map, generating extended feature maps of three scales, and then splicing them in the channel dimension. In the morphological extrusion branch, a sliding window of the same size as the morphological extension branch is used to perform a pixel-by-pixel minimum value replacement operation on the evolutionary input feature map, generating extrusion feature maps of three scales, which are then stitched together in the channel dimension. In the regular convolution branch, a 3×3 standard convolution is performed on the evolving input feature map to generate a convolutional feature map; The outputs of the three branches are concatenated along the channel dimension, input into a 1×1 convolution kernel for channel compression, and then passed through a normalization unit and an activation unit in sequence to generate the evolution output feature map of the current morphological evolution block. The evolutionary output feature map of the current morphological evolution block is added element by element to the evolutionary input feature map to form a residual connection, which is used as the input for the next morphological evolution block. The three morphological evolution blocks are connected in sequence. The third morphological evolution block outputs a morphological feature map, which is used as the input of Stage 3. Stage 4 outputs a fused feature map. In this invention, the multi-stage morphological evolution convolutional fusion module is embedded between Stage 2 and Stage 3 of the InternImage model. It is mainly based on the structural design optimization of the expression rules of crop growth perturbation features in remote sensing images and the distribution of model perception capabilities. As a hierarchical visual backbone network, the InternImage model mainly extracts low-level edge and texture features in Stage 1, while Stage 4 focuses on high-level abstract semantic features. Stage 2 and Stage 3 are in the middle semantic level of the model and have the ability to perceive local details and preliminary global structure at the same time. They are the key stages for fusing periodic perturbation, texture anomaly and structural mutation information.

[0027] Inserting a multi-stage morphological evolution convolutional fusion module between Stage 2 and Stage 3 fully leverages the model's ability to still resolve details, enhancing its perception and response to spatial perturbation boundaries, complex periodic changes, and mesoscale structural abrupt changes. The multi-stage morphological evolution convolutional fusion module performs maximum expansion on spatial edges through a morphological extension branch, improving boundary sensitivity; it performs minimum compression on noisy regions through a morphological compression branch, suppressing weak background perturbations; and combined with a regular convolutional branch, it maintains the expression of the original semantic features. The joint structure of the three branches can significantly improve the discriminative power of the fused features.

[0028] Furthermore, this embedding position helps to inherit the enhanced intermediate features when further modeling deeper semantics in Stage 3 and Stage 4, effectively promoting the deep semantic unfolding of the fusion graph. This not only improves the spatial resolution and multi-feature coupling ability of the overall model, but also ensures training stability and computational efficiency, demonstrating clear engineering practicality and technological innovation.

[0029] In this embodiment, the generation of a pixel-level risk probability map based on the fused feature map specifically involves: The fused feature map is input into the risk scoring module, which consists of two convolutional layers and corresponding activation functions. The first convolutional layer uses a 3×3 kernel with a stride of 1 and a padding size of 1. The number of input channels is consistent with the fused feature map, and the number of output channels is set to 128. The second convolutional layer uses a 1×1 kernel with a stride of 1, padding size of 0, and the number of output channels is set to 1. The output of the first convolutional layer is followed by a ReLU activation function to truncate all negative values; The output of the second convolutional layer is followed by a Sigmoid activation function, which maps the convolutional output value of each pixel position to between 0 and 1, generating a risk probability value for the corresponding pixel position. The risk probability values ​​of all pixel locations are combined to generate a pixel-level risk probability map, which is used to characterize the spatial risk distribution of crop areas.

[0030] In this embodiment, the step of using the K-means algorithm for clustering to generate a risk spatial diffusion prediction map specifically involves: Read the risk probability value of each pixel location from the pixel-level risk probability map, and record the row and column coordinates of the pixel in the image. Construct each pixel into a three-dimensional feature vector including row coordinates, column coordinates and risk probability value. The three-dimensional feature vectors corresponding to all pixels are used as a clustering sample set and input into the K-means clustering algorithm. The number of clusters K is set to 3, representing low risk, medium risk and high risk respectively. All cluster samples are sorted from smallest to largest according to their risk probability values. The mean of the cluster samples with risk probability values ​​in the top 10% is used as the first cluster center, the mean of the cluster samples with risk probability values ​​in the bottom 40% to 60% range is used as the second cluster center, and the mean of the cluster samples with risk probability values ​​in the bottom 10% is used as the third cluster center, thus completing the initial cluster center setting. In each round of clustering iteration, the Euclidean distance between each cluster sample and the three cluster centers is calculated sequentially, and the cluster sample is assigned to the category corresponding to the cluster center with the smallest Euclidean distance. After classifying all clustered samples, the row coordinates, column coordinates, and risk probability values ​​of all clustered samples within each class are averaged to update the cluster center position of the corresponding class. Repeat the process of dividing cluster samples and determining cluster center locations until the changes in cluster center locations in two consecutive rounds are less than a preset threshold, or the preset maximum number of iterations is reached, at which point the clustering process stops. Based on the final clustering results, the category corresponding to each pixel is mapped back to the original spatial location to generate a risk spatial diffusion prediction map with the same size as the pixel-level risk probability map. In generating the risk spatial diffusion prediction map, the K-means clustering algorithm is used to divide regions of different risk levels. This method constructs each pixel in the pixel-level risk probability map into a three-dimensional feature vector including row coordinates, column coordinates, and risk probability values, effectively integrating spatial information and risk intensity information, and improving the discriminative ability of clustering. Initial three cluster centers are set to ensure good distinguishability. During clustering, standard Euclidean distance is used as a similarity measure between cluster samples and cluster centers. Combined with an iterative update strategy, category division and cluster center position adjustments are continuously performed until the change in cluster center position is less than 0.01 for two consecutive rounds, or until the preset maximum number of iterations (100 rounds) is reached. Finally, the clustering result assigns each pixel to a corresponding risk category and restores it according to its spatial location, generating a result consistent with the original. Figure 1 The resulting risk spatial diffusion prediction map provides support for regional risk early warning and analysis.

[0031] In this embodiment, step seven specifically includes: The risk spatial diffusion prediction map is overlaid on the crop plot boundary layer according to pixel location. Based on the proportion of risk categories of pixels contained within each plot boundary, the risk category with the highest proportion is used as the risk level label of the plot, and a plot-level crop risk level map is generated. The information on field numbers, spatial locations, risk level labels, number of pixels and percentage of each risk level is compiled into structured entries and then summarized to generate a structured crop risk warning report file. The structured crop risk warning report file is in CSV or Excel spreadsheet format, with each field recording the corresponding field data on a separate line.

[0032] Example 1: To verify the feasibility of this invention in practice, it was applied to a crop risk monitoring and early warning task in a typical maize-growing area. A multi-period remote sensing image sequence covering a large area of ​​maize fields was selected as input, spanning two consecutive months with a 7-day image interval and a spatial resolution of 1.5 meters. The remote sensing images were acquired from the same orbital satellite platform to ensure consistent image quality. During this period, the study area was affected by continuous rainfall and sudden severe convective weather, leading to abnormal crop growth in some plots and exhibiting typical risk characteristics.

[0033] In this embodiment, the acquired multi-period remote sensing image sequence is first registered. A 21×21 pixel reference patch is used to perform normalized cross-correlation matching with the image to be registered to achieve spatial alignment, with the registration accuracy controlled within ±1 pixel. Subsequently, the main perturbation period and period stability of each pixel are extracted by fast Fourier transform to generate a grayscale period map. It can be seen that the main perturbation period is abnormally prolonged and the period stability is decreased in some areas. Combining image gradient analysis and perturbation entropy calculation, regions with asymmetrical perturbation entropy distribution are further identified, generating an abnormal texture region map, which clearly marks the edge texture abrupt changes in the affected areas. A structural perturbation response map is constructed by using a Gabor filter. The structural perturbation response map is fused with the original registered image of each period according to pixel position to generate a structural abrupt change layer. All structural abrupt change layers are combined in chronological order to form a structural abrupt change channel map.

[0034] The aforementioned grayscale periodic mapping map, abnormal texture region map, and structural mutation channel map are input into a multi-level morphological evolution convolutional fusion module based on the InternImage model backbone structure. In the morphological extension branch and morphological compression branch, the maximum and minimum values ​​of sliding windows at different scales are replaced to enhance the difference between the region edges and centers. Basic spatial texture is extracted in the regular convolutional branch. After the three branches are fused, the channels are further compressed and residual connections are introduced to improve feature stability and expressive power, generating a fused feature map. The fused feature map is input into the risk scoring module, processed by 3×3 and 1×1 convolutional layers, and then the risk probability value between 0 and 1 for each pixel is output through ReLU and Sigmoid activation functions to generate a pixel-level risk probability map.

[0035] To further identify spatially identifiable risk areas, K-means clustering was used to divide the three-dimensional feature vectors (row coordinates, column coordinates, and risk probability values) into high, medium, and low categories. The mean values ​​of the top 10%, the middle 40%-60%, and the bottom 10% of samples were used as the initial cluster centers. After approximately seven rounds of iteration and convergence, a risk spatial diffusion prediction map was generated, clearly marking the distribution boundaries of high-risk areas. Based on this, the map was overlaid with a plot boundary layer. According to the proportion of risk pixels within each plot, a risk level label was assigned to the corresponding plot, and a structured crop risk warning report file (Excel format) was output. This report included plot number, spatial location, risk level label, and information on the number and proportion of pixels at each risk level, organized into structured entries to provide agricultural management departments with intuitive decision-making support.

[0036] To verify the technical effectiveness of the present invention, it was compared with the traditional convolutional neural network method that does not introduce morphological evolution structure. Multiple risk identification accuracy evaluation indicators were set, and the results are shown in Table 1.

[0037] Table 1. Comparison of risk identification performance between traditional convolution methods and the method of this invention. Table 1 above shows a performance comparison between the traditional convolutional method and the method of this invention in crop risk identification. In terms of pixel-level risk identification accuracy, the method of this invention reaches 91.1%, a significant improvement over the 84.7% of the traditional convolutional method. This indicates that the deep feature fusion of the InternImage model's backbone structure and morphological evolution block effectively enhances the ability to capture fine-grained risk information in remote sensing images. Regarding high-risk area discrimination, the accuracy of this invention is 87.5%, significantly better than the 76.3% of the traditional convolutional method, demonstrating the superior performance of this invention in identifying extremely or significantly disturbed areas. It possesses stronger discrimination capabilities; the accuracy rate for medium-risk areas has also improved to 84.9%, an increase of 11.3 percentage points compared to the 73.6% of traditional convolution methods, reflecting the stable performance of this invention even in ambiguous or unclear risk areas; in addition, in terms of the consistency rate of field-level risk level labels, this invention achieves an accuracy rate of 92.6%, far exceeding the 79.2% of traditional convolution methods, indicating that this invention exhibits higher overall consistency and practicality in spatial scale integration and regional clustering result mapping, fully verifying the technical advantages and application prospects of this invention in remote sensing data-driven crop risk early warning scenarios.

[0038] This embodiment constructs a multi-stage morphological evolution convolutional fusion module with the InternImage model as its backbone, fusing grayscale periodic maps, abnormal texture region maps, and structural mutation channels. Figure 3 The method utilizes heterogeneous remote sensing features to achieve deeper and more discriminative spatial feature extraction. Building upon this, a risk scoring module effectively enhances the representation of pixel-level risk probabilities. Furthermore, by introducing the K-means clustering algorithm and combining spatial coordinates with risk probability values ​​to construct a three-dimensional feature vector, the ability to simulate the spatial diffusion of risks is strengthened. Finally, by overlaying the data with crop plot boundary layers, a field-level crop risk level map is generated, and a structured early warning report file is exported, achieving accurate risk identification results and significantly improving information readability. This invention boasts advantages such as reasonable structural design, efficient module collaboration, strong adaptability, and intuitive result interpretation, significantly improving the accuracy and practicality of crop risk early warning and making it suitable for large-scale, continuous-timescale remote sensing monitoring tasks.

[0039] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A crop risk early warning method based on remote sensing data, characterized in that, Includes the following steps: Step 1: Obtain a multi-phase remote sensing image sequence of the crop, and register the multi-phase remote sensing image sequence to obtain a registered image sequence; Step 2: Based on the gray-level time series of each pixel in the registered image sequence, and perform frequency domain analysis, extract the main perturbation period and period stability of the pixel, and generate a gray-level period mapping map. Step 3: Calculate the directional gradient based on the registered image sequence, statistically analyze the perturbation entropy distribution, extract the asymmetric regions of the perturbation entropy distribution, and generate an abnormal texture region map; Step 4: Perform Gabor filtering on the images in the registered image sequence to construct a structural perturbation response map, and fuse it with the original registered image to generate a structural mutation channel map; Step 5: Using the grayscale periodic map, abnormal texture region map, and structural mutation channel map as input, and based on the multi-level morphological evolution convolutional fusion module with the InternImage model as the backbone, spatial domain processing is performed using morphological extension branch and morphological compression branch to generate fused feature map; Step 6: Generate a pixel-level risk probability map based on the fused feature map, and use the K-means algorithm for clustering to generate a risk spatial diffusion prediction map; Step 7: Overlay the risk spatial diffusion prediction map onto the crop plot boundary layer to generate a field-level crop risk level map and output a structured crop risk early warning report file.

2. The crop risk early warning method based on remote sensing data according to claim 1, characterized in that, Step one specifically involves: Acquire a multi-period remote sensing image sequence covering the same crop area. The multi-period remote sensing image sequence is acquired from the same satellite platform, with a spatial resolution of no more than 2 meters and an interval of 5 to 10 days between adjacent images. The earliest remote sensing image acquired in the time series was selected as the reference image, and the remaining remote sensing images were used as images to be registered. Extract a reference patch of size 21×21 pixels from the reference image, and extract candidate patches of the same size by sliding them in each image to be registered. Calculate the normalized cross-correlation coefficient for the reference patch and each candidate patch separately; The candidate patch position with the largest normalized cross-correlation coefficient is selected as the integer pixel-level matching position; The image to be registered is translated and transformed according to the integer pixel-level matching positions to generate a registered image; All registered remote sensing images are arranged in chronological order to construct a registered image sequence.

3. The crop risk early warning method based on remote sensing data according to claim 1, characterized in that, Step two specifically involves: Extract the grayscale value of each pixel in all periods from the registered image sequence to form the grayscale time series corresponding to the pixel; Perform a Fast Fourier Transform on each grayscale time series to obtain a spectrum vector, which includes the amplitude of each frequency. In the spectrum vector, the component with amplitude corresponding to zero frequency is removed, and the frequency component with the largest amplitude is selected from the remaining frequency components. The reciprocal of the corresponding frequency is taken as the main perturbation period of the pixel. The ratio of the amplitude of the frequency corresponding to the main disturbance period to the sum of the amplitudes of all non-zero frequencies is used as the periodic stability of the pixel. The main disturbance period and periodic stability are mapped to grayscale values ​​and transparency, respectively, and combined to generate a grayscale periodic mapping map. In the grayscale periodic inversion mapping map, the grayscale value of each pixel position represents the main disturbance period, and the transparency represents the periodic stability.

4. The crop risk early warning method based on remote sensing data according to claim 1, characterized in that, Step three specifically involves: Each image is selected from the registered image sequence, and the Sobel operator is used to calculate the grayscale change value of each pixel in the horizontal and vertical directions respectively. The gradient magnitude of a pixel is obtained by adding the squares of the horizontal and vertical grayscale changes of each pixel and taking the square root. Divide the vertical grayscale change value by the horizontal grayscale change value, and take the result of the arctangent function to obtain the gradient direction angle of the pixel; The gradient direction of all pixels is divided according to a preset direction range, which is from 0 degrees to 180 degrees and divided into multiple equally spaced direction segments. The gradient magnitude of the corresponding pixel is recorded in each direction segment. For each directional segment, the probability distributions containing pixel gradient magnitudes are statistically analyzed, and each probability value is multiplied by the logarithm of the probability value, then summed and negative to calculate the perturbation entropy of the corresponding directional segment. The perturbation entropy of each directional segment is combined in angular order to form a perturbation entropy distribution vector; The perturbation entropy distribution vector is paired and compared in symmetrical directions. If the difference in perturbation entropy between two directional segments exceeds a preset tolerance threshold, it is marked as an asymmetrical region of perturbation entropy distribution. Combine the pixel locations corresponding to all asymmetrical regions of perturbation entropy distribution to generate an abnormal texture region map.

5. The crop risk early warning method based on remote sensing data according to claim 1, characterized in that, Step four specifically involves: Gabor filtering is performed on each image in the registered image sequence to construct a multi-scale, multi-directional Gabor filter set; Convolution operations are performed on the Gabor filter set for each image to obtain the response values ​​of each pixel at different scales and directions, resulting in a total of 30 filter response layers; For each pixel, the response values ​​in the 30 filtered response layers are combined into a structural response vector of length 30. In the time dimension, the standard deviation of the structural response vector of the corresponding pixel in all periods is calculated to form a structural perturbation response map. The structural disturbance response map is fused with the original registration image of each period according to pixel position to generate a structural mutation layer; Combine all structural mutation layers in chronological order to form a structural mutation channel map.

6. The crop risk early warning method based on remote sensing data according to claim 5, characterized in that, The Gabor filter set specifically includes 5 filter groups, each corresponding to a spatial frequency of 0.05, 0.1, 0.2, 0.3 and 0.4 respectively. Each filter group includes 6 directional filters with directions of 0 degrees, 30 degrees, 60 degrees, 90 degrees, 120 degrees and 150 degrees respectively. For each image in the registered image sequence, a Gabor filter set is applied sequentially for convolution processing to obtain the response results of each pixel under different combinations of directions and frequencies. The response results are in the complex form of the Gabor convolution output. The square of the real part and the square of the imaginary part are added together and the square root is taken to obtain the response value of each pixel under the current direction and spatial frequency, which is used as the structural response intensity.

7. The crop risk early warning method based on remote sensing data according to claim 1, characterized in that, Step five specifically involves: The grayscale periodic map, abnormal texture region map, and structural mutation channel map are aligned according to pixel position and then stitched together along the channel dimension to form the initial input feature map. The initial input feature map is input into the InternImage model for feature extraction. The InternImage model includes four stages: Stage 1, Stage 2, Stage 3 and Stage 4. Stage 1 receives the initial input feature map, which has 64 input channels and is constructed by stacking two InternImage Blocks sequentially. Stage 2 consists of two InternImage Blocks with 64 input channels and 128 output channels. It is connected from Stage 1 to Stage 2 through a downsampling convolution with a stride of 2. Stage 3 consists of three InternImage Blocks with 128 input channels and 256 output channels. It is connected from Stage 2 to Stage 3 through downsampling convolution with a stride of 2. Stage 4 consists of three InternImage Blocks with 256 input channels and 512 output channels. It is connected from Stage 3 to Stage 4 through a downsampling convolution with a stride of 2. The multi-stage morphological evolution convolutional fusion module is inserted between Stage 2 and Stage 3; The multi-stage morphological evolution convolutional fusion module consists of three sequentially connected morphological evolution blocks. Each morphological evolution block includes a morphological extension branch, a morphological compression branch, and a regular convolution branch. The three branches of the first morphological evolution block receive the same Stage2 output feature map as the evolution input feature map. In the morphological extension branch, three sliding windows of 3×3, 5×5 and 7×7 are used to perform a pixel-by-pixel maximum value replacement operation on the evolutionary input feature map, generating extended feature maps of three scales, and then splicing them in the channel dimension. In the morphological extrusion branch, a sliding window of the same size as the morphological extension branch is used to perform a pixel-by-pixel minimum value replacement operation on the evolutionary input feature map, generating extrusion feature maps of three scales, which are then stitched together in the channel dimension. In the regular convolution branch, a 3×3 standard convolution is performed on the evolving input feature map to generate a convolutional feature map; The outputs of the three branches are concatenated along the channel dimension, input into a 1×1 convolution kernel for channel compression, and then passed through a normalization unit and an activation unit in sequence to generate the evolution output feature map of the current morphological evolution block. The evolutionary output feature map of the current morphological evolution block is added element by element to the evolutionary input feature map to form a residual connection, which is used as the input for the next morphological evolution block. The three morphological evolution blocks are connected in sequence. The third morphological evolution block outputs a morphological feature map, which is used as the input of Stage 3. Stage 4 outputs a fused feature map.

8. The crop risk early warning method based on remote sensing data according to claim 1, characterized in that, The generation of pixel-level risk probability maps based on fused feature maps specifically involves: The fused feature map is input into the risk scoring module, which consists of two convolutional layers and corresponding activation functions. The first convolutional layer uses a 3×3 kernel with a stride of 1 and a padding size of 1. The number of input channels is consistent with the fused feature map, and the number of output channels is set to 128. The second convolutional layer uses a 1×1 kernel with a stride of 1, padding size of 0, and the number of output channels is set to 1. The output of the first convolutional layer is followed by a ReLU activation function to truncate all negative values; The output of the second convolutional layer is followed by a Sigmoid activation function, which maps the convolutional output value of each pixel position to between 0 and 1, generating a risk probability value for the corresponding pixel position. The risk probability values ​​for all pixel locations are combined to generate a pixel-level risk probability map.

9. A crop risk early warning method based on remote sensing data according to claim 1, characterized in that, The K-means algorithm is used for clustering to generate a risk spatial diffusion prediction map, specifically as follows: Read the risk probability value of each pixel location from the pixel-level risk probability map, and record the row and column coordinates of the pixel in the image. Construct each pixel into a three-dimensional feature vector including row coordinates, column coordinates and risk probability value. The three-dimensional feature vectors corresponding to all pixels are used as a clustering sample set and input into the K-means clustering algorithm. The number of clusters K is set to 3, representing low risk, medium risk and high risk respectively. All cluster samples are sorted from smallest to largest according to their risk probability values. The mean of the cluster samples with risk probability values ​​in the top 10% is used as the first cluster center, the mean of the cluster samples with risk probability values ​​in the bottom 40% to 60% range is used as the second cluster center, and the mean of the cluster samples with risk probability values ​​in the bottom 10% is used as the third cluster center, thus completing the initial cluster center setting. In each round of clustering iteration, the Euclidean distance between each cluster sample and the three cluster centers is calculated sequentially, and the cluster sample is assigned to the category corresponding to the cluster center with the smallest Euclidean distance. After classifying all clustered samples, the row coordinates, column coordinates, and risk probability values ​​of all clustered samples within each class are averaged to update the cluster center position of the corresponding class. Repeat the process of dividing cluster samples and determining cluster center locations until the changes in cluster center locations in two consecutive rounds are less than a preset threshold, or the preset maximum number of iterations is reached, at which point the clustering process stops. Based on the final clustering results, the category corresponding to each pixel is mapped back to the original spatial location to generate a risk spatial diffusion prediction map with the same size as the pixel-level risk probability map.

10. A crop risk early warning method based on remote sensing data according to claim 1, characterized in that, Step seven specifically involves: The risk spatial diffusion prediction map is overlaid on the crop plot boundary layer according to pixel location. Based on the proportion of risk categories of pixels contained within each plot boundary, the risk category with the highest proportion is used as the risk level label of the plot, and a plot-level crop risk level map is generated. The information on field numbers, spatial locations, risk level labels, number of pixels and percentage of each risk level is compiled into structured entries and then summarized to generate a structured crop risk warning report file. The structured crop risk warning report file is in CSV or Excel spreadsheet format, with each field recording the corresponding field data on a separate line.

Images