A cultivated land change monitoring method and system fusing a deep neural network and remote sensing satellite data

Abstract

The application provides a kind of fusion deep neural network and cultivated land change monitoring method and system of remote sensing satellite data, it is related to cultivated land monitoring technical field, comprising: obtaining the three-dimensional feature grid of the region to be monitored, it contains the identification feature for distinguishing cultivated land area and uncultivated land area;Dimensionality reduction mapping is carried out to three-dimensional feature grid to obtain the differentiated gray scale image;Differentiated gray scale image is handled based on image recognition model to obtain the region distribution diagram with multiple areas to be classified;Based on classification model processing area to be classified, cultivated land area is determined and cultivated land change rate is calculated, the present application is extracted identification feature to facilitate subsequent targeted cultivated land area identification, first image recognition model is used to identify each area to be classified in the region to be monitored, then each area to be classified is identified and judged by classification model to determine the setting of cultivated land area, simplify the data processing difficulty of model, improve the recognition accuracy and recognition efficiency of cultivated land area.

Description

Technical Field

[0001] This invention relates to the field of farmland monitoring technology, specifically to a method and system for monitoring farmland changes that integrates deep neural networks and remote sensing satellite data. Background Technology

[0002] Farmland is the fundamental element for food production. Promoting comprehensive monitoring and supervision of farmland is a condition for ensuring the all-round and orderly development of this fundamental element. This requires the rational use of land resources and ensuring the timeliness, accuracy, and directionality of monitoring and early warning. Currently, farmland conditions are generally monitored and managed through real-time changes in satellite images. For example, the farmland change monitoring method and system based on remote sensing satellite data (classification number G06N) with authorization announcement number "CN117935081B" includes the following steps: based on remote sensing satellite data, using the K-means clustering algorithm, combined with processed graphic information, including radiometric correction and atmospheric correction, land types are divided into farmland and non-farmland, generating farmland regional division information. In this existing technology, by combining K-means clustering algorithm with remote sensing technology, cultivated land areas are identified; support vector machine and spectral analysis are used to assess soil fertility, pH value and organic matter; long short-term memory network and normalized difference moisture index enhancement model are combined to monitor soil moisture; random forest algorithm is used to assess early diseases; convolutional neural network and decision tree algorithm are used to analyze disaster impact; Kriging interpolation and principal component analysis are used to reveal soil variability; and multi-standard decision analysis and artificial neural network are combined to customize multi-dimensional strategies for cultivated land resource management.

[0003] However, existing technologies still have significant drawbacks. For example, existing technologies simply use the K-means clustering algorithm to process remote sensing satellite data to identify cultivated land areas. However, only a portion of the remote sensing satellite image data contains key information for distinguishing between cultivated and non-cultivated land, and this information is not comprehensive enough. When using the K-means clustering algorithm to process remote sensing satellite data, there are problems such as large data processing volume and high interference from irrelevant data, which is not conducive to effectively distinguishing between cultivated and non-cultivated land areas.

[0004] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for monitoring farmland changes that integrates deep neural networks and remote sensing satellite data, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A method for monitoring farmland changes by integrating deep neural networks and remote sensing satellite data includes the following steps: S1. Collect satellite remote sensing images of the area to be monitored and extract features to construct a three-dimensional feature grid of the area to be monitored. The three-dimensional feature grid contains identification features for distinguishing between cultivated land and non-cultivated land. S2, based on the differences in the identification features, the three-dimensional feature grid is reduced to obtain a two-dimensional difference matrix, and the two-dimensional difference matrix is ​​mapped into a differentiated grayscale image. The differentiated grayscale image is used to characterize the differences in the identification features within the area to be monitored. S3, based on convolutional neural networks, constructs and trains an image recognition model with a pre-segmentation module. The pre-segmentation module pre-segments the differential grayscale images based on a quadtree segmentation strategy. During training, the quadtree segmentation strategy is iteratively optimized based on the model performance. The differential grayscale images are input into the trained image recognition model to obtain a region distribution map, which includes at least one region to be classified, used to characterize the distribution of different types of regions within the region to be monitored. S4. A classification model is built and trained based on the random forest algorithm. Combining regional distribution maps, differential grayscale images, and three-dimensional feature grids, classification features of the region to be classified are extracted and input into the classification model to determine the cultivated land area within the region to be monitored. S5 is based on a long short-term memory network to build and train a prediction model to obtain the future cultivated land area, and statistically analyze the cultivated land area of ​​previous years, the current year and the future to generate the current cultivated land change rate and the future cultivated land change rate.

[0007] Furthermore, the logic for obtaining the 3D feature raster is as follows: S11. Acquire satellite remote sensing images of the area to be monitored and perform preprocessing operations on the satellite remote sensing images, including but not limited to radiometric correction, atmospheric correction and noise removal. S12, extract features from the preprocessed satellite remote sensing image to obtain the identification feature vector of each pixel. The identification features are used to distinguish between cultivated land and non-cultivated land areas, including but not limited to near-infrared band reflectance, green band reflectance, red band reflectance, short-wave infrared band reflectance, normalized vegetation index, differential vegetation index, enhanced vegetation index and moisture index. S13. Construct a three-dimensional blank raster with the same size as the satellite remote sensing image, and input the identification features of each pixel into the three-dimensional blank raster to form a three-dimensional feature raster.

[0008] Furthermore, the logic for dimensionality reduction of the three-dimensional feature raster is as follows: for the identification feature vector corresponding to any pixel in the satellite remote sensing image, the identification feature difference coefficient between it and its neighboring pixels is calculated, and the identification feature difference coefficients are summarized to form a two-dimensional difference matrix.

[0009] Furthermore, the method for mapping the two-dimensional difference matrix to a differentiated grayscale image is as follows: 1) For any element value in the two-dimensional difference matrix, normalize it to convert it into a gray level between 0 and 255; 2) Read the spatial resolution, number of rows and columns of the satellite remote sensing image, create a blank grayscale image with the same size as the satellite remote sensing image, and assign the grayscale value of each element in the two-dimensional difference matrix to the corresponding pixel of the blank grayscale image one by one to obtain the differentiated grayscale image.

[0010] Furthermore, the specific architecture of the image recognition model is as follows: The pre-segmentation module pre-segments the differentiated grayscale image based on the quadtree segmentation strategy to generate a pre-segmented image, which is a grayscale image with boundary contours. The input layer is used to receive the pre-segmented image output by the segmentation module; The image processing layer group is used to process the pre-segmented image. It includes three groups of image processing layers connected in sequence, specifically the first image processing layer, the second image processing layer and the third image processing layer. Each image processing layer consists of a convolutional layer, a batch normalization layer and a max pooling layer. Flattening layers are used to flatten the output of an image processing layer group into a one-dimensional array; Fully connected layers are used to further process the flattened features and learn higher-level feature representations; The output layer generates a region distribution map of the same size as the pre-segmented image. The region distribution map is a binarized image, which consists of boundary contours and regions to be classified.

[0011] Furthermore, the process of training the image recognition model is as follows: S31, acquire the differentiated grayscale and color images of the area to be monitored in previous years, and mark the boundary contours on the color image based on expert analysis to divide the color image into several different types of areas; S32, after mapping the labeled boundary contours onto the differential grayscale image, perform binarization to generate a region distribution map corresponding to the differential grayscale image. Summarize multiple sets of one-to-one corresponding differential grayscale images and region distribution maps to construct the first sample set. Divide the first sample set into a training set and a test set according to a 7:3 allocation method. S33, initialize the segmentation parameters of the pre-segmentation module and the hyperparameters of the image recognition model; The segmentation parameters of the pre-segmentation module include block threshold, minimum region threshold, grayscale segmentation threshold and uniformity standard, and hyperparameters include learning rate and batch size. S34. The differentiated grayscale images in the training set are input into the pre-segmentation module for pre-segmentation, and then input into the image recognition model for model training. At the end of each iteration of training, the differentiated grayscale images in the test set are processed by the pre-segmentation module and input into the image recognition model for testing. The segmentation parameters are optimized based on the test results and then trained again until the test results meet the standards.

[0012] Furthermore, the logic for optimizing the image recognition model is as follows: 1) Input the differentiated grayscale images of the same training batch one by one into the pre-segmentation module. The pre-segmentation module first analyzes the distribution of grayscale values ​​of pixels in each differentiated grayscale image to determine the segmentation threshold of each differentiated grayscale image. Based on the segmentation threshold, the differentiated grayscale image is segmented once to form multiple sub-regions. 2) Based on the quadtree segmentation strategy, the sub-regions in the differential grayscale image are simultaneously segmented to generate a pre-segmented image; 3) Input the pre-segmented images from the same training batch into the image recognition model for training. During the training process, use the binary cross-entropy function as the loss function and use the Adam or SGD optimization algorithm to adjust the model parameters to minimize the loss function. 4) After each round of training, the test set is input into the image recognition model for testing. The test results include the crossover ratio and the number deviation. The crossover ratio and the number deviation are combined to determine whether the image recognition model meets the requirements. If it meets the requirements, training is stopped; otherwise, the segmentation parameters are optimized by combining the crossover ratio and the number deviation and training continues. The formula for calculating the intersection-union ratio is as follows: In the formula, This represents the set of pixels belonging to the boundary contour in the region distribution map output by the image recognition model during the t-th iteration of training. This represents the set of pixels belonging to the boundary contour in the actual regional distribution map. This represents the set of pixels belonging to the region to be classified in the region distribution map output by the image recognition model during the t-th iteration of training. This represents the set of pixels in the actual regional distribution map that belong to the region to be classified. This represents the intersection-union ratio (IU) in the t-th iteration of training, where t is the index of the iteration round. The formula for calculating the number deviation is as follows: In the formula, This represents the number of regions to be classified in the region distribution map output by the image recognition model during the t-th iteration of training. This indicates the number of regions to be classified in the actual regional distribution map. This represents the deviation of the number of iterations in the t-th training round; The logic for determining whether an image recognition model meets the requirements is as follows: when the cross-union ratio (CUNR) is greater than the CUNR threshold and the number deviation is within the number deviation range, the image recognition model is considered to meet the requirements; otherwise, it is considered not to meet the requirements.

[0013] Furthermore, the logic for optimizing the segmentation parameters by combining the intersection-union ratio and the number of deviations is as follows: Based on the number deviation, the block threshold is optimized. The optimization logic is as follows: when the number deviation is less than 0, the block threshold is gradually reduced as the absolute value of the number deviation increases; when the number deviation is greater than 0, the block threshold is gradually increased as the number deviation increases. Based on the number deviation, the minimum region threshold is optimized. The optimization logic is as follows: when the number deviation is less than 0, the minimum region threshold is gradually reduced as the absolute value of the number deviation increases; when the number deviation is greater than 0, the minimum region threshold is gradually increased as the number deviation increases. Based on the cross-union ratio, the uniformity standard is optimized. The optimization logic is: as the cross-union ratio decreases, the uniformity standard is gradually reduced. Based on the intersection-over-union ratio (IoU), the grayscale segmentation threshold is optimized. The optimization logic is as follows: as the IoU decreases, the grayscale segmentation threshold is gradually reduced.

[0014] Furthermore, the classification features include, but are not limited to, the roundness, aspect ratio, texture features, and statistical features of the recognition features of the region to be classified. The texture features include, but are not limited to, energy, entropy, contrast, and correlation. The statistical features include, but are not limited to, maximum value, minimum value, mean, and variance.

[0015] A farmland change monitoring system integrating deep neural networks and remote sensing satellite data, used to execute the aforementioned farmland change monitoring method integrating deep neural networks and remote sensing satellite data, includes: The data extraction module is used to acquire satellite remote sensing images of the area to be monitored and extract features to construct a three-dimensional feature grid of the area to be monitored. The three-dimensional feature grid contains identification features for distinguishing between cultivated land and non-cultivated land areas. The grayscale image construction module reduces the dimensionality of the three-dimensional feature raster to obtain a two-dimensional difference matrix based on the differences in the identification features, and then maps the two-dimensional difference matrix into a differentiated grayscale image. The differentiated grayscale image is used to characterize the differences in the identification features within the area to be monitored. An image recognition model is used to process differential grayscale images to obtain a region distribution map. The region distribution map includes at least one region to be classified, which is used to characterize the distribution of different types of regions within the region to be monitored. The image recognition model also has a pre-segmentation module, which pre-segments the differential grayscale images based on a quadtree segmentation strategy. A classification model is used to process the classification features of the region to be classified to determine the cultivated land area within the region to be monitored. The classification features are obtained based on the regional distribution map, the differential grayscale image, and the three-dimensional feature grid. The predictive model is used to obtain the future cultivated land area and to statistically analyze the cultivated land area in previous years, the current year, and the future, generating the current cultivated land change rate and the future cultivated land change rate.

[0016] Compared with the prior art, the beneficial effects of the present invention are: The method and system for monitoring farmland changes that integrates deep neural networks and remote sensing satellite data of the present invention first filters and combines satellite remote sensing data to extract identification features for distinguishing farmland areas from non-farmland areas to construct a three-dimensional feature grid, which facilitates subsequent targeted identification of farmland areas. Then, the three-dimensional feature grid is dimensionality-reduced and mapped to obtain a differentiated grayscale image that characterizes the differences in identification features within the area to be monitored, which facilitates subsequent image processing. Then, a "two-step" approach is adopted, namely, first, the image recognition model identifies each unclassified area in the area to be monitored, and then the classification model identifies and judges each unclassified area to determine the farmland area setting, which simplifies the data processing difficulty of the model and improves the identification accuracy and efficiency of farmland areas. Furthermore, the present invention embeds a pre-segmentation module at the front end of the image recognition model, which reduces the workload of the image recognition model in analyzing the output region distribution map. Based on the feedback from the image recognition model, the pre-segmentation module is optimized to provide the image recognition model with high-quality pre-segmented images, thereby improving the output accuracy of the model. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a modular unit diagram of the present invention; Figure 3 The intersection-union ratio (IU) of an image recognition model with a pre-segmentation module; Figure 4 The deviation in the number of image recognition models with pre-segmentation modules; Figure 5 The intersection-union ratio (IU) of an image recognition model without a pre-segmentation module; Figure 6 The deviation in the number of image recognition models without pre-segmentation modules. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0019] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0020] Example 1: Please see Figure 1 , Figures 3-6 This embodiment provides a method for monitoring farmland changes by fusing deep neural networks and remote sensing satellite data, including the following steps: S1. Acquire satellite remote sensing images of the area to be monitored and extract features to construct a three-dimensional feature raster of the area. The three-dimensional feature raster contains identification features for distinguishing between cultivated land and non-cultivated land areas, including the following steps: S11. Acquire satellite remote sensing images of the area to be monitored and perform preprocessing operations on the satellite remote sensing images, including but not limited to radiometric correction, atmospheric correction and noise removal. As one implementation method, medium-resolution satellites of the Landsat series (such as Landsat 8) are used to acquire satellite remote sensing images of the area to be monitored. The Landsat 8 satellite can acquire spectral information across the entire band, which facilitates subsequent comprehensive analysis and identification of cultivated and non-cultivated areas. Furthermore, the Landsat 8 satellite has a spatial resolution of 15 meters by 15 meters, which can effectively identify the boundaries of cultivated land and is suitable for long-term monitoring and identification in the agricultural field. Among them, radiometric correction is used to convert the original spectral information into reflectance to eliminate errors caused by atmospheric influences, sensor characteristics, and lighting conditions; geometric correction is used to map each pixel in the image to the actual geographical location to correct geometric distortions caused by sensor tilt, terrain changes, etc.; atmospheric correction is used to remove atmospheric interference with the signal, including scattering and absorption effects, to obtain more realistic surface reflectance information. It should be noted that collecting satellite remote sensing images of a certain area and performing radiometric correction, atmospheric correction and noise removal are common knowledge in the art and will not be elaborated here. It should be noted that for any pixel in the preprocessed satellite remote sensing image, it includes geographic coordinates, timestamp, and reflectance data in the visible and near-infrared bands. Some of this data is key information for distinguishing between arable and non-arable land, but it is not comprehensive and lacks specificity. For example, a full-band remote sensing image includes reflectance information in the ultraviolet, blue, green, red, near-infrared, short-wave infrared, mid-infrared, and thermal infrared bands. However, only the reflectance information in the near-infrared, green, red, and short-wave infrared bands can be effectively used to distinguish between arable and non-arable land areas. If satellite remote sensing images are directly used to distinguish between arable and non-arable land areas, there are problems such as large data processing volume and high interference from irrelevant data, which is not conducive to effectively distinguishing between arable and non-arable land areas. S12, extract features from the preprocessed satellite remote sensing image to obtain the identification feature vector of each pixel. The identification features are used to distinguish between cultivated land and non-cultivated land areas, including but not limited to near-infrared band reflectance, green band reflectance, red band reflectance, short-wave infrared band reflectance, normalized vegetation index, differential vegetation index, enhanced vegetation index and moisture index. The identification features used to distinguish between cultivated land areas and non-cultivated land areas are based on the following criteria: Near-infrared reflectance: Plants typically have high reflectance in the near-infrared band. Healthy plants reflect a large amount of near-infrared radiation, while non-arable land such as bare soil or water bodies have low reflectance in the near-infrared band. Therefore, by comparing near-infrared reflectance, arable land (crop growing area) and non-arable land can be effectively identified. Green light band reflectance: The reflectance of the green light band is related to the health status and growth stage of vegetation. Healthy plants have higher reflectance in the green light band, while non-arable land such as bare soil or urban areas often have lower reflectance. By analyzing the green light band reflectance, land use type can be determined. Red light reflectance: The reflectance of the red light band is closely related to the chlorophyll content. Plants have low reflectance in this band. Cultivated land plants absorb a lot of red light through photosynthesis, resulting in low reflectance. Non-cultivated land, such as water bodies or bare land, often shows higher red light reflectance. Therefore, red light reflectance is an important indicator for distinguishing between cultivated land and non-cultivated land. Shortwave infrared reflectance: Shortwave infrared reflectance is sensitive to soil moisture and vegetation moisture. Cultivated land has a high moisture content and usually has a high shortwave infrared reflectance, while non-cultivated land such as dry soil has a low reflectance. Using shortwave infrared reflectance can help identify moist cultivated land and dry non-cultivated land. Normalized Difference Vegetation Index (NDVI): The Normalized Difference Vegetation Index can effectively characterize the health status of vegetation by comparing the reflectance of the near-infrared band and the red band. The NDVI value is between -1 and 1. The higher the value (usually >0.2), the more lush the vegetation. Cultivated land usually has a higher NDVI value, while non-cultivated land (such as urban areas, bare soil, and water bodies) usually has a lower NDVI value. Therefore, NDVI is an important indicator for distinguishing between cultivated land and non-cultivated land. Differential Vegetation Index (DVI): The Differential Vegetation Index characterizes vegetation growth by measuring the difference in reflectance between the shortwave infrared band and the red light band. Higher DVI values ​​usually correspond to denser vegetation. Therefore, DVI can be used to quickly distinguish between cultivated land (high DVI value) and non-cultivated land (low DVI value). Enhanced Vegetation Index (EVI): The Enhanced Vegetation Index can more accurately characterize vegetation health in areas with high vegetation cover and reduce the impact of soil and atmosphere. EVI values ​​are usually higher than NDVI. Healthy cultivated land has a higher EVI value, while non-cultivated land has a lower EVI value, which makes EVI an effective way to identify the differences between cultivated and non-cultivated land. Moisture Index (MVI): The moisture index detects the moisture status of soil and vegetation by comparing the reflectance of short-wave infrared and near-infrared bands. Cultivated land usually has a higher moisture content, so it has a higher MVI value, while non-cultivated land may show a lower MVI value due to drought. This can effectively distinguish between wet cultivated land and dry non-cultivated land. Taking an example where the identification features only include near-infrared reflectance, green reflectance, red reflectance, short-wave infrared reflectance, normalized vegetation index, differential vegetation index, enhanced vegetation index, and moisture index, the representation of the identification feature vector is as follows: .

[0021] In the formula, This represents the feature vector corresponding to the pixel in the i-th row and j-th column of a satellite remote sensing image. Let represent the near-infrared reflectance, green reflectance, red reflectance, short-wave infrared reflectance, normalized difference vegetation index (NDVI), differential vegetation index (DVI), enhanced vegetation index (EDI), and moisture index of the pixel in the i-th row and j-th column of the satellite remote sensing image, respectively. Let i and j be the row and column indices of the pixel in the satellite remote sensing image, respectively. m and n represent the height and width of the satellite remote sensing image, respectively. The indexing method starting from 0 is convenient for consistency with the array indexing rules in multiple programming languages ​​(such as Python, C++, etc.) and computer science. S13. Construct a three-dimensional blank raster with the same size as the satellite remote sensing image, and input the identification features of each pixel into the three-dimensional blank raster to form a three-dimensional feature raster, as follows: S131, Read the spatial resolution, number of rows and columns of the satellite remote sensing image, and create a three-dimensional blank raster with the same size as the satellite remote sensing image; It should be noted that a three-dimensional blank raster with internal elements initialized to 0 can be created based on a NumPy array. The same size means that the number of rows and columns of the created three-dimensional blank raster are the same as those of the satellite remote sensing image, so that each grid of the three-dimensional blank raster corresponds one-to-one with the pixel in the satellite remote sensing image. S132, write the identification features corresponding to each pixel in the satellite remote sensing image into the corresponding grid in the three-dimensional blank raster, and complete the construction of the three-dimensional feature raster; The constructed 3D feature grid includes three dimensions: height, width, and recognition feature vector. The 3D feature grid is represented as follows: In the formula, Represents a three-dimensional feature grid; It should be noted that non-arable land areas include, but are not limited to, wasteland areas, forest areas, grassland areas, wetland areas, water bodies, and urbanized areas.

[0022] S2, based on the differences in the identification features, the three-dimensional feature grid is reduced to obtain a two-dimensional difference matrix, and the two-dimensional difference matrix is ​​mapped into a differentiated grayscale image. The differentiated grayscale image is used to characterize the differences in the identification features within the area to be monitored. The method for dimensionality reduction of three-dimensional feature grids is as follows: For the recognition feature vector of any pixel, the recognition feature difference coefficient between it and its neighboring pixels is calculated using the following formula: In the formula, This represents the k-th identification feature corresponding to the pixel in the i-th row and j-th column of a satellite remote sensing image, where k is the index of the identification feature. K represents the total number of identification features. If the identification features only include near-infrared reflectance, green reflectance, red reflectance, short-wave infrared reflectance, normalized vegetation index, differential vegetation index, enhanced vegetation index, and moisture index, then K=8. In the formula, This represents the set of neighboring pixels of the pixel located in the i-th row and j-th column in a satellite remote sensing image. This represents the number of pixels in the set of adjacent pixels. For example, for a pixel in row 0 and column 0, its corresponding set of adjacent pixels includes the pixel in row 0 and column 1, and the pixel in row 1 and column 0. For example, for a pixel in row 0 and column 1, its corresponding set of adjacent pixels includes the pixel in row 0 and column 0, the pixel in row 0 and column 2, and the pixel in row 1 and column 1. For example, for a pixel in the first row and first column, its corresponding set of adjacent pixels includes the pixel in the zeroth row and first column, the pixel in the second row and first column, the pixel in the first row and zeroth column, and the pixel in the first row and second column. In this case, the corresponding... ; In the formula, Represents the set of adjacent pixels In the diagram, the k-th recognition feature corresponds to the q-th pixel, where q is the set of neighboring pixels. The index of the middle pixel, and ; In the formula, This represents the difference coefficient of recognition features between the pixel in the i-th row and j-th column and its neighboring pixels in a satellite remote sensing image. The larger the value, the greater the difference in recognition features between the pixel in the i-th row and j-th column and its neighboring pixels, and thus the higher the probability that the pixel in the i-th row and j-th column is the boundary contour between different regions. It should be noted that, Used to calculate the pixel in the i-th row and j-th column of a satellite remote sensing image, and the set of its neighboring pixels. The difference in the identification features of the q-th pixel in the satellite remote sensing image is considered. A larger value indicates a higher degree of difference between the pixel in the i-th row and j-th column and its set of neighboring pixels. The greater the probability that the q-th pixel belongs to either cultivated land or non-cultivated land, the more likely it is to be determined by... The average difference in recognition features between the pixels in the i-th row and j-th column and their neighboring pixels is calculated. The larger the value, the less likely the pixels in the i-th row and j-th column and their neighboring pixels belong to the same region. This also indicates that the pixels in the i-th row and j-th column are more likely to be the boundary contours between different regions. In the formula, The preset weight for the k-th identification feature is used to characterize the importance of the k-th identification feature in farmland area identification. The higher the importance, the greater the weight. The larger the value, and the more it satisfies ; As one implementation method, all recognition features are defined to have the same importance in the evaluation of the degree of difference between adjacent pixels, that is, let This setup simplifies the calculation process and improves the efficiency of identifying arable land areas. As another implementation method, the preset weights of each identification feature are determined based on the analytic hierarchy process (AHP). The AHP can effectively quantify the relative importance of different identification features in farmland area identification. The preset weights determined by the AHP are more accurate than uniform settings and have a certain degree of universality, which can effectively improve the accuracy of subsequent farmland area identification. The two-dimensional difference matrix obtained after dimensionality reduction is represented as follows: In the formula, Represents a two-dimensional difference matrix; The method for mapping a two-dimensional difference matrix to a differentiated grayscale image is as follows: 1) For any element value in the two-dimensional difference matrix, normalization is performed to convert it into a grayscale value between 0 and 255. The specific conversion formula is as follows: In the formula, The maximum value in the two-dimensional difference matrix. It is the minimum value in the two-dimensional difference matrix. for The grayscale value obtained after conversion; 2) Read the spatial resolution, number of rows and columns of the satellite remote sensing image, create a blank grayscale image with the same size as the satellite remote sensing image, and assign the grayscale value of each element in the two-dimensional difference matrix to the corresponding pixel of the blank grayscale image one by one to obtain the differentiated grayscale image. It should be noted that blank grayscale images can be constructed using the PIL library in Python. The grayscale value of any pixel in a blank grayscale image is initialized to 0. When processing large images, a transformed grayscale value matrix can be established based on a two-dimensional difference matrix, and then the differentiated grayscale image can be created directly in NumPy. Compared with the method of assigning values ​​one by one in a loop, this speeds up the generation rate of differentiated grayscale images. It should be noted that the method of converting a three-dimensional feature raster into a two-dimensional difference matrix through dimensionality reduction and then mapping it into a differentiated grayscale image not only preserves key difference information but also transforms the data into an image. This facilitates subsequent use of image processing techniques to partition the area to be monitored, significantly improving the efficiency of identifying and partitioning the area to be monitored.

[0023] S3, based on convolutional neural networks, constructs and trains an image recognition model with a pre-segmentation module. The pre-segmentation module pre-segments the differential grayscale images based on a quadtree segmentation strategy. During training, the quadtree segmentation strategy is iteratively optimized based on the model performance. The differential grayscale images are input into the trained image recognition model to obtain a region distribution map, which includes at least one region to be classified, used to characterize the distribution of different types of regions within the region to be monitored. It should be noted that the type of area to be classified is one of the following: cultivated land, wasteland, forest, grassland, wetland, water body, and urbanized buildings; The image recognition model can adopt conventional model architectures such as U-Net, SegNet, or FCN, as detailed below: The pre-segmentation module pre-segments the differentiated grayscale image based on the quadtree segmentation strategy to generate a pre-segmented image. The pre-segmented image is a grayscale image with boundary contours. The grayscale value of the pixels of the boundary contours is 255. In the pre-segmented image, the region surrounded by the boundary contours or the region jointly surrounded by the boundary contours and the boundary of the pre-segmented image is a sub-region. In the identification of the pre-segmentation module, the sub-region is one of the following: cultivated land region, wasteland region, forest region, grassland region, wetland region, water body region, and urbanized building region. The input layer is used to receive the pre-segmented image output by the segmentation module; The image processing layer group comprises three sequentially connected image processing layers: a first image processing layer, a second image processing layer, and a third image processing layer. Each image processing layer consists of a convolutional layer, a batch normalization layer, and a max pooling layer, as detailed below: The first image processing layer includes a convolutional layer 1, a batch normalization layer 1, and a max pooling layer 1, wherein: Convolutional layer 1 consists of 32 3x3 convolutional kernels and uses the ReLU activation function to extract local features in the pre-segmented image, including but not limited to edges and textures. Batch normalization layer 1 is used to normalize the output of convolutional layer 1, reduce internal covariate bias, accelerate the training process, and improve model stability. Max pooling layer 1 uses a 2x2 pooling window. It reduces the dimension of the feature map by taking the maximum value within the window, thereby reducing the amount of computation and the number of parameters, while retaining the main features. The second image processing layer includes a convolutional layer 2, a batch normalization layer 2, and a max pooling layer 2, wherein: Convolutional layer 2 consists of 64 3x3 convolutional kernels, also using the ReLU activation function. Its input is the output of max pooling layer 1, which is used to further extract higher-level features to capture more complex patterns. Batch normalization layer 2 is used to normalize the output of convolutional layer 2, thereby improving the stability of training. Max pooling layer 2 uses a 2x2 pooling window to further downsample the feature map, reduce dimensionality, and reduce computational cost; The third image processing layer includes a convolutional layer 3, a batch normalization layer 3, and a max pooling layer 3, wherein: Convolutional layer 3 consists of 128 3x3 convolutional kernels, also using the ReLU activation function. Its input is the output of max pooling layer 2, used to extract deeper features to enhance the model's learning ability. Batch normalization layer 3 is used to normalize the output of convolutional layer 3 to improve the stability and convergence speed of the model; Max pooling layer 3 uses a 2x2 pooling window to further downsample the feature map, providing a more compact feature representation for subsequent fully connected layers; The flattening layer is used to flatten the output of the max pooling layer 3 into a one-dimensional array to prepare data for the fully connected layer and ensure that the input shape meets the requirements of the fully connected layer. The fully connected layer, consisting of 512 neurons, is used to further process the flattened features and learn higher-level feature representations. The output layer uses the Sigmoid activation function to generate a region distribution map of the same size as the pre-segmented image. The region distribution map is a binary image, which consists of white regions and closed black regions. The white regions are the boundary contours, and the closed black regions are the regions to be classified. In the output layer, the regions to be classified are one of the following: cultivated land, wasteland, forest, grassland, wetland, water body, and urbanized building areas. It should be noted that if a black region is surrounded by a white region, or is jointly surrounded by a white region and the boundary of the binarized image, then this region is defined as a closed black region. The process of training the image recognition model is as follows: S31, acquire the differentiated grayscale and color images of the area to be monitored in previous years, and mark the boundary contours on the color image based on expert analysis to divide the color image into several different types of areas; It should be noted that the color image and the satellite remote sensing image have the same spatial resolution and are spatially aligned, so that the pixels of the color image and the differentiated grayscale image can be aligned one by one. The boundary contours on the color image can be marked using image annotation tools such as OpenCV and Adobe Photoshop. The color image can visually reflect the distribution of different types of areas in the area to be monitored. Therefore, the color image is selected here for expert analysis in order to find accurate boundary contours, which will facilitate the subsequent model training process. S32, after mapping the labeled boundary contours onto the differential grayscale image, perform binarization to generate a region distribution map corresponding to the differential grayscale image. Summarize multiple sets of one-to-one corresponding differential grayscale images and region distribution maps to construct the first sample set. Divide the first sample set into a training set and a test set according to a 7:3 allocation method. The specific logic for mapping the marked boundary contours onto the differentiated grayscale image is as follows: find the pixel corresponding to the boundary contour from the differentiated grayscale image, replace the grayscale value of the pixel with 255, and replace the pixels with grayscale values ​​other than 255 with 0, thereby forming a regional distribution map. S33, initialize the segmentation parameters of the pre-segmentation module and the hyperparameters of the image recognition model; The segmentation parameters of the pre-segmentation module include block threshold, minimum region threshold, grayscale segmentation threshold, and uniformity standard. The hyperparameters include learning rate and batch size. The initial values ​​of learning rate, batch size, minimum region threshold, and uniformity standard can be determined based on past experience or expert analysis. For example, the initial value of learning rate can be set between 0.001 and 0.01, the initial value of batch size can be set to 30, 50, or 80, etc., and the minimum region threshold can be set between 5% and 10% of the total number of pixels in the differential grayscale image. The initial value of the block threshold is calculated using the following formula: In the formula, This is the initial value for the block threshold. The average gray value of pixels in a differential grayscale image. The standard deviation of the gray values ​​of pixels in a differential grayscale image. The threshold adjustment factor has a value between 1 and 3. To avoid over-splitting of the differentiated grayscale image, the threshold adjustment factor can be set to 3 to find pixels that can be clearly used as the dividing contour. S34: The differentiated grayscale images in the training set are input into the pre-segmentation module for pre-segmentation, and then input into the image recognition model for model training. At the end of each training iteration, the differentiated grayscale images in the test set are processed by the pre-segmentation module and input into the image recognition model for testing. Based on the test results, the segmentation parameters are optimized and the model is trained again until the test results meet the requirements. The specific optimization logic is as follows: 1) Differentiated grayscale images from the same training batch are input one by one into the pre-segmentation module. The pre-segmentation module first analyzes the grayscale value distribution of pixels in each differentiated grayscale image to determine the block threshold of each differentiated grayscale image. Based on the block threshold, the differentiated grayscale image is segmented once to form multiple sub-regions. The segmentation logic is as follows: Pixels with gray values ​​higher than the block threshold are selected from the differential grayscale image as boundary pixels. The gray values ​​of the boundary pixels are replaced with 255. The area surrounded by pixels with a gray value of 255, or the area surrounded by pixels with a gray value of 255 and the boundary of the differential grayscale image, is taken as the sub-region. 2) Based on the quadtree segmentation strategy, the sub-regions in the differential grayscale image are simultaneously segmented to generate a pre-segmented image. The specific segmentation strategy is as follows: For each sub-region, the average gray value and standard deviation of the pixels in the sub-region are calculated. A gray-level segmentation threshold and uniformity standard are preset. If the average gray value of the sub-region is less than the gray-level segmentation threshold and the standard deviation of the gray value is less than the uniformity standard, it means that the region corresponding to the sub-region is of the same type and the sub-region is not segmented. Otherwise, it means that the region corresponding to the sub-region is a combination of multiple different types of regions and it is divided into four sub-regions. The same segmentation strategy is used to iteratively segment the four sub-regions until the average gray value is less than the gray-level segmentation threshold and the standard deviation of the gray value is less than the uniformity standard, or the sub-region to be segmented is less than the minimum region threshold. The segmentation boundary formed by the segmentation based on the quadtree segmentation strategy is extracted, and the gray value of the pixels at the segmentation boundary is replaced with 255 to form a pre-segmented image. It should be noted that if the average gray value is less than the gray-level segmentation threshold and the standard deviation of the gray value is less than the uniformity standard, it means that the gray values ​​of the pixels in this sub-region are generally small and the distribution is relatively uniform. This means that the pixels in this sub-region are more likely to belong to the same type of region, so this sub-region is not segmented. Conversely, if the average gray value is less than the gray-level segmentation threshold and the standard deviation of the gray value is less than the uniformity standard, it means that the pixels in this sub-region are more likely to belong to multiple different types of regions, so this sub-region is segmented to further subdivide the region. The gray-level segmentation threshold and uniformity standard can be determined by expert analysis or by combining the regional distribution map and the differentiated gray-level image. For example, based on the regional distribution map of the first sample set, the average gray value of the corresponding differentiated gray-level image in each region to be classified is calculated, and then the mean is used as the gray-level threshold. Similarly, the uniformity standard is calculated. It should be noted that by dividing the differential grayscale image through the above pre-segmentation module, a pre-segmented image with contour features is obtained, which reduces the workload of the output region distribution map of the model analysis in the following text, and is conducive to the iterative optimization training process of the model, reducing the difficulty of model training. In addition, the differential image is first divided into multiple sub-regions, and then quadtree segmentation is performed on each sub-region simultaneously, which speeds up the progress of quadtree segmentation, reduces the segmentation difficulty, and improves the accuracy of quadtree segmentation, so as to provide the model with high-quality pre-segmented images, thereby improving the output accuracy of the model. 3) Input the pre-segmented images from the same training batch into the image recognition model for training. During the training process, use the binary cross-entropy function as the loss function to evaluate the difference between the model output value and the actual value, and use the Adam or SGD optimization algorithm to adjust the model parameters to minimize the loss function. 4) After each round of training, the test set is input into the image recognition model for testing. The test results include the crossover ratio and the number deviation. The crossover ratio and the number deviation are combined to determine whether the image recognition model meets the requirements. If the requirements are met or the predetermined number of iterations is reached, training is stopped. Otherwise, the segmentation parameters are optimized by combining the crossover ratio and the number deviation and training continues. The formula for calculating the intersection-union ratio is as follows: In the formula, This represents the set of pixels belonging to the boundary contour in the region distribution map output by the image recognition model during the t-th iteration of training. This represents the set of pixels belonging to the boundary contour in the actual regional distribution map. This represents the set of pixels belonging to the region to be classified in the region distribution map output by the image recognition model during the t-th iteration of training. This represents the set of pixels in the actual regional distribution map that belong to the region to be classified. This represents the intersection-union ratio (IU) in the t-th iteration of training, where t is the index of the iteration round. It should be noted that, This value is used to characterize the accuracy of the image recognition model in recognizing boundary contours during the t-th training iteration. The larger the value, the higher the accuracy of the image recognition model in recognizing boundary contours during the t-th training iteration. The Cross-Union Ratio (CIRR) is used to characterize the accuracy of the image recognition model in identifying the region to be classified during the t-th training iteration. A higher CIRR indicates higher accuracy. Therefore, the CIRR is represented by averaging the two values. The larger the value, the higher the accuracy of the image recognition model in recognizing the boundary contour and the region to be classified during the t-th iteration of training. The formula for calculating the number deviation is as follows: In the formula, This represents the number of regions to be classified in the region distribution map output by the image recognition model during the t-th iteration of training. This indicates the number of regions to be classified in the actual regional distribution map. This represents the deviation of the number of iterations in the t-th training round; It should be noted that, This value is used to characterize the deviation between the number of regions to be classified in the region distribution map output by the model and the true value. The greater the deviation from 0, the greater the deviation between the number of regions to be classified in the region distribution map output by the model and the true value. This indicates that the image recognition model performs worse in the t-th iteration of training. Specifically: When the absolute value is less than 0, the larger it is, the fewer the number of regions to be classified in the model's output region distribution map compared to the true value. This means the model's output region distribution map has a lower degree of region segmentation, and a greater degree of classifying multiple different types of regions into the same region. Conversely, a smaller absolute value indicates a lower degree of region segmentation. When the value is greater than 0, the larger the value, the more regions to be classified are in the regional distribution map output by the model compared to the true value. In other words, the higher the degree of region splitting in the regional distribution map output by the model, the greater the degree to which the same type of region is split into multiple different types of regions, and the more serious the over-splitting. The logic for determining whether an image recognition model meets the requirements is as follows: if the intersection-union ratio (IU) is greater than the IU threshold and the number deviation is within the number deviation range, the image recognition model is considered to meet the requirements; otherwise, it is considered not to meet the requirements. The specific values ​​of the IU threshold and the number deviation range can be set by the staff according to the actual situation. For example, the IU threshold can be set between 85% and 95%, and the number deviation range can be set to... ; The logic for optimizing the segmentation parameters by combining the intersection-union ratio and the number deviation is as follows: Based on the number deviation, the block threshold is optimized. The optimization logic is as follows: when the number deviation is 0, the block threshold remains unchanged; when the number deviation is less than 0, the block threshold is gradually decreased as the absolute value of the number deviation increases; when the number deviation is greater than 0, the block threshold is gradually increased as the number deviation increases. The specific optimization formula is as follows: In the formula, The block threshold used by the pre-segmentation module in the (t+1)th iteration of training. The bias in the number of pre-segmentation modules used in the t-th iteration of training. This is the block threshold optimization factor used in the t-th iteration of training. Its specific value is between 0.01 and 0.1, determined by the staff based on the changes in the number of biases. In iterative training, if the region block map output by the model in the t-th iteration does not show a significant improvement in the accuracy of the number of regions compared to the (t-1)-th iteration (e.g., ...), ... ), then compared to In terms of raising To expedite the finding of the optimal block threshold, This is the block threshold optimization factor used in the t-th iteration of training. This represents the initial value of the block threshold; It should be noted that a deviation of less than 0 indicates that the model's output region distribution map is too poorly segmented, while a deviation of more than 0 indicates that the model's output region distribution map is too poorly segmented. Therefore, in calculation... At that time, with Based on, adopt right The system is optimized by lowering the segmentation threshold when the region distribution map is too poorly segmented, thereby using more pixels as boundary pixels and increasing the number of sub-regions in the pre-segmented image. Conversely, it raises the segmentation threshold when the region distribution map is too poorly segmented, thereby using fewer pixels as boundary pixels and reducing the number of sub-regions in the pre-segmented image, thus improving the classification accuracy of the pre-segmentation module. Based on the number deviation, the minimum region threshold is optimized. The optimization logic is as follows: when the number deviation is 0, the minimum region threshold remains unchanged; when the number deviation is less than 0, the minimum region threshold is gradually decreased as the absolute value of the number deviation increases; when the number deviation is greater than 0, the minimum region threshold is gradually increased as the number deviation increases. The specific optimization formula is as follows: In the formula, This represents the minimum region threshold used by the pre-segmentation module in the t-th iteration of training. This refers to the minimum region threshold used by the pre-segmentation module in the (t+1)th iteration of training. The initial value for the minimum region threshold. This is the minimum region threshold optimization factor used in the t-th iteration of training. Its specific value is between 0.1 and 0.3, determined by the staff based on the changes in the number of regions. If, in the t-th iteration, the model's output region segmentation map does not show a significant improvement in the accuracy of the number of regions compared to the (t-1)-th iteration, then compared to... In terms of raising To expedite the finding of the optimal minimum region threshold, This is the minimum region threshold optimization factor used in the (t-1)th iteration of training; It should be noted that a deviation of less than 0 indicates that the model's output region distribution map is too poorly segmented, while a deviation of more than 0 indicates that the model's output region distribution map is too poorly segmented. Therefore, in calculation... At that time, through right The system is optimized by lowering the minimum region threshold when the region distribution map is too poorly segmented, so that smaller sub-regions can be segmented when using the quadtree segmentation strategy, thereby increasing the number of sub-regions in the pre-segmented image. Conversely, it raises the minimum region threshold when the region distribution map is too poorly segmented, so that smaller sub-regions cannot be segmented when using the quadtree segmentation strategy, thereby reducing the number of sub-regions in the pre-segmented image and improving the classification accuracy of the pre-segmentation module. Based on the cross-union ratio (CUNR), the uniformity standard is optimized. The optimization logic is as follows: as the CUNR decreases, the uniformity standard is gradually reduced. The specific optimization formula is as follows: In the formula, The uniformity standard used by the pre-segmentation module in the (t+1)th iteration of training. The uniformity criterion used by the pre-segmentation module in the t-th iteration of training. The initial value for uniformity standard, The uniformity standard optimization factor used in the t-th iteration of training has a specific value between 0.05 and 0.2. The specific value is determined by the staff based on the changes in the intersection-union ratio (IU). In iterative training, if the region block map output by the model in the t-th iteration does not show a significant improvement in IU compared to the (t-1)-th iteration (e.g., ...), ... ), then compared to In terms of raising To expedite the search for the optimal uniformity standard, The uniformity standard optimization factor used in the (t-1)th iteration of training. The preset crossover-union ratio threshold; It should be noted that a lower intersection-over-union ratio (IoU) indicates lower accuracy in the regional distribution map output by the model. Therefore, in calculations... At that time, first through This is used to quantify the deviation between the Cross-Union Ratio (CUI) and the Cross-Union Ratio (CUI) in the t-th iteration of training. A larger value indicates poorer recognition accuracy and a greater degree of misidentification of boundary contours and unclassified regions in the t-th iteration of training. Therefore, by... right The correction is made to reduce the uniformity standard when the recognition accuracy of the regional distribution map decreases, so as to improve the segmentation detail when using the quadtree segmentation strategy to segment sub-regions, thereby ensuring the accuracy of sub-region division and improving the classification accuracy of the pre-segmentation module. Based on the intersection-over-union ratio (IoU), the grayscale segmentation threshold is optimized. The optimization logic is as follows: as the IoU decreases, the grayscale segmentation threshold is gradually reduced. The specific optimization formula is as follows: In the formula, This refers to the grayscale segmentation threshold used by the pre-segmentation module in the (t+1)th iteration of training. Let be the grayscale segmentation threshold used by the pre-segmentation module in the t-th round of training iterations. This is the initial value for the grayscale segmentation threshold. The grayscale segmentation threshold optimization factor used in the t-th iteration of training has a specific value between 0.1 and 0.5. The specific value is determined by the staff based on the changes in the intersection-union ratio (IU). If, in the t-th iteration, the model's output region block map does not show a significant improvement in IU compared to the (t-1)-th iteration, then compared to... In terms of raising To expedite the discovery of the optimal grayscale segmentation threshold, This is the grayscale segmentation threshold optimization factor used in the (t-1)th iteration of training; It should be noted that a lower intersection-over-union ratio (IoU) indicates lower accuracy in the regional distribution map output by the model. Therefore, in calculations... At that time, first through This is used to quantify the deviation between the Cross-Union Ratio (CUI) and the Cross-Union Ratio (CUI) in the t-th iteration of training. A larger value indicates poorer recognition accuracy and a greater degree of misidentification of boundary contours and unclassified regions in the t-th iteration of training. Therefore, by... right The correction is made to reduce the grayscale segmentation threshold when the recognition accuracy of the regional distribution map decreases, so as to improve the segmentation detail when using the quadtree segmentation strategy to segment sub-regions, thereby ensuring the accuracy of sub-region division and improving the classification accuracy of the pre-segmentation module. As one implementation, 10 sets of differentiated grayscale images were input into an image recognition model without a pre-segmentation module for testing. The performance of this model is shown in Table 1 below. The same 10 sets of differentiated grayscale images were then input into an image recognition model with a pre-segmentation module for testing. The performance of this model is shown in Table 1 below. Table 1. Performance of Image Recognition Model Without Pre-Segmentation Module Table 2. Performance of Image Recognition Model with Pre-Segmentation Module Analysis of Tables 1 and 2 clearly shows that for the image recognition model without a pre-segmentation module, the mean, maximum, and minimum intersection-over-union (IoU) ratios are 84.79%, 91.46%, and 81.56%, respectively. In contrast, the IoU ratios for the image recognition model with a pre-segmentation module are 92.10%, 87.62%, and 96.43%, respectively. This demonstrates that the image recognition model with a pre-segmentation module is more accurate in recognizing the classification regions and boundary contours compared to the model without a pre-segmentation module. Furthermore, the worst deviation in the number of classification regions for the image recognition model without a pre-segmentation module is -14.22%, while the best is only -6.35%. In contrast, the worst deviation for the number of classification regions for the image recognition model with a pre-segmentation module is only 9.81%, while the best is -3.42%. Therefore, the image recognition model with a pre-segmentation module is more accurate in recognizing the number of classification regions compared to the model without a pre-segmentation module. Thus, the image recognition model with a pre-segmentation module is superior in this solution.

[0024] S4. A classification model is built and trained based on the random forest algorithm. Combining regional distribution maps, differential grayscale images, and three-dimensional feature grids, classification features of the region to be classified are extracted and input into the classification model to determine the cultivated land area within the region to be monitored. As one implementation method, classification features include, but are not limited to, the roundness, aspect ratio, texture features, and statistical features of the recognition features of the region to be classified; The method for obtaining the roundness and aspect ratio of the region to be classified is as follows: extract the boundary outline of the region to be classified from the regional distribution map, and then determine the roundness and aspect ratio of the region to be classified based on the boundary outline. This is existing technology and will not be elaborated here. The texture features of the region to be classified include, but are not limited to, energy, entropy, contrast and correlation. The method for obtaining them is as follows: extract the boundary contour of the region to be classified from the regional distribution map and map it onto the differential grayscale image to select the region corresponding to the region to be classified in the differential grayscale image. Then, calculate and analyze the grayscale co-occurrence matrix of the selected region to obtain the texture features of the region to be classified. This is the existing technology and will not be described in detail here. The statistical features include, but are not limited to, maximum, minimum, mean, and variance. The method for obtaining these features is as follows: grids that are mapped one-to-one with the pixels in the region to be classified are selected from the three-dimensional feature grid. For each identification feature, the statistical features of the identification feature are calculated based on the feature data in the selected grids. For example, taking the mean reflectance of the near-infrared band as an example, the near-infrared band reflectance data in all the selected grids are averaged to obtain the mean reflectance of the near-infrared band. This is existing technology and will not be elaborated here. As an example, the classification model includes an input layer for receiving classification features, a random tree construction module for building trees based on the training set, an ensemble module for calculating the final output based on the prediction results of each tree, an output layer for outputting the classification results, and an evaluation and tuning module for performance evaluation and hyperparameter tuning. The random tree construction module constructs random trees based on the conventional techniques of sample selection (i.e., randomly selecting samples from the training set using a bootstrap method to build multiple independent decision trees) - feature selection (i.e., randomly selecting classification features when splitting nodes in each tree, and then selecting the optimal split point based on the classification features) - decision tree construction (i.e., constructing each tree recursively until a stopping condition is met). The ensemble module obtains the final output by averaging the prediction results of each tree. The specific training process is as follows: The monitoring area was determined from multiple cultivated and non-cultivated land areas in previous years. Classification features of each cultivated land area were extracted and labeled "cultivated land," and classification features of each non-cultivated land area were extracted and labeled "non-cultivated land" to construct a second sample set. This second sample set was divided into training, testing, and validation sets in a 70:15:15 ratio. Hyperparameters of the random forest were set (e.g., number of trees between 100-1000, maximum depth between 10-30, maximum number of splits between 2-10). The classification features from the training set were used as input, and corresponding... Labels are used as output to train the classification model. During training, the random tree building module generates multiple decision trees and integrates them to form the final model. The performance metrics of the classification model (such as mean absolute error, root mean square error, etc.) are monitored during training to ensure the effectiveness of model training. Hyperparameters are adjusted based on the validation results of the validation set (such as coefficient of determination, root mean square error) to avoid overfitting or underfitting. Finally, the test set is input into the classification model for performance testing. If the model's recognition error rate is between 1% and 5%, the training is considered complete; otherwise, training is repeated. It should be noted that the classification features are the salient features that identify cultivated land areas and non-cultivated land areas. Therefore, the classification features are used as the input of the classification model. The classification features are a high-dimensional feature set composed of many features. Therefore, the random forest model framework, which can effectively handle high-dimensional features, is adopted to construct the classification model. The specific construction and training of the classification model are conventional techniques for those skilled in the art and will not be described in detail here. It should be noted that the "two-step" approach of this technical solution involves first identifying the various regions to be classified within the area to be monitored using an image recognition model, and then using a classification model to identify and judge each region to determine the setting of cultivated land areas. Compared to using a single model, which requires both segmenting different types of regions and classifying each region when identifying cultivated land areas, this approach simplifies the data processing difficulty of the model and improves the accuracy and efficiency of cultivated land area identification.

[0025] S5 is based on a long short-term memory network to build and train a prediction model to obtain the future cultivated land area, and to statistically analyze the cultivated land area of ​​previous years, the current year and the future, and generate the current cultivated land change rate and the future cultivated land change rate. The logic for obtaining the cultivated land area is as follows: count the number of pixels belonging to the cultivated land area in the regional distribution map, and then combine the spatial resolution of the pixels to determine the cultivated land area in the regional distribution map. For example, if the spatial resolution of a pixel is 15 meters * 15 meters, then the actual area corresponding to one pixel is 225 square meters, and then the cultivated land area is calculated. The calculation logic for the current change in cultivated land is as follows: First, calculate the difference in cultivated land area between the current year and the previous year in the area to be monitored. Then, divide this difference by the cultivated land area of ​​the previous year to obtain the current rate of change in cultivated land. The larger the value, the greater the expansion of cultivated land area in the current year compared to the previous year. Conversely, the smaller the value, the greater the shrinkage of cultivated land area in the current year compared to the previous year. The calculation logic for future changes in arable land is as follows: First, calculate the difference in arable land area between the monitored area in the next year and the current year. Then, divide this difference by the arable land area in the current year to obtain the future arable land change rate. The larger the value, the greater the predicted expansion of arable land area in the next year. Conversely, the smaller the value, the greater the predicted shrinkage of arable land area in the next year. As one implementation, the prediction model includes an input layer, one or more LSTM layers, a fully connected layer, and an output layer, wherein: The input layer is used to receive the cultivated land area over several consecutive years, specifically the cultivated land area over 5-10 consecutive years. The LSTM layer is used to perform feature processing on the data received from the input layer in order to capture the temporal dependencies in the time series data. The number of LSTM units in the LSTM layer is set between 50 and 100. Fully connected layers are used to map the output of the LSTM layer to the predicted values; The output layer is used to output the cultivated land area for the next year. The training process for the prediction model is as follows: The parameters of the prediction model are defined, and multiple sample areas similar to the ecological and human environment of the area to be monitored are identified. Annual time-series data of cultivated land area in the sample areas are obtained to construct a third sample set. The third sample set is divided into a training set and a test set in a 70:30 ratio. The cultivated land area of ​​the training set for several consecutive years is used as input, and the cultivated land area of ​​the next year is used as output to train the prediction model. During the training process, mean squared error is selected as the loss function, and the Adam optimization algorithm is selected to update the model weights. After training, the prediction model is evaluated using the test set. Specifically, mean squared error can be selected as the evaluation index. If the evaluation index reaches the expected value, the training is considered complete. Otherwise, the parameters such as the number of LSTM units, the number of layers, the learning rate, and the batch size are adjusted based on the evaluation results, and the model is trained again. This is existing technology and will not be elaborated here. The model parameters include batch size, feedback training period, and total training period. The feedback training period is set between one-tenth and one-fifth of the total training period. The batch size is the number of samples used in each training iteration, generally between 32 and 256. The total training period is the total number of training iterations, generally between 30 and 100. The feedback training period is set to determine the time node for obtaining model training feedback results. If this time node is too early, the model training feedback results will lack reference value, while if it is too late, the total model training time will be too long. Therefore, the feedback training period is set between one-tenth and one-fifth of the total training period to balance the reference value of model training feedback results and the efficiency of model training. It should be noted that in constructing the third sample set, selecting areas with similar ecological and human environments to the area to be monitored as sample areas helps improve the accuracy of the prediction model in predicting the cultivated land area of ​​the area to be monitored. When predicting the cultivated land area for the next year, the cultivated land area for the current year and the previous several consecutive years is input into the trained prediction model to obtain the predicted value of the cultivated land area for the next year.

[0026] Example 2: Please refer to the figure. This embodiment provides a farmland change monitoring system that integrates deep neural network and remote sensing satellite data, used to execute the above-described farmland change monitoring method integrating deep neural network and remote sensing satellite data, including: The data extraction module is used to acquire satellite remote sensing images of the area to be monitored and extract features to construct a three-dimensional feature grid of the area to be monitored. The three-dimensional feature grid contains identification features for distinguishing between cultivated land and non-cultivated land areas. The grayscale image construction module reduces the dimensionality of the three-dimensional feature raster to obtain a two-dimensional difference matrix based on the differences in the identification features, and then maps the two-dimensional difference matrix into a differentiated grayscale image. The differentiated grayscale image is used to characterize the differences in the identification features within the area to be monitored. An image recognition model is used to process differential grayscale images to obtain a region distribution map. The region distribution map includes at least one region to be classified, which is used to characterize the distribution of different types of regions within the region to be monitored. The image recognition model also has a pre-segmentation module, which pre-segments the differential grayscale images based on a quadtree segmentation strategy. A classification model is used to process the classification features of the region to be classified to determine the cultivated land area within the region to be monitored. The classification features are obtained based on the regional distribution map, the differential grayscale image, and the three-dimensional feature grid. The predictive model is used to obtain the future cultivated land area and to statistically analyze the cultivated land area in previous years, the current year, and the future, generating the current cultivated land change rate and the future cultivated land change rate.

[0027] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0028] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. 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 by software, 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.

[0029] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0030] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for monitoring farmland change by integrating deep neural networks and remote sensing satellite data, characterized in that, Includes the following steps: S1. Collect satellite remote sensing images of the area to be monitored and extract features to construct a three-dimensional feature grid of the area to be monitored. The three-dimensional feature grid contains identification features for distinguishing between cultivated land and non-cultivated land. S2, based on the differences in the identification features, the three-dimensional feature grid is reduced to obtain a two-dimensional difference matrix, and the two-dimensional difference matrix is ​​mapped into a differentiated grayscale image. The differentiated grayscale image is used to characterize the differences in the identification features within the area to be monitored. S3, based on convolutional neural networks, constructs and trains an image recognition model with a pre-segmentation module. The pre-segmentation module pre-segments the differential grayscale images based on a quadtree segmentation strategy. During training, the quadtree segmentation strategy is iteratively optimized based on the model performance. The differential grayscale images are input into the trained image recognition model to obtain a region distribution map, which includes at least one region to be classified, used to characterize the distribution of different types of regions within the region to be monitored. S4. A classification model is built and trained based on the random forest algorithm. Combining regional distribution maps, differential grayscale images, and three-dimensional feature grids, classification features of the region to be classified are extracted and input into the classification model to determine the cultivated land area within the region to be monitored. S5 is based on a long short-term memory network to build and train a prediction model to obtain the future cultivated land area, and statistically analyze the cultivated land area of ​​previous years, the current year and the future to generate the current cultivated land change rate and the future cultivated land change rate.

2. The method for monitoring farmland changes by fusing deep neural networks and remote sensing satellite data according to claim 1, characterized in that, The logic for obtaining the 3D feature raster is as follows: S11. Acquire satellite remote sensing images of the area to be monitored and perform preprocessing operations on the satellite remote sensing images, including but not limited to radiometric correction, atmospheric correction and noise removal. S12, extract features from the preprocessed satellite remote sensing image to obtain the identification feature vector of each pixel. The identification features are used to distinguish between cultivated land and non-cultivated land areas, including but not limited to near-infrared band reflectance, green band reflectance, red band reflectance, short-wave infrared band reflectance, normalized vegetation index, differential vegetation index, enhanced vegetation index and moisture index. S13. Construct a three-dimensional blank raster with the same size as the satellite remote sensing image, and input the identification features of each pixel into the three-dimensional blank raster to form a three-dimensional feature raster.

3. The method for monitoring farmland changes by fusing deep neural networks and remote sensing satellite data according to claim 2, characterized in that, The logic for dimensionality reduction of the three-dimensional feature raster is as follows: For the identification feature vector corresponding to any pixel in the satellite remote sensing image, the identification feature difference coefficient between it and its neighboring pixels is calculated, and the identification feature difference coefficients are summarized to form a two-dimensional difference matrix.

4. The method for monitoring farmland changes by fusing deep neural networks and remote sensing satellite data according to claim 1, characterized in that, The method for mapping a two-dimensional difference matrix to a differential grayscale image is as follows: 1) For any element value in the two-dimensional difference matrix, normalize it to convert it into a gray level between 0 and 255; 2) Read the spatial resolution, number of rows and columns of the satellite remote sensing image, create a blank grayscale image with the same size as the satellite remote sensing image, and assign the grayscale value of each element in the two-dimensional difference matrix to the corresponding pixel of the blank grayscale image one by one to obtain the differentiated grayscale image.

5. The method for monitoring farmland changes by fusing deep neural networks and remote sensing satellite data according to claim 1, characterized in that, The specific architecture of the image recognition model is as follows: The pre-segmentation module pre-segments the differentiated grayscale image based on the quadtree segmentation strategy to generate a pre-segmented image, which is a grayscale image with boundary contours. The input layer is used to receive the pre-segmented image output by the segmentation module; The image processing layer group is used to process the pre-segmented image. It includes three groups of image processing layers connected in sequence, specifically the first image processing layer, the second image processing layer and the third image processing layer. Each image processing layer consists of a convolutional layer, a batch normalization layer and a max pooling layer. Flattening layers are used to flatten the output of an image processing layer group into a one-dimensional array; Fully connected layers are used to further process the flattened features and learn higher-level feature representations; The output layer generates a region distribution map of the same size as the pre-segmented image. The region distribution map is a binarized image, which consists of boundary contours and regions to be classified.

6. The method for monitoring farmland changes by fusing deep neural networks and remote sensing satellite data according to claim 5, characterized in that, The process of training an image recognition model is as follows: S31, acquire the differentiated grayscale and color images of the area to be monitored in previous years, and mark the boundary contours on the color image based on expert analysis to divide the color image into several different types of areas; S32, after mapping the labeled boundary contours onto the differential grayscale image, perform binarization to generate a region distribution map corresponding to the differential grayscale image. Summarize multiple sets of one-to-one corresponding differential grayscale images and region distribution maps to construct the first sample set. Divide the first sample set into a training set and a test set according to a 7:3 allocation method. S33, initialize the segmentation parameters of the pre-segmentation module and the hyperparameters of the image recognition model; The segmentation parameters of the pre-segmentation module include block threshold, minimum region threshold, grayscale segmentation threshold and uniformity standard, and hyperparameters include learning rate and batch size. S34. The differentiated grayscale images in the training set are input into the pre-segmentation module for pre-segmentation, and then input into the image recognition model for model training. At the end of each iteration of training, the differentiated grayscale images in the test set are processed by the pre-segmentation module and input into the image recognition model for testing. The segmentation parameters are optimized based on the test results and then trained again until the test results meet the standards.

7. The method for monitoring farmland changes by fusing deep neural networks and remote sensing satellite data according to claim 6, characterized in that, The logic for optimizing the image recognition model is as follows: 1) Input the differentiated grayscale images of the same training batch one by one into the pre-segmentation module. The pre-segmentation module first analyzes the distribution of grayscale values ​​of pixels in each differentiated grayscale image to determine the segmentation threshold of each differentiated grayscale image. Based on the segmentation threshold, the differentiated grayscale image is segmented once to form multiple sub-regions. 2) Based on the quadtree segmentation strategy, the sub-regions in the differential grayscale image are simultaneously segmented to generate a pre-segmented image; 3) Input the pre-segmented images from the same training batch into the image recognition model for training. During the training process, use the binary cross-entropy function as the loss function and use the Adam or SGD optimization algorithm to adjust the model parameters to minimize the loss function. 4) After each round of training, the test set is input into the image recognition model for testing. The test results include the crossover ratio and the number deviation. The crossover ratio and the number deviation are combined to determine whether the image recognition model meets the requirements. If it meets the requirements, training is stopped; otherwise, the segmentation parameters are optimized by combining the crossover ratio and the number deviation and training continues. The formula for calculating the intersection-union ratio is as follows: In the formula, This represents the set of pixels belonging to the boundary contour in the region distribution map output by the image recognition model during the t-th iteration of training. This represents the set of pixels belonging to the boundary contour in the actual regional distribution map. This represents the set of pixels belonging to the region to be classified in the region distribution map output by the image recognition model during the t-th iteration of training. This represents the set of pixels in the actual regional distribution map that belong to the region to be classified. This represents the intersection-union ratio (IU) in the t-th iteration of training, where t is the index of the iteration round. The formula for calculating the number deviation is as follows: In the formula, This represents the number of regions to be classified in the region distribution map output by the image recognition model during the t-th iteration of training. This indicates the number of regions to be classified in the actual regional distribution map. This represents the deviation of the number of iterations in the t-th training round; The logic for determining whether an image recognition model meets the requirements is as follows: when the cross-union ratio (CUNR) is greater than the CUNR threshold and the number deviation is within the number deviation range, the image recognition model is considered to meet the requirements; otherwise, it is considered not to meet the requirements.

8. The method for monitoring farmland changes by fusing deep neural networks and remote sensing satellite data according to claim 7, characterized in that, The logic for optimizing segmentation parameters by combining intersection-over-union ratio and number deviation is as follows: Based on the number deviation, the block threshold is optimized. The optimization logic is as follows: when the number deviation is less than 0, the block threshold is gradually reduced as the absolute value of the number deviation increases; when the number deviation is greater than 0, the block threshold is gradually increased as the number deviation increases. Based on the number deviation, the minimum region threshold is optimized. The optimization logic is as follows: when the number deviation is less than 0, the minimum region threshold is gradually reduced as the absolute value of the number deviation increases; when the number deviation is greater than 0, the minimum region threshold is gradually increased as the number deviation increases. Based on the cross-union ratio, the uniformity standard is optimized. The optimization logic is: as the cross-union ratio decreases, the uniformity standard is gradually reduced. Based on the intersection-over-union ratio (IoU), the grayscale segmentation threshold is optimized. The optimization logic is as follows: as the IoU decreases, the grayscale segmentation threshold is gradually reduced.

9. The method for monitoring farmland changes by fusing deep neural networks and remote sensing satellite data according to claim 1, characterized in that, The classification features include, but are not limited to, the roundness, aspect ratio, texture features, and statistical features of the recognition features of the region to be classified. The texture features include, but are not limited to, energy, entropy, contrast, and correlation. The statistical features include, but are not limited to, maximum value, minimum value, mean, and variance.

10. A farmland change monitoring system integrating deep neural networks and remote sensing satellite data, used to execute the farmland change monitoring method integrating deep neural networks and remote sensing satellite data as described in any one of claims 1-9, characterized in that, include: The data extraction module is used to acquire satellite remote sensing images of the area to be monitored and extract features to construct a three-dimensional feature grid of the area to be monitored. The three-dimensional feature grid contains identification features for distinguishing between cultivated land and non-cultivated land areas. The grayscale image construction module reduces the dimensionality of the three-dimensional feature raster to obtain a two-dimensional difference matrix based on the differences in the identification features, and then maps the two-dimensional difference matrix into a differentiated grayscale image. The differentiated grayscale image is used to characterize the differences in the identification features within the area to be monitored. An image recognition model is used to process differential grayscale images to obtain a region distribution map. The region distribution map includes at least one region to be classified, which is used to characterize the distribution of different types of regions within the region to be monitored. The image recognition model also has a pre-segmentation module, which pre-segments the differential grayscale images based on a quadtree segmentation strategy. A classification model is used to process the classification features of the region to be classified to determine the cultivated land area within the region to be monitored. The classification features are obtained based on the regional distribution map, the differential grayscale image, and the three-dimensional feature grid. The predictive model is used to obtain the future cultivated land area and to statistically analyze the cultivated land area in previous years, the current year, and the future, generating the current cultivated land change rate and the future cultivated land change rate.

Images