Large-scale remote sensing fragmented cultivated land change detection algorithm based on coarse screening and fine inspection

Abstract

The present invention discloses a large-scale remote sensing fragmented farmland change detection algorithm based on coarse screening and fine inspection, comprising the following steps: Step S1: Acquire two large-scale farmland remote sensing images of the same location at different times and preprocess them separately; then crop the preprocessed images to obtain a pre-phase sub-image group and a post-phase sub-image group; Step S2: Select a histogram standardization method to unify the style of the pre-phase sub-image group and the post-phase sub-image group; Step S3: Coarsely screen the farmland change images to obtain a new image dataset, wherein the farmland change images are the pre-phase sub-image group and the post-phase sub-image group; Step S4: Finely detect and locate the farmland change areas in the new image dataset to obtain the farmland change status. This algorithm solves the problems of existing change detection methods used to detect farmland changes, such as the difficulty in obtaining label information, the small number of farmland change areas and the difficulty in detecting them, and the complexity of image acquisition and application.

Description

Technical field

[0001] The present invention belongs to the technical field of cultivated land change detection, and in particular relates to a large-scale remote sensing fragmented cultivated land change detection algorithm based on coarse screening and fine inspection. [Background Technology]

[0002] Land resources are the material foundation for human survival and development. Researching cultivated land and its spatiotemporal changes, and gaining a deeper understanding of the current status of regional cultivated land resources and the causes of these changes, plays a crucial role in revealing the patterns of cultivated land change, ensuring food security, and promoting social stability. As a highly practical remote sensing analysis technique, change detection aims to identify interesting change information while filtering out irrelevant interference. It has been widely applied in various fields, including land cover and land use change analysis, urban planning, forest monitoring, and disaster assessment.

[0003] With the continuous development of convolutional neural networks, many deep learning-based methods have been applied to the field of change detection in recent years. These methods can be divided into supervised and unsupervised change detection methods, depending on whether the training data contains label information. Supervised methods, with their high-precision classification capabilities, have achieved accurate change detection in fields such as urban planning and environmental monitoring. Unsupervised methods, on the other hand, have demonstrated strong practicality in scenarios such as remote sensing image processing and anomaly detection due to their lack of label information. However, both methods face challenges in application, such as data acquisition and computational complexity. Therefore, they need to be selected and improved based on specific needs to promote the further development of change detection technology.

[0004] In recent years, remote sensing technology has been developing towards multi-sensor, multi-platform, multi-angle, and high spatial, spectral, and temporal resolution. The amount of image data available daily is increasing by terabytes. Regarding the current large-scale and large-scale nature of cultivated land detection, the above methods have the following drawbacks:

[0005] (1) Difficulty in obtaining label information: Due to the diversity and complexity of land cover types, it is particularly difficult to obtain sufficient and representative label information. This not only involves the investment of manpower and material resources, but also involves ensuring the quality and consistency of labeling. At the same time, due to the seasonal changes in cultivated land and the differences between different regions, the acquisition of labeled samples also needs to consider temporal and spatial factors, which further increases the difficulty and cost of labeling.

[0006] (2) Few areas of cultivated land change and difficult to detect: The relative stability and periodic changes of cultivated land result in relatively few areas experiencing actual changes. This results in the number of positive samples (i.e., areas of change) being far less than the number of negative samples (i.e., areas that have not changed) when training deep learning models, making it difficult for the model to effectively learn the characteristics of the changes. In addition, due to the subtlety and complexity of cultivated land changes, even if changes do occur, they may be difficult to accurately detect due to the influence of multiple factors (such as climate, crop type, planting method, etc.).

[0007] (3) The image acquisition and application process are complex: Remote sensing images are usually high-resolution and contain a large amount of geographic information. For specific application scenarios, the original images often need to go through a series of processing steps before they can be applied to extract the geographic area of ​​interest and focus on the detailed features of the area of ​​interest. [Summary of the invention]

[0008] The purpose of the present invention is to provide a large-scale remote sensing fragmented farmland change detection algorithm based on coarse screening and fine inspection, so as to solve the problems of existing change detection methods when used to detect farmland changes, such as difficulty in obtaining label information, small number of farmland change areas and difficulty in detection, and complex image acquisition and application.

[0009] The present invention adopts the following technical solution: a large-scale remote sensing fragmented farmland change detection algorithm based on coarse screening and fine inspection, comprising the following steps:

[0010] Step S1, collecting two large-scale farmland remote sensing images of the same location at different times, the two large-scale farmland remote sensing images being a front-phase image and a back-phase image; and preprocessing the front-phase image and the back-phase image respectively;

[0011] Then, the pre-processed front-phase image and the back-phase image are cropped according to a selected size to obtain two groups of small-sized farmland remote sensing images; the two groups of small-sized farmland remote sensing images are respectively a front-phase sub-image group and a back-phase sub-image group;

[0012] Step S2: selecting a histogram normalization method to unify the styles of the front-phase sub-image group and the back-phase sub-image group;

[0013] Step S3, coarsely screening the cultivated land change image to obtain a new image dataset, wherein the cultivated land change image comprises a front-phase sub-image group and a back-phase sub-image group;

[0014] Step S4: Finely detect and locate the cultivated land change area in the new image dataset to obtain the cultivated land change status.

[0015] Furthermore, the specific method of step S3 is:

[0016] Step 311: Using a feature extraction algorithm, extract features of the same dimension from all sub-images in the new front-phase sub-image group and the back-phase sub-image group under the same algorithm. The features of the front-phase image and the back-phase image can be represented as an N×K matrix, where N is the feature dimension and K is the number of sub-images.

[0017] Step 312: Take the i-th sub-image feature in the later phase sub-image group as the object to be reconstructed, and select k sub-image features around the previous phase sub-image feature at the position corresponding to the i-th sub-image feature in the later phase group. 2 The image features are used to build a dictionary to achieve sparse reconstruction;

[0018] Step 313: Calculate the reconstruction error according to formula (1) and sort the reconstruction errors from large to small. A larger error indicates a higher possibility of change. That is, sort according to the reconstruction error and discard image pairs with an error less than the error threshold. The remaining images are the new image dataset.

[0019]

[0020] Among them, ||·||2 is the l2-norm, ε is the error threshold. If the error threshold is less than the error threshold, it can be considered as no change. The number of words in the dictionary M = k 2 ≤K, K is the number of sub-images obtained after cropping each large-scale farmland remote sensing image, k is the width of the surrounding neighborhood, x i is a sparse representation matrix.

[0021] Furthermore, the specific method of step S3 is:

[0022] Step 321: grayscale all sub-images in the new front-phase sub-image group and the back-phase sub-image group, and then form a front-phase sub-image group matrix and a back-phase sub-image group matrix respectively;

[0023] Step 322: using PCA to reduce the dimensions of the matrix of the front-phase sub-image group and the matrix of the back-phase sub-image group to the same N-dimensionality, and obtain the reduced-dimensionality key feature matrix A of the front-phase sub-image group and the reduced-dimensionality key feature matrix B of the back-phase sub-image group;

[0024] Step 323: According to formulas (2) to (4), the corresponding positions of the key feature matrix A of the previous phase sub-image group and the key feature matrix B of the next phase sub-image group are subtracted to obtain the key feature difference matrix C;

[0025] Then, the key features are made into a difference matrix C and the absolute values ​​are summed up by column to obtain the key feature difference threshold matrix D. The matrix is ​​sorted according to the absolute value summation results, and the image pairs whose absolute value summation results are less than the difference threshold are removed; the remaining images are the new image dataset;

[0026]

[0027] in, δ is the difference threshold, below which the change is negligible; A and B are the key feature matrices of the pre-phase sub-image group and the post-phase sub-image group, respectively; C is the key feature difference matrix; D is the key feature difference threshold matrix; N is the matrix of the pre-phase sub-image group and the post-phase sub-image group, the two image matrices are reduced to the same dimension; K is the number of sub-images obtained after cropping each pre-phase image or post-phase image; and δ is the difference threshold.

[0028] Furthermore, the specific method of step S2 is:

[0029] First, the grayscale histograms of the front-phase sub-image group and the back-phase sub-image group are analyzed to obtain the brightness and contrast distribution of the front-phase sub-image group and the back-phase sub-image group;

[0030] Secondly, according to the histogram of the subsequent phase sub-image group and the histogram of the previous phase sub-image group, the pixel value of the previous phase sub-image group is mapped to a new value by a mapping function, so that the new value is within the dynamic range of the histogram of the subsequent phase sub-image group;

[0031] Finally, the mapping function is applied to each pixel of the previous phase sub-image group to generate a new previous phase sub-image group.

[0032] Furthermore, the specific method of step S4 is:

[0033] Step 4.1: Feature Extraction

[0034] The paired image data in the new image dataset obtained after the coarse screening in step 2 is used as input, and the spatial features of the image are extracted through a series of convolutional layers; the features are downsampled using a maximum pooling or average pooling layer to retain the main feature information; activation functions such as ReLU are used to increase the nonlinearity of the features; and the feature maps of the front-phase sub-image and the feature maps of the back-phase sub-image are obtained;

[0035] Step 4.2: Change Extraction

[0036] Aligning the feature map of the front-phase sub-image and the feature map of the back-phase sub-image obtained in step 4.1; calculating the difference between the feature map of the front-phase sub-image and the feature map of the back-phase sub-image to obtain a change feature map; performing threshold processing on the change feature map;

[0037] Step 4.3: Generate a change graph

[0038] The change feature map is input into a classification network to generate a probability map of changed and unchanged areas; and the probability map is subjected to threshold segmentation to obtain a binary change map.

[0039] The beneficial effects of the present invention are as follows: First, the present invention targets the non-change areas of interest that are widely present in remote sensing images of cultivated land. The present invention introduces two highly efficient coarse detection algorithms, which can quickly eliminate those areas that are irrelevant to the target of interest or have insignificant changes, thereby improving the overall analysis efficiency. Secondly, in response to the problem of insufficient unlabeled or labeled training samples that are prevalent in remote sensing image analysis, the coarse screening module of the present invention does not rely on sample labels, and the fine detection module adopts an unsupervised or weakly supervised learning model to complete the detection of change areas by extracting and learning the intrinsic features of the image. Finally, in response to the application complexity problem from image acquisition to landing, the coarse screening module quickly identifies and screens images of non-cultivated land changes, and the fine detection module specifically extracts and locates the areas of interest. The cascade processing of the two modules realizes full-process change detection. This innovation not only solves the problem of lack of labeled data, but also makes the change detection algorithm more versatile and flexible, and can adapt to remote sensing image analysis tasks of different scenarios and needs.

Brief Description of the Drawings

[0040] Figure 1 A diagram of a coarse screening method for characterizing changes by reconstructing characteristic errors according to the present invention;

[0041] Figure 2 A diagram of a coarse screening method for characterizing changes in key characteristic errors according to the present invention;

[0042] Figure 3 A diagram of the method for fine detection and positioning of cultivated land change areas according to the present invention. [Specific implementation method]

[0043] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0044] The present invention provides a large-scale remote sensing fragmented farmland change detection algorithm based on coarse screening and fine inspection, comprising the following steps:

[0045] Step S1, collecting two large-scale remote sensing images of cultivated land at the same location at different times, the two large-scale remote sensing images of cultivated land being a front-phase image and a back-phase image respectively; preprocessing the front-phase image and the back-phase image respectively; the preprocessing includes geometric correction and image registration;

[0046] Then, the pre-processed front-phase image and the back-phase image are cropped according to a selected size to obtain two groups of small-sized farmland remote sensing images; the two groups of small-sized farmland remote sensing images are respectively a front-phase sub-image group and a back-phase sub-image group;

[0047] Step S2: selecting a histogram normalization method to unify the styles of the front-phase sub-image group and the back-phase sub-image group;

[0048] Step S3, coarsely screening the cultivated land change image to obtain a new image dataset, wherein the cultivated land change image comprises a front-phase sub-image group and a back-phase sub-image group;

[0049] Step S4: Finely detect and locate the cultivated land change area in the new image dataset to obtain the cultivated land change status.

[0050] This invention aims to detect farmland changes using a "coarse screening" followed by "fine inspection" strategy. Traditional machine learning methods are used to analyze and quantify the differences between the previous and next time-phase sub-images, discarding sub-image pairs with relatively small differences. This allows for rapid identification of areas with significant changes, reducing the amount of data required for subsequent processing and improving efficiency. In the fine inspection phase, deep learning methods are used to perform detailed inspections on the retained image pairs and locate changes in areas of farmland damage.

[0051] The proposed dual-module change detection method, which involves "coarse screening" followed by "fine inspection," makes it possible to use raw remote sensing imagery (large-scale pre- and post-phase images) directly for detection. The coarse screening module first operates on the raw remote sensing imagery (large-scale pre- and post-phase images), using an efficient algorithm to quickly identify and screen images of non-arable land changes. This step effectively reduces the amount of data to be processed later and initially eliminates a large amount of irrelevant information. Subsequently, the fine inspection module conducts an in-depth analysis of candidate change areas from the coarse screening results, employing a deep learning model to meticulously classify and verify these candidate areas to ensure the accuracy and reliability of the detection results.

[0052] Guided by practical application needs, the coarse screening module uses traditional methods to analyze and quantify the differences between the previous and next phase images to address the difficulty of obtaining sample change labels in images, completely eliminating the reliance on labels. The fine detection module can consider adopting an unsupervised or weakly supervised learning model, using a small amount of incomplete or noisy label information to guide model learning, thereby reducing reliance on labels while achieving good change detection effects.

[0053] In some embodiments, as Figure 1 As shown, the first coarse screening method in step S3 is to characterize the change by the reconstruction error, and the specific process is as follows:

[0054] Step 311: Using a feature extraction algorithm, extract features of the same dimension from all sub-images in the new front-phase sub-image group and the back-phase sub-image group under the same algorithm. The features of the front-phase image and the back-phase image can be represented as an N×K matrix, where N is the feature dimension and K is the number of sub-images.

[0055] Step 312: Take the i-th sub-image feature in the later phase sub-image group as the object to be reconstructed, and select k sub-image features around the previous phase sub-image feature at the position corresponding to the i-th sub-image feature in the later phase group. 2 The dictionary is constructed based on the features of each image to achieve sparse reconstruction of the features of the i-th sub-image in the subsequent phase sub-image group.

[0056] Step 313: Calculate the reconstruction error according to formula (1) and sort the reconstruction errors from large to small. A larger error indicates a higher possibility of change. That is, by sorting according to the reconstruction error, image pairs with an error less than the error threshold can be discarded; the remaining images are the new image dataset;

[0057]

[0058] Among them, ||·||2 is the l2-norm, ε is the error threshold. If the error threshold is less than the error threshold, it can be considered as no change. The number of words in the dictionary M = k 2 ≤K, K is the number of sub-images obtained after cropping each original image, k is the width of the surrounding neighborhood, x i is a sparse representation matrix.

[0059] This method focuses on identifying features that are sparsely distributed in the dataset but potentially significant. It can quickly identify significant changes in sparse features in cultivated land images, thereby quickly screening cultivated land images for further analysis. The advantage of this method is its high efficiency, making it suitable for processing large datasets. It can quickly identify areas of potential change, saving significant computing resources for subsequent detailed analysis. Furthermore, sparse features are often associated with anomalies or unique phenomena in the data. In cultivated land monitoring, significant changes in sparse features may indicate changes in land use or soil degradation. This method is generally insensitive to noise, helping to reduce false positives and showing high sensitivity for promptly identifying cultivated land issues.

[0060] In some embodiments, as Figure 2 As shown, the second coarse screening method in step S3 is to characterize the changes based on the key feature differences, and the specific process is as follows:

[0061] Step 321: grayscale all sub-images in the new front-phase sub-image group and the back-phase sub-image group, and then form a front-phase sub-image group matrix and a back-phase sub-image group matrix respectively;

[0062] Step 322: PCA is used to reduce the dimensions of the matrix of the front-phase sub-image group and the matrix of the back-phase sub-image group to the same N-dimensionality, and the reduced dimensions are obtained, respectively, to obtain the key feature matrix A of the front-phase sub-image group and the key feature matrix B of the back-phase sub-image group, so as to highlight the key information and eliminate the interference information. PCA is a principal component analysis method; PCA is a principal component analysis method;

[0063] Step 323: According to formulas (2) to (4), the corresponding positions of the key feature matrix A of the previous phase sub-image group and the key feature matrix B of the next phase sub-image group are subtracted to obtain the key feature difference matrix C;

[0064] Then, the key features are made into a difference matrix C and the absolute values ​​are summed up by column to obtain the key feature difference threshold matrix D. The matrix is ​​sorted according to the absolute value summation results, and the image pairs whose absolute value summation results are less than the difference threshold are removed; the remaining images are the new image dataset;

[0065]

[0066] in, δ is the difference threshold, below which the change is negligible; A and B are the key feature matrices of the pre-phase sub-image group and the post-phase sub-image group, respectively; C is the key feature difference matrix; D is the key feature difference threshold matrix; N is the matrix of the pre-phase sub-image group and the post-phase sub-image group, the two image matrices are reduced to the same dimension; K is the number of sub-images obtained after cropping each pre-phase image or post-phase image; and δ is the difference threshold.

[0067] The second approach focuses on extracting key features, identifying those that significantly influence changes in the surface morphology or structure of cultivated land. These features are often closely related to factors such as crop growth, soil quality, and topography. This approach is suitable for detailed analysis of physical or morphological changes in cultivated land, capturing subtle yet significant surface fluctuations. Compared to sparse feature analysis, this approach focuses more on detail, making it more reliable in assessing changes in images.

[0068] In some embodiments, the specific method of step S2 is:

[0069] First, the grayscale histograms of the front-phase sub-image group and the back-phase sub-image group are analyzed to obtain the brightness and contrast distribution of the front-phase sub-image group and the back-phase sub-image group;

[0070] Secondly, according to the histogram of the subsequent phase sub-image group and the histogram of the previous phase sub-image group, the pixel value of the previous phase sub-image group is mapped to a new value by a mapping function, so that the new value is within the dynamic range of the histogram of the subsequent phase sub-image group;

[0071] Finally, the mapping function is applied to each pixel of the previous phase sub-image group to generate a new previous phase sub-image group.

[0072] In some embodiments, as Figure 3 As shown, the specific method of step S4 is:

[0073] Step 4.1: Feature Extraction

[0074] The paired image data in the new image dataset obtained after the coarse screening in step 2 is used as input. The paired image data is the front phase sub-image and the back phase sub-image. The spatial features of the image are extracted through a series of convolutional layers.

[0075] Use the maximum pooling or average pooling layer to downsample features and retain the main feature information;

[0076] Use activation functions such as ReLU to increase the nonlinearity of features; obtain feature maps of the previous phase sub-image and feature maps of the next phase sub-image for subsequent change extraction;

[0077] Step 4.2: Change Extraction

[0078] Align the feature map of the front-phase sub-image and the feature map of the back-phase sub-image obtained in step 4.1 to ensure the consistency of corresponding pixel positions; calculate the difference between the feature map of the front-phase sub-image and the feature map of the back-phase sub-image to obtain a change feature map; perform threshold processing on the change feature map to extract significantly changed areas; and use an attention mechanism to strengthen attention to the changed areas and improve the accuracy of change detection.

[0079] Step 4.3: Generate a change graph

[0080] The change feature map is input into a classification network to generate a probability map of changed and unchanged areas; the probability map is threshold segmented to obtain a binary change map, from which the change status of the cultivated land can be known.

[0081] Example

[0082] The effects of the present invention can be further illustrated by the following experimental results and analysis:

[0083] 1. Experimental data: The experimental data of the present invention are two pairs of optical remote sensing images located in Shangyu District, Shaoxing City, Zhejiang Province. The acquisition satellites are Gaofen-1B and Beijing-2, respectively. The sizes of the experimental areas are 40598×35178 and 21243×19041, respectively. During the experiment, they are uniformly cropped to 512×512 size to form their respective sub-image groups.

[0084] 2. Experimental methods

[0085] Two methods proposed by the present invention:

[0086] (1) A coarse screening method that uses reconstruction feature errors to characterize changes;

[0087] (2) A coarse screening method that uses key characteristic errors to characterize changes.

[0088] Table 1 Comparison of the results of rough screening methods for optical remote sensing images of Shangyu District (taken by Gaofen 1B)

[0089]

[0090]

[0091] Table 2 Comparison of optical remote sensing image coarse screening methods in Shangyu District (photographed by Beijing No. 2)

[0092]

[0093] As can be seen from Tables 1 and 2, as the number of words k in the dictionary increases 2 Or as the value of dimension reduction dimension N increases, the image removal rate of the two coarse screening methods increases. In the ShangYu-GF1b dataset, when k 2 =7×7 or N=512, the image removal rate reaches the maximum, which is 37.07% and 46.59% respectively. In the ShangYu-BJ2 dataset, when k 2 =9×9 and N=1024, the image removal rates are 15.79% and 19.80%.

[0094] Therefore, the above analysis shows that the two coarse screening methods proposed in the present invention can both screen out sub-images without any cultivated land changes, which indicates that the method proposed in the present invention is effective and can provide a suitable new image dataset for subsequent change detection to achieve fine detection and positioning of cultivated land change areas, thereby obtaining the cultivated land change status.

Claims

1. A large-scale remote sensing fragmented farmland change detection algorithm based on coarse screening and fine inspection, characterized by: The following steps are involved: Step S1, collecting two large-scale farmland remote sensing images of the same location at different times, the two large-scale farmland remote sensing images being a front-phase image and a back-phase image; and preprocessing the front-phase image and the back-phase image respectively; Then, the pre-processed front-phase image and the back-phase image are cropped according to a selected size to obtain two groups of small-sized farmland remote sensing images; the two groups of small-sized farmland remote sensing images are respectively a front-phase sub-image group and a back-phase sub-image group; Step S2: selecting a histogram normalization method to unify the styles of the front-phase sub-image group and the back-phase sub-image group; Step S3, performing a coarse screening of the cultivated land change images to obtain a new image dataset, wherein the cultivated land change images are a group of sub-images of the front phase after style unification and a group of sub-images of the back phase after style unification; The specific method of step S3 is: Step 311: Using a feature extraction algorithm, extract features of the same dimension from all sub-images in the style-unified front-phase sub-image group and the style-unified back-phase sub-image group. The features of both the front-phase image and the style-unified back-phase image can be represented as an n×K matrix, where n is the feature dimension and K is the number of sub-images. Step 312: Take the i-th sub-image feature in the later phase sub-image group as the object to be reconstructed, and select k sub-image features around the previous phase sub-image feature at the position corresponding to the i-th sub-image feature in the later phase group. 2 The image features are used to build a dictionary to achieve sparse reconstruction; Step 313: Calculate the reconstruction error according to formula (1) and sort the reconstruction errors from large to small. A larger error indicates a higher possibility of change. That is, sort according to the reconstruction error and discard image pairs with an error less than the error threshold. The remaining images are the new image dataset. Among them, ||·||2 is the l2-norm, ε is the error threshold. If the error threshold is less than the error threshold, it can be considered as no change. The number of words in the dictionary M = k 2 ≤K, K is the number of sub-images obtained after cropping each large-scale farmland remote sensing image, k is the width of the surrounding neighborhood, x i is a sparse representation matrix; Step S4: finely detecting and locating the cultivated land change area in the new image dataset to obtain the cultivated land change status; The specific method of step S4 is: Step 4.1: Feature Extraction The paired image data in the new image dataset obtained after the coarse screening in step 3 is used as input, and the spatial features of the image are extracted through a series of convolutional layers; the features are downsampled using the maximum pooling or average pooling layer to retain important feature information; the ReLU activation function is used to increase the nonlinearity of the features; and the feature maps of the front phase sub-image and the feature maps of the back phase sub-image are obtained; Step 4.2: Change Extraction Aligning the feature map of the front-phase sub-image and the feature map of the back-phase sub-image obtained in step 4.1; calculating the difference between the feature map of the front-phase sub-image and the feature map of the back-phase sub-image to obtain a change feature map; performing threshold processing on the change feature map; Step 4.3: Generate a change graph The change feature map is input into a classification network to generate a probability map of changed and unchanged areas; and the probability map is subjected to threshold segmentation to obtain a binary change map.

2. The large-scale remote sensing fragmented farmland change detection algorithm based on coarse screening and fine inspection according to claim 1 is characterized in that: The specific method of step S3 may also be: Step 321: grayscale all sub-images in the front-phase sub-image group and the back-phase sub-image group after the style unification, and then form a front-phase sub-image group matrix and a back-phase sub-image group matrix respectively; Step 322: using PCA to reduce the dimensions of the matrix of the front-phase sub-image group and the matrix of the back-phase sub-image group to the same N-dimensionality, and obtain the reduced-dimensionality key feature matrix A of the front-phase sub-image group and the reduced-dimensionality key feature matrix B of the back-phase sub-image group; Step 323: According to formulas (2) to (4), the corresponding positions of the key feature matrix A of the previous phase sub-image group and the key feature matrix B of the next phase sub-image group are subtracted to obtain the key feature difference matrix C; Then, the key features are made into a difference matrix C and the absolute values ​​are summed up by column to obtain the key feature difference threshold matrix D. The matrix is ​​sorted according to the absolute value summation results, and the image pairs whose absolute value summation results are less than the difference threshold are removed; the remaining images are the new image dataset; in, δ is the difference threshold, below which the change is negligible; A and B are the key feature matrices of the pre-phase sub-image group and the post-phase sub-image group, respectively; C is the key feature difference matrix; D is the key feature difference threshold matrix; N is the matrix of the pre-phase sub-image group and the post-phase sub-image group, the two image matrices are reduced to the same dimension; K is the number of sub-images obtained after cropping each pre-phase image or post-phase image; and δ is the difference threshold.

3. The large-scale remote sensing fragmented farmland change detection algorithm based on coarse screening and fine inspection according to claim 1 or 2, characterized in that: The specific method of step S2 is: First, the grayscale histograms of the front-phase sub-image group and the back-phase sub-image group are analyzed to obtain the brightness and contrast distribution of the front-phase sub-image group and the back-phase sub-image group; Secondly, according to the histogram of the subsequent phase sub-image group and the histogram of the previous phase sub-image group, the pixel value of the previous phase sub-image group is mapped to a new value by a mapping function, so that the new value is within the dynamic range of the histogram of the subsequent phase sub-image group; Finally, the mapping function is applied to each pixel of the previous phase sub-image group to generate the previous phase sub-image group with unified style.

Images