A cultivated land change detection method and device based on satellite remote sensing data

Abstract

The application discloses a kind of based on satellite remote sensing data's cultivated land change detection method and equipment, the method includes: obtaining the current satellite remote sensing image of target cultivated land area and the information of equinox and weather when the image is collected;Obtain the satellite remote sensing image of target cultivated land area matched with the equinox and weather information as reference image;The image and the reference image are divided, obtain M first sub-image and corresponding M second sub-image;The contour similarity of first sub-image and second sub-image in each pair of sub-image is calculated, when the contour similarity of a pair of sub-image is less than first threshold, the pair of sub-image is detected, and the class similarity between the pair of sub-image is determined according to the detection result;If class similarity is less than second threshold, and the detection result of the first sub-image of the pair of sub-image contains object that does not appear in cultivated land, it indicates that there is the situation of illegal occupation of cultivated land in the cultivated land area of the first sub-image.The application can improve detection accuracy.

Description

Technical Field

[0001] This invention belongs to the field of satellite remote sensing technology, specifically relating to a method and equipment for detecting changes in cultivated land based on satellite remote sensing data. Background Technology

[0002] Farmland is a core resource for ensuring food security, ecological security, and social stability. The scarcity and non-renewability of farmland resources dictate that its protection must adhere to the bottom line of "no reduction in quantity, no decline in quality, and no ecological damage." However, with the acceleration of urbanization and industrialization, the problem of farmland being used for non-agricultural and non-grain purposes is becoming increasingly serious, specifically manifested in three aspects: a decrease in farmland quantity, degradation of farmland quality, and ecological damage. Statistics show that the majority of the reasons for the decrease in farmland quantity are illegal occupations, such as building factories, digging ponds for fish farming, and mining. The majority of the reasons for the degradation of farmland quality are also due to soil pollution and organic matter loss caused by illegal occupation. When farmland is encroached upon, the original agricultural ecosystem (such as wetlands and vegetation buffer zones) is destroyed, which exacerbates soil erosion, causes a decline in biodiversity, and ultimately leads to ecological damage.

[0003] Currently, monitoring of illegal occupation of arable land mainly relies on manual inspections or computer vision image recognition technologies, but these have significant drawbacks, such as poor adaptability to complex scenarios and low detection accuracy, thus failing to meet the needs of precise supervision in complex scenarios. Summary of the Invention

[0004] To address the aforementioned problems in the existing technology, this invention provides a method and device for detecting changes in cultivated land based on satellite remote sensing data.

[0005] The technical problem to be solved by this invention is achieved through the following technical solution: This invention provides a method for detecting farmland changes based on satellite remote sensing data, comprising: Acquire the current satellite remote sensing image of the target cultivated land area and the corresponding solar term and weather information of the current satellite remote sensing image; Obtain satellite remote sensing images of the target cultivated land area that match the solar term and weather information from the reference image library, and use them as the current reference images; The current satellite remote sensing image is divided into M first sub-images, and the current reference image is divided into M second sub-images using the same division method as the current satellite remote sensing image. Each sub-image contains at least one cultivated land area. The z-th first sub-image and the z-th second sub-image form a pair of sub-images. The value of z is from 1 to M. z is a positive integer, and M is a positive integer greater than or equal to 2. Calculate the contour similarity between each pair of sub-images, and when the contour similarity of a pair of sub-images is less than a first threshold, use the first sub-image and the second sub-image in the pair of sub-images as the first image to be detected and the second image to be detected, respectively. The trained first object detection network is used to perform object detection on the first image to be detected and the second image to be detected, respectively. The category similarity between the first image to be detected and the second image to be detected is determined based on the object detection results. The category similarity is determined based on the number of objects with different categories and the number of objects with the same category. When the categories of objects with different categories are all of the first preset category, the number of objects with different categories is 0. If the category similarity is less than the second threshold, it is determined whether the target detection result of the first image to be detected contains an object of the second preset category. If so, it indicates that there is illegal occupation of farmland in the farmland area of ​​the first image to be detected.

[0006] The present invention also provides a farmland change detection device based on satellite remote sensing data, including a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other through the communication bus; The memory is used to store computer programs; When the processor executes the program stored in the memory, it implements the steps of the above-mentioned method for detecting changes in cultivated land based on satellite remote sensing data.

[0007] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention detects changes in cultivated land by introducing reference images matched with solar terms and weather information, reducing the impact of environmental factors on the detection results. It is applicable to various complex weather conditions, thereby improving the accuracy of detection and adaptability to complex scenarios. Furthermore, by dividing the image and precisely comparing the contour features of objects in each pair of sub-images, this invention enhances the ability to identify subtle changes in cultivated land areas, effectively improving the sensitivity and reliability of detection. In addition, based on contour feature detection, this invention further determines whether cultivated land areas have changed by calculating category similarity values, effectively reducing false alarms caused by farming activities, lowering the false detection rate, and further improving the accuracy of detection.

[0008] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating a method for detecting changes in cultivated land based on satellite remote sensing data, provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the feature fusion module provided in an embodiment of the present invention. Detailed Implementation

[0010] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0011] Figure 1 This is a flowchart illustrating a method for detecting farmland changes based on satellite remote sensing data, as provided in an embodiment of the present invention. Figure 1 As shown, the method includes: S101. Obtain the current satellite remote sensing image of the target cultivated land area and the corresponding solar term and weather information.

[0012] Here, the solar term and weather information indicates the current solar term, as well as current visibility, light intensity, temperature, humidity, wind speed, etc. Specifically, by establishing a real-time communication connection with the meteorological station in the target cultivated land area, the solar term and weather information of the target cultivated land area collected at the time of satellite remote sensing image acquisition can be automatically obtained.

[0013] In some embodiments, the current satellite remote sensing image is a satellite remote sensing image that has undergone image quality assessment and whose image quality meets a preset quality threshold.

[0014] S102. Obtain satellite remote sensing images of the target cultivated land area that match the solar term and weather information from the reference image library, and use them as the current reference images.

[0015] Here, the solar term and weather information at the time of acquiring the current satellite remote sensing image of the target cultivated land area are crucial for matching the corresponding reference image. For example, if the current satellite remote sensing image was acquired during the Spring Equinox in spring, and the corresponding visibility is X, light intensity is Y, and wind speed is Z, then a historical satellite remote sensing image of the target cultivated land area that also belongs to the Spring Equinox and whose corresponding visibility, light intensity, and wind speed are closest to visibility X, light intensity Y, and wind speed Z is selected from the reference image library as the reference image for the current satellite remote sensing image (i.e., the current reference image). This approach not only considers the seasonal impact but also the influence of real-time weather conditions on image quality, thereby improving the accuracy of cultivated land change detection.

[0016] S103. Divide the current satellite remote sensing image into M first sub-images, and divide the current reference image into M second sub-images using the same division method as the current satellite remote sensing image. Each sub-image contains at least one cultivated land area. The z-th first sub-image and the z-th second sub-image form a pair of sub-images. The value of z is from 1 to M. z is a positive integer and M is a positive integer greater than or equal to 2.

[0017] For example, when the target cultivated land area includes M cultivated land areas, the current satellite remote sensing image and the current reference image can be divided into M sub-images using the same division method, so that each sub-image contains a cultivated land area.

[0018] S104. Calculate the contour similarity between each pair of sub-images, and when the contour similarity between a pair of sub-images is less than the first threshold, take the first sub-image and the second sub-image in the pair of sub-images as the first image to be detected and the second image to be detected, respectively.

[0019] S105. The trained first object detection network is used to perform object detection on the first image to be detected and the second image to be detected respectively. The category similarity between the first image to be detected and the second image to be detected is determined based on the object detection results. The category similarity is determined based on the number of objects with different categories and the number of objects with the same category. When the categories of objects with different categories are all of the first preset category, the number of objects with different categories is 0.

[0020] Here, the objects in the first preset category refer to objects commonly found in farmland, such as people, animals, agricultural machinery, and field ridges.

[0021] S106. If the category similarity is less than the second threshold, determine whether the target detection result of the first image to be detected contains an object of the second preset category. If so, it indicates that there is illegal occupation of farmland in the farmland area of ​​the first image to be detected.

[0022] Here, the objects in the second preset category refer to illegal objects that damage farmland, such as buildings, ponds, mines, and roads.

[0023] Here, when the contour similarity between a first image to be detected and a second image to be detected is less than a first threshold, and the category similarity between the first image to be detected and the second image to be detected is also less than a second threshold, it indicates that a new object that should not appear in the cultivated land area of ​​the first image to be detected may have appeared. Therefore, it is necessary to further determine whether the target detection result of the first image to be detected contains an object of the second preset category. If yes, it indicates that there is illegal occupation of cultivated land in the cultivated land area of ​​the first image to be detected, that is, the cultivated land area of ​​the first image to be detected has changed. If no, it indicates that there is no illegal occupation of cultivated land in the cultivated land area of ​​the first image to be detected, that is, the cultivated land area of ​​the first image to be detected has not changed.

[0024] In some embodiments, the "calculating the contour similarity between each pair of sub-images" in S104 above can be achieved through steps S1041 to S1046: S1041. For each pair of sub-images, extract the contour point set of each object in the first and second sub-images of the pair of sub-images respectively.

[0025] Here, for each pair of sub-images, existing contour extraction algorithms can be used to extract the contour point set of each object in the first sub-image of the pair, and the contour point set of each object in the second sub-image of the pair.

[0026] S1042. For each object, calculate the geometric features of the object based on the object's contour point set, whereby the geometric features include: area, perimeter, and aspect ratio.

[0027] Specifically, for an object, before calculating its area, perimeter, and aspect ratio, median filtering of its contour point set is required to remove redundant points and improve the accuracy of the calculated area, perimeter, and aspect ratio. When calculating the area of ​​an object, first check if the denoised contour point set is closed. If not, connect the last point to the first point to close the point set. If closed, the area can be calculated using the Shoelace Theorem. Similarly, when calculating the perimeter of an object, first check if the denoised contour point set is closed. If closed, the perimeter is obtained by calculating the sum of the Euclidean distances between adjacent points. If not closed, treat the last point and the first point as adjacent points, and then calculate the sum of the Euclidean distances between these adjacent points to obtain the perimeter. Specifically, for an object, when calculating the aspect ratio of the object, the minimum bounding rectangle of the object is determined based on the denoised contour point set of the object, and the aspect ratio of the object is obtained based on the minimum bounding rectangle.

[0028] S1043. After normalizing the contour point set of each object, input it into the pre-trained texture feature extraction network to obtain the texture features of each object.

[0029] For example, the texture feature extraction network could be ResNet18.

[0030] S1044. The geometric features and texture features of all objects in the first sub-image are fused to obtain the geometric features and texture features of the first sub-image, and the geometric features and texture features of the first sub-image are used as the contour features of the first sub-image.

[0031] Specifically, the geometric features of each object in the first sub-image are a row matrix consisting of area, perimeter, and aspect ratio. This row matrix can first be standardized (Z-score) to eliminate unit differences in area, perimeter, and aspect ratio. Then, the standardized row matrices of all objects in the first sub-image are combined to form a geometric feature matrix. This geometric feature matrix has the number of rows equal to the number of objects in the first sub-image, and three columns. Thus, this geometric feature matrix represents the geometric features of the first sub-image. Similarly, the same processing is performed on the geometric features of all objects in the second sub-image to obtain the geometric features of the second sub-image.

[0032] Specifically, existing feature fusion methods can be used to fuse the texture features of all objects in the first sub-image to obtain the texture features of the first sub-image. Similarly, the texture features of all objects in the second sub-image can be fused to obtain the texture features of the second sub-image.

[0033] S1045. The geometric features and texture features of all objects in the second sub-image are fused to obtain the geometric features and texture features of the second sub-image, and the geometric features and texture features of the second sub-image are used as the contour features of the second sub-image.

[0034] S1046. Calculate the similarity between the contour features of the first sub-image and the contour features of the second sub-image to obtain the contour similarity between the first sub-image and the second sub-image.

[0035] Specifically, before calculating the contour similarity between the first and second sub-images, the geometric features of the first sub-image are sorted in ascending order according to the area column (i.e., the row with the largest area is used as the last row of the entire matrix, and the row with the smallest area is used as the first row of the entire matrix) to obtain the updated geometric features of the first sub-image. This reduces the error caused by differences in object order. Similarly, the geometric features of the second sub-image are processed in the same way to obtain the updated geometric features of the second sub-image. Then, the similarity between the texture features of the first and second sub-images, as well as the similarity between the updated geometric features of the first and second sub-images, are calculated separately. Finally, the similarity between the texture features and the similarity between the updated geometric features are weighted and fused to obtain the contour similarity between the first and second sub-images.

[0036] Specifically, the geometric features of the first sub-image are determined by... The area, perimeter, and aspect ratio of an object constitute its composition. The matrix, the geometric features of the second sub-image are formed by... The area, perimeter, and aspect ratio of an object constitute its composition. Based on this matrix, the expression for the contour similarity between the first and second sub-images is as follows: ; ; ; in, For texture weights, The texture features of the first sub-image, For the texture features of the second sub-image, The cosine similarity function is used. The updated geometric features representing the first sub-image. The updated geometric features of the second sub-image represent the updated geometric features of the first sub-image, which are based on... The size of the area in each column of the matrix, for The matrix obtained by sorting the rows of the matrix in ascending order, and the updated geometric features of the second sub-image are according to... The size of the area in each column of the matrix, for The matrix obtained by sorting the rows of the matrix in ascending order. The value is , represent The OK, The value is , represent The OK, Indicates from From the starting point to the line and from From the starting point to the The optimal cumulative distance between rows, where, The starting point is The first element, The starting point is The first element, , and Similarly, Represents the L2 norm. This indicates taking the minimum value. This indicates taking the maximum value, when hour, ,when In and hour, yes .here, It is a normalization factor that can eliminate the influence of sequence length differences on similarity calculation. It can reflect the overall difference between two sequences. The contour similarity score ranges from [0,1], which can intuitively reflect the overall similarity between two geometric feature matrices.

[0037] This invention effectively solves the problem of inconsistent row counts in geometric feature matrices through standardization, sorting, and dynamic time warping, while preserving object-level feature information. Furthermore, it can maintain computational efficiency while retaining details, making it suitable for scenarios where the number of objects changes dynamically in real-world situations.

[0038] In some embodiments, the "determining the category similarity between the first image to be detected and the second image to be detected based on the target detection result" in S105 above can be achieved through steps S1051 to S1056: S1051. Determine the number of objects of the same category in the target detection results of the first image to be detected and the target detection results of the second image to be detected, and obtain the first number.

[0039] S1052. Determine whether there are objects of different categories in the target detection results of the first image to be detected and the target detection results of the second image to be detected.

[0040] S1053. If there are objects of different categories, determine whether the categories of the objects of different categories are all people, animals, agricultural machinery, or field ridges.

[0041] S1054. If among objects of different categories, some objects belong to the categories of people, agricultural machinery, or field ridges, then the number of objects among the objects of different categories that are not of the categories of people, animals, agricultural machinery, or field ridges shall be taken as the second quantity.

[0042] S1055. If the categories of objects with different categories are all people, animals, agricultural machinery, or field ridges, then set the number of objects with different categories to 0, and obtain a second quantity with a value of 0.

[0043] S1056. If the categories of objects with different categories are not human, animal, agricultural machinery, or field ridge, then the number of objects with different categories shall be taken as the second quantity.

[0044] In this invention, since people, animals, agricultural machinery, and field ridges are all common objects in cultivated land, when detecting illegal occupation of cultivated land in the first image to be detected, the number of objects of different categories that are useful for detecting illegal occupation of cultivated land can be obtained by removing the number of objects of the different categories among the target detection results of the first image to be detected and the target detection results of the second image to be detected, which are people, animals, agricultural machinery, or field ridges.

[0045] S1057. Based on the first quantity and the second quantity, determine the category similarity between the first image to be detected and the second image to be detected.

[0046] Specifically, the sum of the first quantity and the second quantity is calculated, and then the ratio of the first quantity to the sum of the first quantity and the second quantity is used as the category similarity between the first image to be detected and the second image to be detected.

[0047] In some embodiments, the above method further includes S107~S109: S107. When the contour similarity of M pairs of sub-images consisting of M first sub-images and M second sub-images is greater than or equal to the first threshold, or when the category similarity of each pair of sub-images with a contour similarity less than the first threshold is greater than or equal to the second threshold, it indicates that all cultivated land areas in the current satellite remote sensing image have not changed.

[0048] S108. Mark all cultivated land areas in the current satellite remote sensing image, and use the trained second target detection network to detect objects of the second preset category in the marked current satellite remote sensing image. The objects of the second preset category include: buildings, ponds, mines and roads. The trained first target detection network and the trained second target detection network have the same structure, but the training sets of the trained first target detection network and the trained second target detection network are different.

[0049] Specifically, the training samples used to train both the first and second object detection networks are images collected from farmland areas. However, when training the first object detection network, the labels on the sample images include people, agricultural machinery, field ridges, buildings, ponds, mines, and roads, so that the first object detection network can identify people, agricultural machinery, field ridges, buildings, ponds, mines, and roads. When training the second object detection network, the labels on the sample images only include buildings, ponds, mines, and roads in farmland areas, so that the second object detection network can only identify buildings, ponds, mines, and roads in farmland.

[0050] S109. When no object of the second preset category is detected in all the marked cultivated land areas, the current satellite remote sensing image is used as the satellite remote sensing image of the target cultivated land area that matches the solar term and weather information and then placed into the reference image library.

[0051] Here, by performing target detection again on the current satellite remote sensing image, it can be ensured that the cultivated land area in the current satellite remote sensing image does not contain illegal objects such as buildings, ponds, mines, and roads, thus guaranteeing the accuracy of the reference image. This invention ensures the timeliness and accuracy of the reference data by automatically updating the reference image library, reducing the need for manual intervention.

[0052] In some embodiments, both the first and second object detection networks include: a multi-layer feature extraction module, a feature fusion module, and a detection head module. The multi-layer feature extraction module extracts features at different levels from the input satellite remote sensing image. The feature fusion module performs adaptive feature fusion based on adaptive weights on the features at different levels to obtain fused features. The detection head module generates object detection results based on the fused features.

[0053] Specifically, a multi-layer feature extraction module can be used to extract features at three different levels, where the scale of the first-level features is larger than that of the second-level features, and the scale of the second-level features is larger than that of the third-level features. For example, the multi-layer feature extraction module could be a ResNet-50. For instance, Figure 2 This is a schematic diagram of the feature fusion module provided in an embodiment of the present invention. Figure 2 As shown, the feature fusion module includes: three channel alignment networks, two weight generation networks, and two adjustment and fusion modules. The first channel alignment network performs channel alignment on the first-level features to obtain the first channel aligned feature; the second channel alignment network performs channel alignment on the second-level features to obtain the second channel aligned feature; and the third channel alignment network performs channel alignment on the third-level features to obtain the third channel aligned feature. The first weight generation network generates adaptive weights for the second channel aligned feature, and the second weight generation network generates adaptive weights for the third channel aligned feature. The first adjustment and fusion module performs weighted adjustment on the third channel aligned feature based on its adaptive weights before fusing it with the second channel aligned feature to obtain the initial fused feature. The second adjustment and fusion module performs weighted adjustment on the initial fused feature based on its adaptive weights before fusing it with the first channel aligned feature to obtain the final fused feature. It should be noted that weighted adjustment of the channel aligned feature using adaptive weights means multiplying the adaptive weights by the channel aligned feature. The fusion described in this invention refers to pixel-by-pixel fusion.

[0054] This invention adaptively generates weights, which can more finely adjust the weight relationship between features at different levels. In addition, the adaptive weights generated by this invention enable deep interaction and information fusion between multi-level features, which can ensure that the detailed information of low-level features and the semantic information of high-level features are effectively combined. This allows for more effective feature fusion, fully leverages the complementary advantages of multi-level features, and helps improve the detection performance of the network in complex scenes.

[0055] Specifically, the weight generation network consists of the following layers in sequence: a first convolutional layer, a batch normalization layer, a ReLU activation function, a global average pooling layer, a second convolutional layer, and a sigmoid normalization function. The kernel size of the first convolutional layer is 3×3, and the kernel size of the second convolutional layer is 1×1. The inputs and outputs of the first convolutional layer, batch normalization layer, ReLU activation function, global average pooling layer, second convolutional layer, and sigmoid normalization function are connected sequentially. Furthermore, the input of the first convolutional layer is the input of the weight generation network, and the output of the sigmoid normalization function is the output of the weight generation network.

[0056] The present invention also provides a farmland change detection device based on satellite remote sensing data, including a processor, a communication interface, a memory, and a communication bus. The processor, communication interface, and memory communicate with each other through the communication bus. The memory is used to store computer programs. When the processor executes the program stored in the memory, it implements the steps of the above-mentioned farmland change detection method based on satellite remote sensing data.

[0057] It should be noted that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, features defined as "first" or "second" may explicitly or implicitly include one or more features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0058] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Furthermore, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.

[0059] In this specification, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude multiple instances. While different embodiments may describe certain measures, this does not mean that these measures cannot be combined to produce a good effect.

[0060] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A method for detecting changes in cultivated land based on satellite remote sensing data, characterized in that, include: Acquire the current satellite remote sensing image of the target cultivated land area and the corresponding solar term and weather information of the current satellite remote sensing image; Obtain satellite remote sensing images of the target cultivated land area that match the solar term and weather information from the reference image library, and use them as the current reference images; The current satellite remote sensing image is divided into M first sub-images, and the current reference image is divided into M second sub-images using the same division method as the current satellite remote sensing image. Each sub-image contains at least one cultivated land area. The z-th first sub-image and the z-th second sub-image form a pair of sub-images. The value of z is from 1 to M. z is a positive integer, and M is a positive integer greater than or equal to 2. Calculate the contour similarity between each pair of sub-images, and when the contour similarity of a pair of sub-images is less than a first threshold, use the first sub-image and the second sub-image in the pair of sub-images as the first image to be detected and the second image to be detected, respectively. The trained first object detection network is used to perform object detection on the first image to be detected and the second image to be detected, respectively. The category similarity between the first image to be detected and the second image to be detected is determined based on the object detection results. The category similarity is determined based on the number of objects with different categories and the number of objects with the same category. When the categories of objects with different categories are all of the first preset category, the number of objects with different categories is 0. If the category similarity is less than the second threshold, it is determined whether the target detection result of the first image to be detected contains an object of the second preset category. If so, it indicates that there is illegal occupation of farmland in the farmland area of ​​the first image to be detected.

2. The method according to claim 1, characterized in that, The calculation of contour similarity between each pair of sub-images includes: For each pair of sub-images, extract the set of contour points for each object in the first and second sub-images respectively; For each object, the geometric features of the object are calculated based on the object's contour point set, wherein the geometric features include: area, perimeter, and aspect ratio; The contour point set of the object is normalized and then input into a pre-trained texture feature extraction network to obtain the texture features of the object; The geometric and texture features of all objects in the first sub-image are fused to obtain the geometric features and texture features of the first sub-image, and the geometric and texture features of the first sub-image are used as the contour features of the first sub-image. The geometric and texture features of all objects in the second sub-image are fused to obtain the geometric features and texture features of the second sub-image, and the geometric and texture features of the second sub-image are used as the contour features of the second sub-image. Calculate the similarity between the contour features of the first sub-image and the contour features of the second sub-image to obtain the contour similarity between the first sub-image and the second sub-image.

3. The method according to claim 2, characterized in that, The geometric features of the first sub-image are... The area, perimeter, and aspect ratio of an object constitute its composition. The matrix, the geometric features of the second sub-image are formed by... The area, perimeter, and aspect ratio of an object constitute its composition. The matrix, the formula for calculating the contour similarity is as follows: ； ； ； in, For texture weights, The texture features of the first sub-image. The texture features of the second sub-image. The cosine similarity function is used. The updated geometric features representing the first sub-image. The updated geometric features of the second sub-image represent the updated geometric features of the first sub-image, which are based on the aforementioned... The size of the area in each column of the matrix, for the... The matrix obtained by sorting the rows of the matrix in ascending order, and the updated geometric features of the second sub-image are according to the above. The size of the area in each column of the matrix, for the... The matrix obtained by sorting the rows of the matrix in ascending order. The value is , represent The OK, The value is , represent The OK, Indicates from From the starting point to the line and from From the starting point to the The optimal cumulative distance between rows The starting point is The first element, The starting point is The first element, Indicates from From the starting point to the line and from From the starting point to the The optimal cumulative distance between rows Indicates from From the starting point to the line and from From the starting point to the The optimal cumulative distance between rows Indicates from From the starting point to the line and from From the starting point to the The optimal cumulative distance between rows Represents the L2 norm. This indicates taking the minimum value. This indicates taking the maximum value, when hour, ,when In and hour, yes .

4. The method according to claim 1, characterized in that, The objects in the first preset category include: people, animals, agricultural machinery, and field ridges; determining the category similarity between the first image to be detected and the second image to be detected based on the target detection results includes: The number of objects of the same category in the target detection results of the first image to be detected and the target detection results of the second image to be detected is determined to obtain a first number; Determine whether there are objects of different categories in the target detection results of the first image to be detected and the target detection results of the second image to be detected; If there are objects of different categories, determine whether all objects of different categories are people, animals, agricultural machinery, or field ridges; If among objects of different categories, some objects belong to the categories of people, animals, agricultural machinery, or field ridges, then the number of objects among the objects of different categories that are not in the categories of people, animals, agricultural machinery, or field ridges shall be used as the second quantity. If the categories of objects with different categories are all people, agricultural machinery, or field ridges, then the number of objects with different categories is set to 0, resulting in the second number with a value of 0. If the categories of objects that are different are not human, animal, agricultural machinery, or field ridge, then the number of objects that are different in category shall be taken as the second number; Based on the first quantity and the second quantity, the category similarity between the first image to be detected and the second image to be detected is determined.

5. The method according to claim 4, characterized in that, Determining the category similarity between the first image to be detected and the second image to be detected based on the first quantity and the second quantity includes: Calculate the sum of the first quantity and the second quantity; The ratio of the first quantity to the sum of the first quantity and the second quantity is used as the category similarity between the first image to be detected and the second image to be detected.

6. The method according to claim 1, characterized in that, The method further includes: When the contour similarity of the M pairs of sub-images consisting of the M first sub-images and the M second sub-images is greater than or equal to the first threshold, or when the category similarity of each pair of sub-images with a contour similarity less than the first threshold is greater than or equal to the second threshold, it indicates that all cultivated land areas in the current satellite remote sensing image have not changed. All cultivated land areas are marked in the current satellite remote sensing image. A trained second target detection network is used to detect objects of the second preset category in the marked current satellite remote sensing image. The objects of the second preset category include: buildings, ponds, mines and roads. The trained first target detection network and the trained second target detection network have the same structure, but the training sets of the trained first target detection network and the trained second target detection network are different. When no object of the second preset category is detected in any of the marked cultivated land areas, the current satellite remote sensing image is placed into the reference image library as the satellite remote sensing image of the target cultivated land area that matches the solar term and weather information.

7. The method according to claim 6, characterized in that, Both the first target detection network and the second target detection network include: a multi-layer feature extraction module, a feature fusion module, and a detection head module; wherein, the multi-layer feature extraction module is used to extract features of different levels from the input satellite remote sensing image; the feature fusion module is used to perform adaptive feature fusion based on adaptive weights on the features of different levels to obtain fused features; and the detection head module is used to generate target detection results based on the fused features.

8. The method according to claim 7, characterized in that, The multi-layer feature extraction module is used to extract features at three different levels; the feature fusion module includes: three channel alignment networks, two weight generation networks, and two adjustment and fusion modules; The first channel alignment network is used to perform channel alignment on the first level of features to obtain the first channel aligned features; the second channel alignment network is used to perform channel alignment on the second level of features to obtain the second channel aligned features; and the third channel alignment network is used to perform channel alignment on the third level of features to obtain the third channel aligned features. The first weight generation network is used to generate adaptive weights for the second channel alignment feature; The second weight generation network is used to generate adaptive weights for the third channel alignment feature; The first adjustment and fusion module is used to adjust the third channel alignment feature according to the adaptive weight of the third channel alignment feature and then fuse it with the second channel alignment feature to obtain the initial fused feature; The second adjustment and fusion module is used to adjust the initial fusion feature according to the adaptive weight of the second channel alignment feature and then fuse it with the first channel alignment feature to obtain the final fusion feature.

9. The method according to claim 8, characterized in that, The weight generation network comprises, in sequence: a first convolutional layer, a batch normalization layer, a ReLU activation function, a global average pooling layer, a second convolutional layer, and a Sigmoid normalization function. The kernel size of the first convolutional layer is 3×3, and the kernel size of the second convolutional layer is 1×1.

10. A farmland change detection device based on satellite remote sensing data, comprising a processor, a communication interface, a memory, and a communication bus, characterized in that, The processor, the communication interface, and the memory communicate with each other via the communication bus; The memory is used to store computer programs; When the processor executes a program stored in the memory, it implements the steps of the method described in any one of claims 1-9.

Images