Deep Learning-Based Land Expansion Detection Method, Device, Equipment, and Medium

Abstract

Embodiments of the present disclosure disclose a land expansion detection method, apparatus, device, and medium based on deep learning. A specific implementation of the method includes: obtaining each remote sensing image that meets preset conditions as initial data; performing preliminary preprocessing on the initial data to obtain preliminarily processed data; performing first preprocessing on the first-phase preliminary remote sensing image group to obtain first-phase segmented remote sensing images; based on a semantic segmentation model, performing semantic segmentation on the first-phase segmented remote sensing images to obtain a land mask; performing second preprocessing on the preliminarily processed data to obtain a remote sensing image set; based on the land mask, performing mask coverage on each detection remote sensing image in the remote sensing image set to obtain a covered remote sensing image set; based on a change detection model and the covered remote sensing image set, generating a visualization picture representing land expansion changes. This implementation realizes high-precision and high-robustness land expansion detection by using deep learning methods.

Description

Technical Field

[0001] Embodiments of the present disclosure relate to the field of computer technology, and a method, apparatus, device, and medium for land expansion detection based on deep learning. Background Art

[0002] The method for land expansion detection based on deep learning is an important technical means in the fields of remote sensing image analysis and geographic information processing. It realizes the accurate identification and visualization of land expansion areas through the processing and analysis of remote sensing images of different time phases. Currently, land expansion detection mainly relies on manual visual interpretation and traditional image processing algorithms. The above-mentioned manual visual interpretation requires professionals to observe and judge remote sensing images based on experience and knowledge. Traditional image processing algorithms such as threshold segmentation and vegetation index calculation can automatically identify some land expansion areas.

[0003] However, when using the above methods, there are often the following technical problems: Manual visual interpretation is not only time-consuming and laborious, but also highly subjective. The results of different interpreters may vary, making it difficult to ensure consistency. Traditional image processing algorithms are prone to being interfered by factors such as noise and ground object similarity when facing complex land expansion scenarios, resulting in poor adaptability to complex surface cover types and being unable to accurately identify the subtle expansion or changes of the land, thus leading to misjudgment and missed judgment.

[0004] The above information disclosed in this background art section is only used to enhance the understanding of the background of the inventive concept, and thus, it may include information that does not form the prior art known to those of ordinary skill in the art in this country. Summary of the Invention

[0005] This summary of the disclosure is used to introduce concepts in a brief form, and these concepts will be described in detail in the following detailed implementation section. This summary of the disclosure is not intended to identify the key features or essential features of the claimed technical solution, nor is it intended to be used to limit the scope of the claimed technical solution.

[0006] Some embodiments of the present disclosure propose a method, apparatus, electronic device, and computer-readable medium for land expansion detection based on deep learning to solve one or more of the technical problems mentioned in the above background art section.

[0007] In a first aspect, some embodiments of the present disclosure propose a method for detecting land expansion based on deep learning, the method comprising: obtaining each remote sensing image that meets a preset condition as initial data, wherein the initial data includes a first-phase remote sensing image group and a second-phase remote sensing image group; performing preliminary preprocessing on the initial data to obtain preliminary processed data, wherein the preliminary processed data includes a first-phase preliminary remote sensing image group and a second-phase preliminary remote sensing image group; performing first preprocessing on the first-phase preliminary remote sensing image group to obtain a first-phase segmented remote sensing image; performing semantic segmentation on the first-phase segmented remote sensing image based on a semantic segmentation model to obtain a land mask; performing second preprocessing on the preliminary processed data to obtain a remote sensing image set, wherein the remote sensing image set includes a first-phase detected remote sensing image and a second-phase detected remote sensing image; performing mask coverage on each detected remote sensing image in the remote sensing image set based on the land mask to obtain a covered remote sensing image set, wherein the covered remote sensing image set includes a first-phase covered remote sensing image and a second-phase covered remote sensing image; generating a visualization picture representing land expansion changes based on a change detection model and the covered remote sensing image set.

[0008] In a second aspect, some embodiments of the present disclosure propose a device for detecting land expansion based on deep learning, comprising: an obtaining unit configured to obtain each remote sensing image that meets a preset condition as initial data, wherein the initial data includes a first-phase remote sensing image group and a second-phase remote sensing image group; a preliminary processing unit configured to perform preliminary preprocessing on the initial data to obtain preliminary processed data, wherein the preliminary processed data includes a first-phase preliminary remote sensing image group and a second-phase preliminary remote sensing image group; a first processing unit configured to perform first preprocessing on the first-phase preliminary remote sensing image group to obtain a first-phase segmented remote sensing image; a segmentation unit configured to perform semantic segmentation on the first-phase segmented remote sensing image based on a semantic segmentation model to obtain a land mask; a second processing unit configured to perform second preprocessing on the preliminary processed data to obtain a remote sensing image set, wherein the remote sensing image set includes a first-phase detected remote sensing image and a second-phase detected remote sensing image; a covering unit configured to perform mask coverage on each detected remote sensing image in the remote sensing image set based on the land mask to obtain a covered remote sensing image set, wherein the covered remote sensing image set includes a first-phase covered remote sensing image and a second-phase covered remote sensing image; a generating unit configured to generate a visualization picture representing land expansion changes based on a change detection model and the covered remote sensing image set.

[0009] In a third aspect, some embodiments of the present disclosure provide an electronic device, including: one or more processors; a storage device storing one or more programs thereon, and when the one or more programs are executed by the one or more processors, the one or more processors implement the method described in any implementation manner of the first aspect above.

[0010] In a fourth aspect, some embodiments of the present disclosure provide a computer-readable medium storing a computer program thereon, wherein when the program is executed by a processor, the method described in any implementation manner of the first aspect above is implemented.

[0011] In a fifth aspect, some embodiments of the present disclosure provide a computer program product including a computer program, and when the computer program is executed by a processor, the method described in any implementation manner of the first aspect above is implemented.

[0012] The above-mentioned various embodiments of the present disclosure have the following beneficial effects: Through the deep learning-based land expansion detection method of some embodiments of the present disclosure, efficient, accurate, and highly robust land expansion detection can be achieved, significantly improving the reliability and practicality of the detection results. Specifically, traditional land expansion detection methods, such as relying on manual visual interpretation, may cause problems such as low efficiency, strong subjectivity, and poor result consistency when dealing with large-scale remote sensing images; if relying on traditional image processing algorithms, there may be problems such as poor adaptability to complex surface cover types and inability to accurately identify subtle expansion changes. Based on this, for the deep learning-based land expansion detection method of some embodiments of the present disclosure, first, each remote sensing image that meets the preset conditions is obtained as initial data, where the above initial data includes a first-phase remote sensing image group and a second-phase remote sensing image group. Thus, a targeted data basis is provided for subsequent land expansion detection, ensuring the timeliness and relevance of the data. Then, the above initial data is preliminarily preprocessed to obtain preliminarily processed data, where the above preliminarily processed data includes a first-phase preliminarily remote sensing image group and a second-phase preliminarily remote sensing image group. Thus, through operations such as cloud masking construction and image size adjustment, noise and interference factors in the data are effectively removed, making the image data more in line with the input requirements of the subsequent deep learning model and improving the data quality. Next, the first-phase preliminarily remote sensing image group is subjected to a first preprocessing to obtain a first-phase segmented remote sensing image. Thus, through operations such as band synthesis and image synthesis, the multi-band remote sensing image information is organically integrated, providing richer and more comprehensive feature information for the generation of the land mask. Then, based on the semantic segmentation model, the first-phase segmented remote sensing image is semantically segmented to obtain a land mask. Thus, by using the deep learning model to automatically learn the complex features in the image, the accurate recognition and segmentation of the land area are realized. Compared with traditional methods, the generation accuracy and efficiency of the land mask are greatly improved. Then, the above preliminarily processed data is subjected to a second preprocessing to obtain a remote sensing image set, where the above remote sensing image set includes a first-phase detected remote sensing image and a second-phase detected remote sensing image. Thus, through operations such as targeted band synthesis and resolution adjustment on the image again, it is ensured that the image data used for change detection meets the requirements of the model in terms of features and resolution, providing data guarantee for accurately detecting land expansion changes subsequently. Then, based on the above land mask, each detected remote sensing image in the remote sensing image set is covered with the mask to obtain a covered remote sensing image set, where the above covered remote sensing image set includes a first-phase covered remote sensing image and a second-phase covered remote sensing image. Thus, through the effective coverage of the land mask, the change detection is focused on the land area, reducing the interference of non-land areas and improving the pertinence and accuracy of the change detection. Finally, based on the change detection model and the above covered remote sensing image set, a visualization picture representing the land expansion change is generated.Therefore, by leveraging the powerful feature extraction and analysis capabilities of the deep learning model and combining the information in the post-coverage remote sensing image set, the land expansion change areas are accurately detected and presented in the form of intuitive visualization pictures, facilitating users to quickly and accurately understand and analyze the scope, degree, and pattern of land expansion, and providing strong support for relevant decision-making. In addition, since scientific and rigorous strategies are adopted in all aspects such as data screening, preprocessing, and model training, this method can effectively handle remote sensing image data obtained from different regions, different time phases, and different sensors, enhancing the generality and adaptability of the model. At the same time, through continuous training and optimization of the deep learning model, the performance and robustness of the model can be continuously improved, enabling it to maintain a stable detection effect in complex and changing actual application scenarios. Thus, by combining deep learning technology with remote sensing image analysis, efficient and accurate land expansion detection is achieved, providing a more reliable and practical decision-making basis for fields such as global land resource management, urban planning, and ecological environment protection. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] In combination with the accompanying drawings and with reference to the following specific embodiments, the above and other features, advantages, and aspects of the various embodiments of the present disclosure will become more apparent. Throughout the accompanying drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic and the elements and elements are not necessarily drawn to scale.

[0014] Figure 1 is a flowchart of some embodiments of the deep learning-based land expansion detection method according to the present disclosure;

[0015] Figure 2 is a schematic structural diagram of some embodiments of the deep learning-based land expansion detection device according to the present disclosure;

[0016] Figure 3 is a schematic structural diagram of an electronic device suitable for implementing some embodiments of the present disclosure. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0017] Embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although some embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. On the contrary, these embodiments are provided to more thoroughly and completely understand the present disclosure. It should be understood that the drawings and embodiments of the present disclosure are only for exemplary purposes and are not used to limit the protection scope of the present disclosure.

[0018] It should also be noted that, for the sake of convenience of description, only parts related to the relevant invention are shown in the drawings. Without conflict, the embodiments in the present disclosure and the features in the embodiments can be combined with each other.

[0019] It should be noted that concepts such as "first" and "second" mentioned in this disclosure are only used to distinguish different devices, modules or units, and are not used to limit the order of functions performed by these devices, modules or units or their interdependent relationships.

[0020] It should be noted that the modifications of "one" and "multiple" mentioned in this disclosure are illustrative rather than restrictive. Those skilled in the art should understand that, unless otherwise clearly specified in the context, it should be understood as "one or more".

[0021] The names of the messages or information exchanged between multiple devices in the embodiments of this disclosure are only for illustrative purposes and are not used to limit the scope of these messages or information.

[0022] The following will detail this disclosure with reference to the accompanying drawings and in conjunction with embodiments.

[0023] Figure 1 Flow 100 of some embodiments of a deep learning-based land expansion detection method according to this disclosure is shown. The deep learning-based land expansion detection method includes the following steps:

[0024] Step 101, obtaining each remote sensing image that meets a preset condition as initial data.

[0025] In some embodiments, the execution subject (such as a computing device) of the deep learning-based land expansion detection method may obtain each remote sensing image that meets preset conditions as initial data. Among them, the above-mentioned each remote sensing image is usually multi-spectral image data of the Earth's surface obtained by sensors on satellites or aircraft. The above-mentioned preset conditions include that the spatial range of the remote sensing image is greater than or equal to a preset target area, the cloud coverage rate of the remote sensing image is less than or equal to a preset cloud coverage rate, and the shooting time of the remote sensing image is within a preset time range. The above-mentioned preset target area may be an area marked or selected by the identified researcher or user on a platform (such as GEE). The above-mentioned preset cloud coverage rate is usually the proportion of the area of the remote sensing image blocked by clouds in the entire image area (such as 5%). The above-mentioned preset time range is usually a time interval set by the identified researcher or user. The above-mentioned time interval is usually one year or ten years. The above-mentioned preset time range usually includes a first-phase time range and a second-phase time range. The above-mentioned first-phase time range is usually an earlier period (such as 1984). The above-mentioned second-phase time range is usually a later period (such as 2023). The above-mentioned initial data may include a first-phase remote sensing image group and a second-phase remote sensing image group. The above-mentioned first-phase remote sensing image group may include each remote sensing image that meets the preset conditions and meets the above-mentioned first-phase time range. The above-mentioned second-phase remote sensing image group may include each remote sensing image that meets the preset conditions and meets the above-mentioned second-phase time range.

[0026] In some optional implementation manners of some embodiments, the execution subject may obtain each remote sensing image that meets the preset conditions as initial data through the following steps:

[0027] Step 1, obtain remote sensing image data as an initial image set. Among them, the above-mentioned remote sensing image data includes each remote sensing image and the attribute data corresponding to each remote sensing image. The above-mentioned attribute data corresponding to each remote sensing image may include the shooting time, spatial range (represented by geographical coordinates), cloud coverage rate, band information, and resolution of each remote sensing image. The above-mentioned initial image set includes each remote sensing image obtained by the platform. In practice, the execution subject may obtain remote sensing image data collected from satellites such as Landsat-5 and Sentinel-2 through the GEE platform. Then, each remote sensing image in the above-mentioned remote sensing image data may be integrated into an initial image set.

[0028] Step 2: According to the pre-set target area, screen each remote sensing image in the above initial image set to obtain a target area image set. Among them, the above target area image set may include each remote sensing image in the initial image set that is greater than or equal to the pre-set target area. In practice, the above execution entity can identify the geographic coordinate information of the pre-set target area. Then, according to the attribute data corresponding to each remote sensing image in the above initial image set, screen out the remote sensing images in the above initial image set whose spatial range is greater than or equal to the pre-set target area, and obtain each remote sensing image whose spatial range is greater than or equal to the pre-set target area. Finally, integrate each remote sensing image whose spatial range is greater than or equal to the pre-set target area into a target area image set.

[0029] Step 3: According to the pre-set cloud coverage rate, screen each remote sensing image in the above target area image set to obtain a preliminary screening image set. Among them, the above preliminary screening image set may include each remote sensing image in the above target area image set that is less than or equal to the pre-set cloud coverage rate. In practice, the above execution entity can screen out the remote sensing images in the above target area image set whose cloud coverage rate is less than or equal to the pre-set cloud coverage rate according to the attribute data corresponding to each remote sensing image in the above target area image set, and obtain each remote sensing image whose cloud coverage rate is less than or equal to the pre-set cloud coverage rate. Finally, integrate each remote sensing image whose cloud coverage rate is less than or equal to the pre-set cloud coverage rate into a preliminary screening image set.

[0030] Step 4: According to the pre-set time range, screen each remote sensing image in the above preliminary screening image set to obtain a target time image set. Among them, the above target time image set may include each remote sensing image in the above preliminary screening image set that meets the pre-set time range. In practice, the above execution entity can screen out the remote sensing images in the above preliminary screening image set whose shooting time meets the pre-set time range according to the attribute data corresponding to each remote sensing image in the above preliminary screening image set, and obtain each remote sensing image whose shooting time meets the pre-set time range. Finally, integrate each remote sensing image whose shooting time meets the pre-set time range into a target time image set. For example, currently the pre-set time range includes 1984 and 2023, and the above execution entity screens each remote sensing image in the above preliminary screening image set according to the attribute information of each remote sensing image in the above preliminary screening image set, and obtains each remote sensing image whose shooting time is between 1984 and 2023. Then integrate each remote sensing image whose shooting time is between 1984 and 2023 into a target time image set.

[0031] Step 5: Based on the attribute data corresponding to each remote sensing image in the above target time image set, classify each remote sensing image in the above target time image set to obtain a first-phase remote sensing image group and a second-phase remote sensing image group. The above first-phase remote sensing image group can be each remote sensing image in the above target time image set that satisfies the above first-phase time range. The above second-phase remote sensing image group can be each remote sensing image in the above target time image set that satisfies the above second-phase time range. In practice, the above execution entity can classify each remote sensing image in the above target time image set according to the first-phase time range and the second-phase time range in the pre-set time range to obtain a first-phase remote sensing image group and a second-phase remote sensing image group, that is, the initial data. For example, the currently pre-set time range includes 1984 and 2023. Among them, 1984 is the first-phase time range; 2023 is the second-phase time range. Integrate each remote sensing image in the above target time image set with a shooting time in 1984 into a first-phase remote sensing image group. Similarly, a second-phase remote sensing image group can be obtained. Finally, integrate the above first-phase remote sensing image group and the above second-phase remote sensing image group into the initial data.

[0032] Step 102: Perform preliminary preprocessing on the initial data to obtain preliminary processed data.

[0033] In some embodiments, the above execution entity can perform preliminary preprocessing on the above initial data to obtain preliminary processed data. Among them, the above preliminary preprocessing includes cloud cover removal and size adjustment. The above cloud cover removal can include constructing a cloud mask and applying the cloud mask to the remote sensing image. The above size adjustment can be to adjust the size of the remote sensing image to a preset size (such as 512×512). The above preliminary processed data includes a first-phase preliminary remote sensing image group and a second-phase preliminary remote sensing image group. The above first-phase preliminary remote sensing image group can be obtained by performing preliminary preprocessing on the above first-phase remote sensing image group. The above second-phase preliminary remote sensing image group can be obtained by performing preliminary preprocessing on the above second-phase remote sensing image group.

[0034] In some optional implementation manners of some embodiments, the above execution entity can perform preliminary preprocessing on the above initial data through the following steps to obtain preliminary processed data:

[0035] Step 1: For each remote sensing image in the above initial data, perform the following steps:

[0036] Sub-step 1: Obtain the attribute data of the above remote sensing image. Among them, the attribute data of the above remote sensing image includes the shooting time of the above remote sensing image, the spatial range of the above remote sensing image, the cloud coverage rate of the above remote sensing image, the band information of the above remote sensing image, and the resolution of the above remote sensing image. In practice, the above execution entity can retrieve the attribute data of the above remote sensing image from the attribute data corresponding to each of the above remote sensing images.

[0037] Sub-step 2: Construct a cloud mask based on the attribute data of the above remote sensing image. Among them, the above cloud mask can be a binary map marking cloud regions and non-cloud regions. In practice, the above execution entity can construct cloud mask conditions (usually the band information to be removed) based on the band information of the above remote sensing image. Then, combine the above cloud mask conditions to obtain a binary map, that is, the cloud mask. For example, for a second-phase remote sensing image obtained from the Sentinel-2 satellite with a shooting time of 2023, the above execution entity can extract the SCL band of the above remote sensing image from the band information of the above remote sensing image. Then, the classification values in the SCL band of the above remote sensing image can be extracted. Next, logical operations (such as neq representing not equal) can be used to exclude specific classification values, and the steps include: excluding pixels of cloud shadow (SCL value is 3), medium probability cloud (SCL value is 8), high probability cloud (SCL value is 9), and thin cirrus cloud (SCL value is 10). Finally, after logical operations, each pixel is marked as 1 (representing non-cloud region) or 0 (representing cloud region), and a binary map with "1" representing non-cloud region and "0" representing cloud region can be obtained, that is, the cloud mask. Another example, for a first-phase remote sensing image obtained from the Landsat-5 satellite with a shooting time of 1984, the above execution entity can extract the QA_PIXEL band of the above remote sensing image from the band information of the above remote sensing image. Then, the classification values in the QA_PIXEL band of the above remote sensing image can be extracted. Next, logical operations (such as bitwiseAnd and eq) can be used to exclude specific classification values, and the steps include: excluding pixels of cloud (Bit3 of QA_PIXEL), cloud shadow (Bit4 of QA_PIXEL), and dilated cloud (Bit2 of QA_PIXEL). Similarly, a binary map with "1" representing non-cloud region and "0" representing cloud region can also be obtained according to logical operations, that is, the cloud mask.

[0038] Sub-step 3: Generate a preliminarily processed remote sensing image based on the above cloud mask. Among them, the above preliminarily processed remote sensing image can be the above remote sensing image to which the above cloud mask is applied. In practice, the above execution entity can apply the above cloud mask to each band of the above remote sensing image. Then, pixels including specific classification values can be removed through masking operations, and only pixels in non-cloud regions are retained to obtain a preliminarily processed remote sensing image.

[0039] Sub-step 4: Adjust the above-mentioned preliminarily processed remote sensing image to a pre-set size to obtain a preliminary remote sensing image. Among them, the above-mentioned pre-set size can be the pre-set size magnitude. The above-mentioned preliminary remote sensing image can be obtained by adjusting the above-mentioned preliminarily processed remote sensing image to the pre-set size. In practice, the above-mentioned execution entity can adjust the above-mentioned preliminarily processed remote sensing image to a size of 512×512 to obtain a preliminary remote sensing image.

[0040] Step 2: Integrate each preliminary remote sensing image into preliminarily processed data. In practice, the above-mentioned execution entity can integrate each preliminary remote sensing image into a first-phase preliminary remote sensing image group and a second-phase preliminary remote sensing image group according to the shooting time of each preliminary remote sensing image and the pre-set time range (for example, the first-phase time range is 1984, and the second-phase time range is 2023). Then, the above-mentioned first-phase preliminary remote sensing image group and the above-mentioned second-phase preliminary remote sensing image group can be integrated into preliminarily processed data.

[0041] Step 103: Perform a first preprocessing on the first-phase preliminary remote sensing image group to obtain a first-phase segmented remote sensing image.

[0042] In some embodiments, the above-mentioned execution entity can perform a first preprocessing on the above-mentioned first-phase preliminary remote sensing image group to obtain a first-phase segmented remote sensing image. Among them, the above-mentioned first preprocessing includes band synthesis and image synthesis. The above-mentioned band synthesis can be a process of extracting data of multiple different bands from a single multi-spectral image and combining these bands into a new composite image. The above-mentioned band synthesis includes true color synthesis and false color synthesis. The above-mentioned true color synthesis usually uses the red, green, and blue bands corresponding to the red, green, and blue primary colors respectively. The above-mentioned false color synthesis usually uses non-visible light bands (such as near-infrared, mid-infrared) for synthesis. The above-mentioned image synthesis usually refers to combining multiple images into a new image through a certain method. The above-mentioned first-phase segmented remote sensing image can be a remote sensing image obtained by performing a first preprocessing on the first-phase preliminary remote sensing image group.

[0043] In some optional implementation manners of some embodiments, the above-mentioned execution entity can perform a first preprocessing on the above-mentioned first-phase preliminary remote sensing image group through the following steps to obtain a first-phase segmented remote sensing image:

[0044] Step 1: According to the band conditions of the segmentation model, perform band synthesis on each first-phase preliminary remote sensing image in the above first-phase preliminary remote sensing image group to obtain each first-phase remote sensing image that meets the band conditions of the segmentation model. Among them, the above band conditions of the segmentation model can be that the remote sensing image processed by the semantic segmentation model is an image obtained through false color synthesis (such as fusing the information of the near-infrared, red, and green bands). In practice, the above execution entity can use the method of additive color fusion to perform band synthesis on each first-phase preliminary remote sensing image to obtain each first-phase remote sensing image that meets the band conditions of the segmentation model. For example, for each first-phase preliminary remote sensing image in the above first-phase preliminary remote sensing images, the above execution entity can extract the near-infrared band, red band, and green band of any first-phase preliminary remote sensing image. Then, assign the near-infrared band to the red channel, the red band to the green channel, and the green band to the blue channel. Finally, the band information of these three channels can be superimposed to obtain the first-phase remote sensing image corresponding to the first-phase preliminary remote sensing image that meets the band conditions of the segmentation model. Similarly, each first-phase remote sensing image that meets the band conditions of the segmentation model can be obtained.

[0045] Step 2: Perform image synthesis on each of the above first-phase remote sensing images that meet the band conditions of the segmentation model to obtain a first-phase segmented remote sensing image. Among them, the above first-phase segmented remote sensing image can be a remote sensing image obtained by performing image synthesis on each of the above first-phase remote sensing images that meet the band conditions of the segmentation model. In practice, the above execution entity can adopt the method of median composite image to synthesize the above first-phase remote sensing images that meet the band conditions of the segmentation model into a first-phase segmented remote sensing image. For example, the above execution entity can first perform alignment processing on each of the above first-phase remote sensing images that meet the band conditions of the segmentation model (such as by constructing a coordinate system) to ensure that the same pixel positions correspond to the same scene content. Then, the position of each pixel in any first-phase remote sensing image that meets the band conditions of the segmentation model can be extracted. Next, for any position, the pixel values at this position in each of the above first-phase remote sensing images that meet the band conditions of the segmentation model can be extracted. Then, the extracted pixel values at this position are sorted, and the median value is taken as the median pixel at this position. Then, the median pixels at each position can be obtained. Finally, the median pixels at each position can be combined to obtain the first-phase segmented remote sensing image.

[0046] Step 104: Based on the semantic segmentation model, perform semantic segmentation on the first-phase segmented remote sensing image to obtain a land mask.

[0047] In some embodiments, the above-mentioned execution entity may perform semantic segmentation on the above-mentioned first-phase segmented remote sensing image based on a semantic segmentation model to obtain a land mask. Among them, the above-mentioned semantic segmentation model may be a deep learning model (such as FCN or Res-Unet) that takes an image synthesized with false colors as input data and the land mask of the image as output. The above-mentioned semantic segmentation refers to segmenting regions in an image. The above-mentioned land mask is usually a binary image with the same size as the input first-phase segmented remote sensing image and representing the land area, where the pixel values are 0 or 1 (0 represents a non-land area, and 1 represents a land area).

[0048] In some optional implementation manners of some embodiments, the above-mentioned execution entity may perform semantic segmentation on the above-mentioned first-phase segmented remote sensing image based on a semantic segmentation model through the following steps to obtain a land mask:

[0049] First step, according to the band conditions of the segmentation model, perform band synthesis on each second-phase preliminary remote sensing image in the above-mentioned second-phase preliminary remote sensing image group to obtain each second-phase remote sensing image that meets the band conditions of the segmentation model. In practice, the specific implementation manner may refer to the steps of obtaining each first-phase remote sensing image that meets the band conditions of the segmentation model, which will not be elaborated here.

[0050] Second step, according to the resolution condition of the segmentation model, adjust the resolution of each second-phase remote sensing image that meets the band condition of the segmentation model to obtain each second-phase remote sensing image that meets the resolution condition of the segmentation model. Among them, the above-mentioned resolution condition of the segmentation model can be that the remote sensing image processed by the semantic segmentation model is an image with a relatively low resolution. In practice, the above-mentioned execution entity can adjust the resolution of each second-phase remote sensing image that meets the band condition of the segmentation model to be the same as the resolution of each first-phase remote sensing image that meets the band condition of the segmentation model, so as to obtain each second-phase remote sensing image that meets the resolution condition of the segmentation model. For example, currently, each second-phase remote sensing image that meets the band condition of the segmentation model is a 10m (resolution) remote sensing image that meets the second-phase time range (such as 2023) and comes from the Sentinel-2 satellite. Currently, each first-phase remote sensing image that meets the band condition of the segmentation model is a 30m (resolution) remote sensing image that meets the first-phase time range (such as 1984) and comes from the Landsat-5 satellite. Each second-phase remote sensing image that meets the band condition of the segmentation model can be input into an image processing software (such as Photoshop). Then, by using the image processing software to resample each second-phase remote sensing image that meets the band condition of the segmentation model, each second-phase remote sensing image that meets the band condition of the segmentation model with a resolution of 30m is obtained. Finally, the size of each second-phase remote sensing image that meets the band condition of the segmentation model with a resolution of 30m can be adjusted by using the PIL library of Python, so that the size of each second-phase remote sensing image that meets the band condition of the segmentation model with a resolution of 30m is still the pre-set size (such as 512×512), and each second-phase remote sensing image that meets the resolution condition of the segmentation model is obtained.

[0051] Step 3: Integrate the above-mentioned second-phase remote sensing images that meet the resolution conditions of the segmentation model and the above-mentioned first-phase remote sensing images that meet the band conditions of the segmentation model into the dataset of the semantic segmentation model. For example, the above-mentioned execution entity can first use geographic information system software (such as ArcGIS) to uniformly convert the remote sensing images that meet the resolution conditions of the segmentation model into the GCS_WGS_1984 coordinate system to ensure image alignment. Then, a Shapefile file of the polygon type can be created in the geographic information system software, and a "class_id" field can be added to distinguish land (1) and sea (0). Next, the "Create Fishnet" tool in ArcGIS can be used to create sampling points at a certain interval within the study area to obtain sampling points. Then, the sampling points can be overlaid with the Shapefile file so that each sampling point obtains the corresponding "class_id" value, and the overlaid feature data can be converted into a raster format to generate semantic segmentation labels. Then, remote sensing images with a size of 512×512 that contain both land and sea are filtered out. Finally, the remote sensing images corresponding to the second phase (such as 2023) are divided into a training set and a validation set at a ratio of 80% and 20%, the remote sensing images corresponding to the first phase (such as 1984) are all used for the test set, and the images of each remote sensing image are organized into the dataset folder to obtain the dataset of the semantic segmentation model.

[0052] Step 4: Train the constructed semantic segmentation model with the above-mentioned dataset of the semantic segmentation model to obtain the semantic segmentation model. Among them, the above-mentioned constructed semantic segmentation model can be an untrained semantic segmentation model. In practice, the above-mentioned execution entity can use a loss function (such as cross entropy) to detect the ability of the constructed semantic segmentation model to process the dataset of the semantic segmentation model for optimization and adjustment, and add residual connections and skip connections to the constructed semantic segmentation model. Finally, the semantic segmentation model can be obtained.

[0053] Step 5: Input the above-mentioned first-phase segmented remote sensing image into the segmentation model input layer of the above-mentioned semantic segmentation model to obtain the input image. Among them, the above-mentioned semantic segmentation model can be a Res-Unet model. The above-mentioned semantic segmentation model can include the above-mentioned segmentation model input layer, segmentation model encoder, segmentation model bottleneck layer, segmentation model decoder, and segmentation model output layer. In practice, the above-mentioned execution entity can set the size of the image to 512×512 and the number of channels (convolution kernel channels) to 3 in the above-mentioned segmentation model input layer to correspond to the near-infrared, red, and green bands in the remote sensing image. The above-mentioned first-phase segmented remote sensing image can be received as the input image through the above-mentioned segmentation model input layer.

[0054] Step 6: Input the above input image into the downsampling part of the above segmentation model encoder to downsample the above input image and obtain a downsampled feature map. Among them, the above segmentation model encoder includes the downsampling part of the segmentation model encoder and the feature extraction layer of the segmentation model encoder. The downsampling part of the above segmentation model encoder includes a preset number of segmentation model downsampling layers. In practice, the above execution entity can input the above input image into the downsampling part of the above segmentation model encoder to perform a max pooling operation and obtain a downsampled feature map. Among them, the downsampling part of the above segmentation model encoder includes three segmentation model downsampling layers. The pooling kernel of each segmentation model downsampling layer is set to 2×2. An activation function (such as ReLU) is introduced in each segmentation model downsampling layer. The output size of the first segmentation model downsampling layer is 256×256×64 (64 represents the output channels of this layer); the output size of the second segmentation model downsampling layer is 128×128×128 (the last 128 represents the output channels of this layer); the output size of the third segmentation model downsampling layer is 64×64×256 (256 represents the output channels of this layer). For example, the above execution entity can first input the above input data into the first segmentation model downsampling layer to obtain the output of the first downsampling layer of the segmentation model. Then input the output of the first downsampling layer of the above segmentation model into the second segmentation model downsampling layer to obtain the output of the second downsampling layer of the segmentation model. Finally, input the output of the second downsampling layer of the segmentation model into the third segmentation model downsampling layer to obtain a downsampled feature map with a size of 64×64×256. In this way, the max pooling operation is completed in each segmentation model downsampling layer.

[0055] Step 7: Input the above downsampled feature map into the feature extraction layer of the above segmentation model encoder to extract features from the above downsampled feature map and obtain an encoder feature map. Among them, the above feature extraction layer includes convolutional processing and max pooling operation. In practice, the above execution entity can perform convolutional processing on the above downsampled feature map by using a convolutional kernel, and then obtain an encoder feature map through a pooling operation. Among them, the above convolutional processing includes two convolutional operations. Each convolutional operation can be a 3×3 convolutional kernel operation and uses ReLU as the activation function. The output channels of each convolutional operation are 512. The pooling kernel in the max pooling operation is set to 2×2. For example, the above execution entity can perform two convolutional processes on the above downsampled feature map to obtain the output of the convolutional process. Then, perform a max pooling operation on the output of the above convolutional process to obtain an encoder feature map with a size of 32×32×512.

[0056] In the eighth step, input the above encoder feature map into the bottleneck layer of the above segmentation model to obtain a bottleneck feature map. Among them, the bottleneck layer of the above segmentation model may include feature extraction and compression operations. In practice, the above execution entity can increase the number of channels of the above encoder feature map by increasing the number of channels, so as to achieve feature extraction and compression and obtain a bottleneck feature map. For example, the above execution entity can increase the number of channels of the above encoder feature map from 512 to 1024 to obtain a bottleneck feature map with a size of 32×32×1024.

[0057] In the ninth step, input the above bottleneck feature map into the decoder of the above segmentation model to upsample the above bottleneck feature map to obtain a decoder feature map. Among them, the decoder of the above segmentation model includes a preset number of segmentation model upsampling layers. In practice, the above execution entity can input the above bottleneck feature map into the above preset number of segmentation model upsampling layers for upsampling operations to obtain a decoder feature map. Among them, the decoder of the above segmentation model may include four segmentation model upsampling layers. The scaling factor of each segmentation model upsampling layer is set to 2×2. Each segmentation model upsampling layer introduces an activation function (such as ReLU) and an adaptive dilated convolution (such as the dilation rate of the convolution kernel is set to 1, 3, 6, 12) to capture multi-scale information and improve the accuracy of small-range processing. The output size of the first segmentation model upsampling layer is 64×64×512 (512 represents the output channels of this layer); the output size of the second segmentation model upsampling layer is 128×128×256 (256 represents the output channels of this layer); the output size of the third segmentation model upsampling layer is 256×256×128 (128 represents the output channels of this layer). The output size of the fourth segmentation model upsampling layer is 512×512×64 (64 represents the output channels of this layer). For example, the above execution entity can input the above bottleneck feature map with a size of 32×32×1024 into the first segmentation model upsampling layer to obtain the output of the first upsampling layer of the segmentation model. Then input the output of the first upsampling layer of the segmentation model into the second segmentation model upsampling layer to obtain the output of the second upsampling layer of the segmentation model. Then, input the output of the second upsampling layer of the segmentation model into the third segmentation model upsampling layer to obtain the output of the third upsampling layer of the segmentation model; finally, input the output of the third upsampling layer of the segmentation model into the fourth segmentation model upsampling layer to obtain a decoder feature map with a size of 512×512×64. In this way, upsampling operations are completed in each segmentation model upsampling layer.

[0058] Step 10: Input the above decoder feature map into the output layer of the above segmentation model to obtain a land mask. Among them, the output size of the output layer of the above segmentation model is 512×512×1 (1 represents the output channels of this layer). The above land mask can be a single-channel binary image. In practice, the above execution entity can perform a convolution operation on the above decoder feature map using a 1×1 convolutional kernel and use Sigmoid as the activation function. In addition, a boundary-aware loss function can be combined. By jointly optimizing pixel classification (such as traditional cross-entropy loss to ensure the overall classification correctness of land and non-land areas) and boundary regression tasks (that is, based on the distance between the predicted boundary and the true boundary, such as Hausdorff distance or L1 loss, to penalize edge blurring or misalignment), the edge accuracy of the land mask can be improved to obtain a land mask with an output size of 512×512×1.

[0059] The above first to tenth steps are an inventive point of the embodiments of the present disclosure, which solve the problems of "when the existing land expansion detection method processes complex remote sensing images, there are problems of insufficient segmentation accuracy, incomplete capture of subtle ground feature characteristics, and inability to efficiently generate land masks". The difficulties in meeting the high-precision land expansion detection requirements in the prior art are as follows: the existing methods have deficiencies in dynamic feature extraction, multi-scale feature fusion, and end-to-end segmentation result generation, and are difficult to effectively adapt to image data obtained from different regions and different sensors. At the same time, there are defects in model robustness and computational efficiency, resulting in low detection accuracy and inability to fully meet the requirements of high-precision land expansion detection. If the above factors are solved, it is possible to improve the land expansion detection accuracy, optimize the model robustness, and achieve high-precision semantic segmentation. To achieve this effect, the present disclosure adopts a land expansion detection method based on deep learning, synthesizing the above-mentioned second-phase preliminary remote sensing images into second-phase remote sensing images that meet the band conditions of the segmentation model; adjusting the resolution of the second-phase remote sensing images that meet the band conditions of the segmentation model to obtain second-phase remote sensing images that meet the resolution conditions of the segmentation model; integrating the second-phase remote sensing images that meet the resolution conditions of the segmentation model and the above-mentioned first-phase remote sensing images that meet the band conditions of the segmentation model into a dataset of the semantic segmentation model; training the constructed semantic segmentation model using the dataset of the semantic segmentation model to obtain a semantic segmentation model; inputting the first-phase segmented remote sensing image into the input layer of the semantic segmentation model to obtain a standardized input image; performing downsampling on the input image using the downsampling part of the segmentation model encoder to obtain a downsampled feature map; performing feature extraction on the downsampled feature map using the feature extraction layer of the segmentation model encoder to obtain an encoder feature map; performing feature extraction and compression on the encoder feature map using the bottleneck layer of the segmentation model to obtain a bottleneck feature map; performing upsampling on the bottleneck feature map using the decoder of the segmentation model to obtain a decoder feature map; and finally obtaining a single-channel land mask through the output layer of the segmentation model. Thus, an end-to-end optimization from the original remote sensing image to the land mask can be realized. Thereby, it is possible to effectively extract land features, flexibly adapt to different image data, optimize all aspects of land expansion detection, improve the detection accuracy and model robustness, and thus achieve high-precision semantic segmentation.

[0060] Step 105: Perform second preprocessing on the preliminarily processed data to obtain a set of remote sensing images.

[0061] In some embodiments, the above-mentioned execution entity may perform a second preprocessing on the above-mentioned preliminary processed data to obtain a remote sensing image set. Wherein, the above-mentioned second preprocessing includes band synthesis, image synthesis, resolution adjustment and integration processing. The above-mentioned remote sensing image set includes a first-phase detection remote sensing image and a second-phase detection remote sensing image. The above-mentioned remote sensing image set may be a remote sensing image set obtained by performing a second preprocessing on the above-mentioned preliminary processed data.

[0062] In some optional implementation manners of some embodiments, the above-mentioned execution entity may perform a second preprocessing on the above-mentioned preliminary processed data through the following steps to obtain a remote sensing image set:

[0063] Step 1, according to the band conditions of the detection model, perform band synthesis on each preliminary remote sensing image in the above-mentioned preliminary processed data to obtain each preliminary remote sensing image that meets the band conditions of the detection model. Wherein, the above-mentioned band conditions of the detection model may be that the remote sensing image processed by the change detection model is an image obtained by true color synthesis (such as fusing the information of three bands of red light, green light, and blue light). Each of the above-mentioned preliminary remote sensing images that meet the band conditions of the detection model includes each first-phase remote sensing image that meets the band conditions of the detection model and each second-phase remote sensing image that meets the band conditions of the detection model. In practice, the above-mentioned execution entity may use the method of additive color fusion to perform band synthesis on each preliminary remote sensing image in the above-mentioned first-phase preliminary remote sensing image group and the above-mentioned second-phase preliminary remote sensing image group to obtain each first-phase remote sensing image that meets the band conditions of the detection model and each second-phase remote sensing image that meets the band conditions of the detection model, that is, each preliminary remote sensing image that meets the band conditions of the detection model. For example, for each preliminary remote sensing image in the above-mentioned preliminary remote sensing images, the above-mentioned execution entity may extract the red light band, green light band, and blue light band of any preliminary remote sensing image. Then, assign the red light band to the red channel, the green light band to the green channel, and the blue light band to the blue channel. Finally, by superimposing the band information of these three channels, the preliminary remote sensing image corresponding to the detection model band condition can be obtained. Similarly, each preliminary remote sensing image that meets the band conditions of the detection model can be obtained, and according to the attribute information corresponding to each remote sensing image, each first-phase remote sensing image that meets the band conditions of the detection model and each second-phase remote sensing image that meets the band conditions of the detection model can be further obtained.

[0064] Step 2: Synthesize the first-phase remote sensing images that meet the detection model's band conditions and the second-phase remote sensing images that meet the detection model's band conditions to obtain the first-phase target remote sensing image and the second-phase detection remote sensing image. Among them, the above-mentioned first-phase target remote sensing image can be a remote sensing image obtained by synthesizing the above-mentioned first-phase remote sensing images that meet the detection model's band conditions. The above-mentioned second-phase detection remote sensing image can be a remote sensing image obtained by synthesizing the above-mentioned second-phase remote sensing images that meet the detection model's band conditions. In practice, the specific implementation method can refer to the steps of obtaining the above-mentioned first-phase segmented remote sensing image, which will not be elaborated here.

[0065] Step 3: Adjust the resolution of the above-mentioned first-phase target remote sensing image according to the detection model's resolution condition to obtain the first-phase detection remote sensing image. Among them, the above-mentioned detection model's resolution condition can be that the remote sensing image processed by the change detection model is an image with a higher resolution. In practice, the above-mentioned execution entity can use GAN (Generative Adversarial Network) to adjust the resolution of the above-mentioned first-phase target remote sensing image to be the same as that of the second-phase detection remote sensing image to obtain the first-phase detection remote sensing image. For example, the current second-phase detection remote sensing image is a 10m (resolution) remote sensing image that meets the second-phase time range (such as 2023) and comes from the Sentinel-2 satellite, and the current first-phase target remote sensing image is a 30m (resolution) remote sensing image that meets the first-phase time range (such as 1984) and comes from the Landsat-5 satellite. The above-mentioned execution entity can use ESRGAN (Super-Resolution Reconstruction Model) to increase the resolution of the current first-phase target remote sensing image from 30m to 10m. The first-phase target remote sensing image after resolution adjustment is determined as the first-phase detection remote sensing image.

[0066] Step 4: Integrate the above-mentioned first-phase detection remote sensing image and the second-phase detection remote sensing image into a remote sensing image set. In practice, the above-mentioned execution entity can integrate the above-mentioned first-phase detection remote sensing image and the above-mentioned second-phase detection remote sensing image into a remote sensing image set.

[0067] Step 106: Based on the land mask, perform mask coverage on each detection remote sensing image in the remote sensing image set to obtain the covered remote sensing image set.

[0068] In some embodiments, the above-mentioned execution entity may perform masking coverage on each detected remote sensing image in the above-mentioned remote sensing image set based on the above-mentioned land mask to obtain a set of post-coverage remote sensing images. Among them, the above-mentioned set of post-coverage remote sensing images includes the first-phase post-coverage remote sensing images and the second-phase post-coverage remote sensing images. In practice, the above-mentioned execution entity may use the land area in the above-mentioned land mask as a color coverage layer, that is, fill the land area with a single color. Then, the color coverage layer can be superimposed on each detected remote sensing image according to the coordinates of each pixel to ensure correct superimposition, and the first-phase post-coverage remote sensing images and the second-phase post-coverage remote sensing images, that is, the set of post-coverage remote sensing images, can be obtained.

[0069] Step 107, generate a visualization picture representing land expansion changes based on the change detection model and the set of post-coverage remote sensing images.

[0070] In some embodiments, the above-mentioned execution entity may generate a visualization picture representing land expansion changes based on the change detection model and the above-mentioned set of post-coverage remote sensing images. Among them, the above-mentioned change detection model may be a deep learning model (such as CNN or Siamese) that takes two images synthesized in true color as input data and the change mask of these two images as output. The above-mentioned change mask may be a binary image used to represent the change information between two images. The pixel value therein is 1 (or 255) or 0 (1 indicates that the pixel at this position has changed between the two input images, and 0 indicates that the pixel at this position has not changed between the two input images).

[0071] In some optional implementation manners of some embodiments, the above-mentioned execution entity may generate a visualization picture representing land expansion changes based on the change detection model and the above-mentioned set of post-coverage remote sensing images through the following steps:

[0072] Step 1, according to the resolution conditions of the detection model, adjust the resolution of each first-phase remote sensing image that meets the band conditions of the detection model to obtain each first-phase remote sensing image that meets the resolution conditions of the detection model. Among them, the above-mentioned first-phase remote sensing images that meet the resolution conditions of the detection model are obtained by adjusting the resolution of each first-phase remote sensing image that meets the band conditions of the detection model. In practice, the specific implementation manner may refer to the steps for obtaining the above-mentioned first-phase detected remote sensing images and will not be elaborated here.

[0073] Step 2: Based on the above-mentioned land mask, perform mask coverage on each of the first-phase remote sensing images that meet the resolution conditions of the detection model and each of the second-phase remote sensing images that meet the band conditions of the detection model to obtain a dataset for the change detection model. Among them, the dataset of the change detection model includes each subunit. In practice, the above-mentioned execution entity can respectively overlay the above-mentioned land mask on each of the first-phase remote sensing images that meet the resolution conditions of the detection model and each of the second-phase remote sensing images that meet the band conditions of the detection model to obtain the first-phase remote sensing images that meet the resolution conditions of the detection model after overlay and the second-phase remote sensing images that meet the band conditions of the detection model after overlay. The specific implementation method can refer to the steps of obtaining the set of remote sensing images after coverage, which will not be elaborated here. Then, each of the first-phase remote sensing images that meet the resolution conditions of the detection model after overlay can be integrated into a first-phase detection data group, and each of the second-phase remote sensing images that meet the band conditions of the detection model after overlay can be integrated into a second-phase detection data group, and a first-phase detection data group and a second-phase detection data group can be obtained. Then, image registration technology can be used to ensure the spatial alignment of the two groups of data, and control points can be used for precise registration. Next, one overlaid remote sensing image can be randomly selected from each of the two detection data groups for change detection. Among them, a change detection algorithm (such as differential analysis, ratio analysis, or principal component analysis) can be selected, or a machine learning method (such as random forest or support vector machine), or even a deep learning model (such as CNN) can be used to perform change detection on the randomly selected two image data. Then, the change area can be manually or semi-automatically marked through a GIS tool, and training and test labels can be created. The randomly selected two labeled image data are used as a subunit of the dataset of the change detection model. Similarly, each subunit can be obtained. Integrate each subunit, and the above-mentioned subunits can be divided into a training set and a test set according to the ratio of 80% and 20% to obtain a dataset for the change detection model.

[0074] Step 3: Train the constructed change detection model through the dataset of the change detection model to obtain a change detection model. Among them, the constructed change detection model can be an untrained change detection model. In practice, the above-mentioned execution entity can use a loss function (such as cross-entropy) to detect the ability of the constructed change detection model to process the dataset of the change detection model for optimization and adjustment, and add residual connections and skip connections to the constructed change detection model. Finally, a change detection model can be obtained.

[0075] Step 4: Input the above-mentioned set of remote sensing images after coverage into the above-mentioned change detection model to obtain a change mask for the first-phase remote sensing images after coverage and the second-phase remote sensing images after coverage.

[0076] Step 5: Superimpose the above change mask on the above remotely sensed image after the first-phase coverage and the above remotely sensed image after the second-phase coverage to obtain a visualization picture representing the land expansion change. The above visualization picture can be the remotely sensed image after the first-phase coverage with the change mask superimposed or the remotely sensed image after the second-phase coverage with the change mask superimposed. In practice, the above execution entity can use the changed area in the above change mask as the change coverage layer, that is, fill the changed area with a single color (different from the color of the land mask). Then, the change coverage layer can be superimposed on the above remotely sensed image after the first-phase coverage and the above remotely sensed image after the second-phase coverage according to the coordinates of each pixel to ensure accurate superimposition, and a visualization picture that can represent the land expansion change between the two phases is obtained.

[0077] In some optional implementation manners of some embodiments, the above execution entity can input the above set of remotely sensed images after coverage into the above change detection model through the following steps to obtain a change mask for the remotely sensed image after the first-phase coverage and the remotely sensed image after the second-phase coverage:

[0078] First step, input the remotely sensed image after the first - phase coverage and the remotely sensed image after the second - phase coverage into the sub - network of the change detection model to obtain the first - phase output feature map and the second - phase output feature map. Among them, the above - mentioned change detection model can be a Siamese model. The above - mentioned change detection model can include a sub - network part, each feature fusion layer, and a detection model output layer. The above - mentioned sub - network part includes a first sub - network and a second sub - network. Each sub - network includes a detection model input layer, a convolutional layer, and each feature generation layer (for down - sampling), and the weights and the number of channels of each sub - network are the same (that is, the first sub - network and the second sub - network have separate input layers, convolutional layers, and each feature generation layer, but the input layers, convolutional layers, and each feature generation layer of the first sub - network and the second sub - network have the same weights and the number of channels). In practice, for the remotely sensed image after the first - phase coverage and the remotely sensed image after the second - phase coverage, the execution subject can input the remotely sensed image after the first - phase coverage into the first sub - network and the remotely sensed image after the second - phase coverage into the second sub - network, and the first - phase output feature map and the second - phase output feature map can be obtained. For example, for the first sub - network, the execution subject can input the remotely sensed image after the first - phase coverage into the input layer of the first sub - network. Among them, the detection model input layer can be set to have 3 channels and a size of 512×512, and a first - sub - network input image with an output size of 512×512×3 can be obtained. Then, the first - sub - network input image can be input into the convolutional layer of the first sub - network to obtain a convolutional output image with an output size of 512×512×64. Among them, the above - mentioned convolutional layer includes 64 3x3 convolutional kernels and uses same padding to keep the spatial dimension unchanged. Then, the convolutional output image can be input into each feature generation layer to obtain a first - phase output feature map with an output size of 64×64×512. Among them, the first sub - network includes three feature generation layers. Each feature generation layer includes a pooling kernel (2×2) with a stride of 2 and a convolutional kernel (3×3). A Residual Attention Block is embedded in the convolutional kernels of each sub - network to enhance feature discriminability through residual connections and channel attention mechanisms, and the ReLU activation function is introduced. Among them, the number of channels of the first feature generation layer is 128, and the output feature map size is 256×256×128 (128 represents the output channels of this layer); the number of channels of the second feature generation layer is 256, and the output feature map size is 128×128×256 (the last 256 represents the output channels of this layer); the number of channels of the third feature generation layer is 512, and the output feature map size is 64×64×512 (512 represents the output channels of this layer). Similarly, the second - phase output feature map can be obtained.

[0079] In the second step, input the above first-phase output feature map and the above second-phase output feature map into each of the above feature fusion layers to obtain a change feature map. Among them, the scaling factors of each of the above feature fusion layers are all set to 2×2. Each of the above feature fusion layers introduces an activation function (such as ReLU) and a Difference-Guided Feature Pyramid Network (Difference-Guided FPN) to dynamically fuse multi-scale features by calculating the pixel-level difference weights of the output feature maps of each sub-network feature generation layer, so as to achieve upsampling. In practice, the above execution entity can first fuse the first-phase output feature map and the above second-phase output feature map into a deep change feature map. Then, the deep change feature map can be upsampled through each feature fusion layer, and the features in the output feature map of the second feature generation layer, the output feature map of the first feature generation layer of the first sub-network and the second sub-network, and the convolutional output image are spliced and refined. In order to restore the deep feature map to a pre-set size and integrate multi-scale features to obtain a change feature map. Among them, the above change detection model can include 3 feature fusion layers. For example, the above execution entity can perform feature fusion on the above first-phase output feature map and the above second-phase output feature map in a differential fusion manner to obtain a deep change feature map with a size of 64×64×512. Then, input the deep change feature map into the first feature fusion layer, and introduce the output feature map of the second feature generation layer of the first sub-network and the second sub-network to perform channel splicing on the deep change feature map with the current size of 64×64×512 to retain local shape details, and obtain an output feature map of the first feature fusion layer with a size of 128×128×256 (256 represents the output channels of this layer). Next, input the output feature map of the first feature fusion layer into the second feature fusion layer, and introduce the output feature map of the first feature generation layer of the first sub-network and the second sub-network to perform texture contrast (such as road breaks) on the output feature map of the first feature fusion layer with the current size of 128×128×256 to enhance the texture, and obtain an output feature map of the second feature fusion layer with a size of 256×256×128 (128 represents the output channels of this layer). Finally, input the output feature map of the second feature fusion layer into the third feature fusion layer, and introduce the convolutional output images of the first sub-network and the second sub-network to perform accuracy restoration (pixel-level) on the output feature map of the second feature fusion layer with the current size of 256×256×128, and obtain a change feature map with a size of 512×512×64 (64 represents the output channels of this layer).

[0080] In the third step, input the above-mentioned change feature map into the output layer of the above-mentioned detection model to obtain a change mask. Among them, the above-mentioned change mask can be a single-channel binary image representing the land change between the remotely sensed image after the first-phase coverage and the remotely sensed image after the second-phase coverage. In practice, the above-mentioned execution entity can perform a convolution operation on the change feature map using a 1×1 convolution kernel and use Sigmoid as the activation function to obtain a change mask with an output size of 512×512×1.

[0081] The above steps 1 to 3 are an inventive point of the embodiments of the present disclosure, which solve the problem that "existing land expansion detection methods have incomplete capture of subtle ground object features and inability to efficiently generate change masks when dealing with complex remotely sensed images". This makes it difficult to meet the requirements of high-precision change detection in the prior art as follows: existing methods have deficiencies in dynamic feature extraction, multi-scale feature fusion, and change mask generation, and it is difficult to effectively adapt to image data obtained from different regions and different sensors. At the same time, there are defects in model robustness and computational efficiency, resulting in low detection accuracy and inability to fully meet the requirements of high-precision change detection. If the above factors are solved, the requirements of improving change detection accuracy, optimizing model robustness, and achieving high-precision change detection can be met. To achieve this effect, the present disclosure adopts a Siamese model change detection method based on deep learning, and realizes the end-to-end optimization from the original remotely sensed image to the change mask through the following steps: First, input the remotely sensed images after the first-phase and second-phase coverages into two sub-networks of the Siamese model to obtain the first-phase output feature map and the second-phase output feature map. The sub-network includes a detection model input layer, a convolution layer, and multiple feature generation layers, which are used for downsampling and extracting multi-scale features, and at the same time keep the weights of the two sub-networks the same to ensure the consistency of feature extraction. Then, input the first-phase output feature map and the second-phase output feature map into the feature fusion layer to obtain a change feature map. The feature fusion layer refines and integrates multi-scale features through upsampling and feature splicing, restores the deep feature map to a pre-set size to retain local shape details and enhance texture. Finally, input the upsampled change feature map into the output layer of the detection model to obtain a change mask. The change mask is a binary image representing the land change between the two phases, and is generated through a 1×1 convolution operation and a Sigmoid activation function, realizing the precise mapping from the feature map to the change mask. Thus, the present disclosure can effectively extract land features, flexibly adapt to different image data, optimize all links of change detection, improve detection accuracy and model robustness, and thus meet the requirements of high-precision change detection.

[0082] The above-mentioned various embodiments of the present disclosure have the following beneficial effects: Through the deep learning-based land expansion detection method of some embodiments of the present disclosure, efficient, accurate, and highly robust land expansion detection can be achieved, significantly improving the reliability and practicality of the detection results. Specifically, traditional land expansion detection methods, such as relying on manual visual interpretation, may cause problems such as low efficiency, strong subjectivity, and poor result consistency when dealing with large-scale remote sensing images; if relying on traditional image processing algorithms, there may be problems such as poor adaptability to complex surface cover types and inability to accurately identify subtle expansion changes. Based on this, for the deep learning-based land expansion detection method of some embodiments of the present disclosure, first, each remote sensing image that meets the preset conditions is obtained as initial data, where the above initial data includes a first-phase remote sensing image group and a second-phase remote sensing image group. Thus, a targeted data basis is provided for subsequent land expansion detection, ensuring the timeliness and relevance of the data. Then, the above initial data is preliminarily preprocessed to obtain preliminarily processed data, where the above preliminarily processed data includes a first-phase preliminary remote sensing image group and a second-phase preliminary remote sensing image group. Thus, through operations such as cloud masking construction and image size adjustment, the noise and interference factors in the data are effectively removed, making the image data more in line with the input requirements of the subsequent deep learning model and improving the data quality. Next, the first-phase preliminary remote sensing image group is subjected to a first preprocessing to obtain a first-phase segmented remote sensing image. Thus, through operations such as band synthesis and image synthesis, the multi-band remote sensing image information is organically integrated, providing richer and more comprehensive feature information for the generation of the land mask. Then, based on the semantic segmentation model, the first-phase segmented remote sensing image is semantically segmented to obtain a land mask. Thus, by using the deep learning model to automatically learn the complex features in the image, the accurate identification and segmentation of the land area are realized. Compared with traditional methods, the generation accuracy and efficiency of the land mask are greatly improved. Then, the above preliminarily processed data is subjected to a second preprocessing to obtain a remote sensing image set, where the above remote sensing image set includes a first-phase detection remote sensing image and a second-phase detection remote sensing image. Thus, the image is again subjected to targeted processing such as band synthesis and resolution adjustment, ensuring that the image data used for change detection meets the requirements of the model in terms of features and resolution, providing data guarantee for accurately detecting land expansion changes subsequently. Then, based on the above land mask, each detection remote sensing image in the remote sensing image set is covered with the mask to obtain a covered remote sensing image set, where the above covered remote sensing image set includes a first-phase covered remote sensing image and a second-phase covered remote sensing image. Thus, through the effective coverage of the land mask, the change detection is focused on the land area, reducing the interference of non-land areas and improving the pertinence and accuracy of the change detection. Finally, based on the change detection model and the above covered remote sensing image set, a visual picture representing the land expansion change is generated.Thus, by leveraging the powerful feature extraction and analysis capabilities of the deep learning model and combining the information in the post-coverage remote sensing image set, the land expansion change areas are accurately detected and presented in the form of intuitive visualization pictures, facilitating users to quickly and accurately understand and analyze the scope, degree, and pattern of land expansion, and providing strong support for relevant decision-making. In addition, since scientific and rigorous strategies are adopted in all aspects such as data screening, preprocessing, and model training in this method, it can effectively handle remote sensing image data obtained from different regions, different time phases, and different sensors, enhancing the generality and adaptability of the model. At the same time, through continuous training and optimization of the deep learning model, the performance and robustness of the model can be continuously improved, enabling it to maintain a stable detection effect in complex and changing practical application scenarios. Therefore, by combining deep learning technology with remote sensing image analysis, efficient and accurate land expansion detection is achieved, providing a more reliable and practical decision-making basis for fields such as global land resource management, urban planning, and ecological environment protection.

[0083] Further reference is made to Figure 2 , as an implementation of the methods shown in the above figures, the present disclosure provides some embodiments of a land expansion detection device based on deep learning. These device embodiments correspond to Figure 1 the method embodiments shown, and the device can be specifically applied to various electronic devices.

[0084] As Figure 2As shown in the figure, the deep learning-based land expansion detection device 200 of some embodiments includes: an acquisition unit 201, a preliminary processing unit 202, a first processing unit 203, a segmentation unit 204, a second processing unit 205, a coverage unit 206, and a generation unit 207. Among them, the acquisition unit 201 is configured to acquire each remote sensing image that meets the preset conditions as initial data, where the initial data includes a first-phase remote sensing image group and a second-phase remote sensing image group; the preliminary processing unit 202 is configured to perform preliminary preprocessing on the initial data to obtain preliminary processed data, where the preliminary processed data includes a first-phase preliminary remote sensing image group and a second-phase preliminary remote sensing image group; the first processing unit 203 is configured to perform first preprocessing on the first-phase preliminary remote sensing image group to obtain a first-phase segmented remote sensing image; the segmentation unit 204 is configured to perform semantic segmentation on the first-phase segmented remote sensing image based on a semantic segmentation model to obtain a land mask; the second processing unit 205 is configured to perform second preprocessing on the preliminary processed data to obtain a remote sensing image set, where the remote sensing image set includes a first-phase detected remote sensing image and a second-phase detected remote sensing image; the coverage unit 206 is configured to perform mask coverage on each detected remote sensing image in the remote sensing image set based on the land mask to obtain a covered remote sensing image set, where the covered remote sensing image set includes a first-phase covered remote sensing image and a second-phase covered remote sensing image; the generation unit 207 is configured to generate a visualization picture representing the land expansion change based on a change detection model and the covered remote sensing image set.

[0085] It can be understood that the units described in the device 200 correspond to the respective steps in the method described with reference to Figure 1 Therefore, the operations, features, and beneficial effects described above for the method also apply to the device 200 and the units included therein, and will not be repeated here.

[0086] The following refers to Figure 3 , which shows a schematic structural diagram of an electronic device 300 suitable for implementing some embodiments of the present disclosure. Figure 3 The electronic device shown is only an example and should not impose any limitations on the functions and usage scopes of the embodiments of the present disclosure.

[0087] As Figure 3As shown, the electronic device 300 may include a processing device 301 (such as a central processing unit, a graphics processing unit, etc.), which may perform various appropriate actions and processes according to a program stored in the read-only memory (ROM) 302 or a program loaded from the storage device 308 into the random access memory (RAM) 303. In the RAM 303, various programs and data required for the operation of the electronic device 300 are also stored. The processing device 301, the ROM 302, and the RAM 303 are connected to each other through a bus 304. The input / output (I / O) interface 305 is also connected to the bus 304.

[0088] Generally, the following devices may be connected to the I / O interface 305: an input device 306 including, for example, a touch screen, a touchpad, a keyboard, a mouse, a camera, a microphone, an accelerometer, a gyroscope, etc.; an output device 307 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; and a communication device 309. The communication device 309 may allow the electronic device 300 to communicate with other devices wirelessly or wiredly to exchange data. Although Figure 3 an electronic device 300 with various devices is shown, it should be understood that it is not required to implement or include all the shown devices. Instead, more or fewer devices may be implemented or included. Figure 3 Each block shown in may represent one device or, as required, multiple devices.

[0089] In particular, according to some embodiments of the present disclosure, the processes described above with reference to the flowcharts may be implemented as computer software programs. For example, some embodiments of the present disclosure include a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes program codes for performing the methods shown in the flowcharts. In such some embodiments, the computer program may be downloaded and installed from a network through the communication device 309, or installed from the storage device 308, or installed from the ROM 302. When the computer program is executed by the processing device 301, the above functions defined in the methods of some embodiments of the present disclosure are executed.

[0090] It should be noted that the computer-readable media described in some embodiments of the present disclosure may be a computer-readable signal medium, a computer-readable storage medium, or any combination of the two. A computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination of the above. More specific examples of the computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In some embodiments of the present disclosure, the computer-readable storage medium may be any tangible medium that contains or stores a program, which can be used by or in conjunction with an instruction execution system, apparatus, or device. In some embodiments of the present disclosure, the computer-readable signal medium may include a data signal propagated in a baseband or as part of a carrier wave, which carries computer-readable program code. Such a propagated data signal may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. The computer-readable signal medium may also be any computer-readable medium other than the computer-readable storage medium, which can send, propagate, or transmit a program for use by or in conjunction with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium may be transmitted using any appropriate medium, including but not limited to: wires, optical cables, RF (radio frequency), etc., or any suitable combination of the above.

[0091] In some embodiments, the client and the server can communicate using any currently known or future-developed network protocol such as HTTP (HyperText Transfer Protocol), and can be interconnected with digital data communication in any form or medium (e.g., a communication network). Examples of communication networks include local area networks ("LAN"), wide area networks ("WAN"), the Internet (e.g., the Internet), and end-to-end networks (e.g., ad hoc end-to-end networks), as well as any currently known or future-developed networks.

[0092] The above computer-readable medium may be included in the above electronic device; or it may exist independently without being assembled into the electronic device. The above computer-readable medium carries one or more programs. When the one or more programs are executed by the electronic device, the electronic device is caused to: obtain each remote sensing image that meets a preset condition as initial data, where the initial data includes a first-phase remote sensing image group and a second-phase remote sensing image group; perform preliminary preprocessing on the initial data to obtain preliminary processed data, where the preliminary processed data includes a first-phase preliminary remote sensing image group and a second-phase preliminary remote sensing image group; perform first preprocessing on the first-phase preliminary remote sensing image group to obtain a first-phase segmented remote sensing image; perform semantic segmentation on the first-phase segmented remote sensing image based on a semantic segmentation model to obtain a land mask; perform second preprocessing on the preliminary processed data to obtain a remote sensing image set, where the remote sensing image set includes a first-phase detected remote sensing image and a second-phase detected remote sensing image; perform mask coverage on each detected remote sensing image in the remote sensing image set based on the land mask to obtain a covered remote sensing image set, where the covered remote sensing image set includes a first-phase covered remote sensing image and a second-phase covered remote sensing image; generate a visualization picture representing land expansion change based on a change detection model and the covered remote sensing image set.

[0093] Computer program code for performing the operations of some embodiments of the present disclosure may be written in one or more programming languages or combinations thereof. The programming languages include object-oriented programming languages such as Java, Smalltalk, C++; and also include conventional procedural programming languages such as the "C" language or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, executed as an independent software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., by using an Internet service provider to connect through the Internet).

[0094] The flowcharts and block diagrams in the accompanying drawings illustrate the possible architectures, functions, and operations of systems, methods, and computer program products according to various embodiments of the present disclosure. In this regard, each block in the flowchart or block diagram may represent a module, a segment of a program, or a part of code that contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than that marked in the accompanying drawings. For example, two consecutive blocks shown may actually be executed substantially in parallel, and they may sometimes be executed in the reverse order, depending on the functions involved. It should also be noted that each block in the block diagram and / or flowchart, as well as combinations of blocks in the block diagram and / or flowchart, can be implemented by a dedicated hardware-based system that performs the specified functions or operations, or can be implemented by a combination of dedicated hardware and computer instructions.

[0095] The units described in some embodiments of the present disclosure can be implemented in software or in hardware. The described units can also be provided in a processor. For example, it can be described as: a processor includes an acquisition unit, a preliminary processing unit, a first processing unit, a segmentation unit, a second processing unit, an overlay unit, and a generation unit. Among them, the names of these units do not constitute a limitation to the unit itself in some cases. For example, the acquisition unit can also be described as "the unit that acquires each remote sensing image that meets the preset conditions as the initial data".

[0096] The functions described above herein can be performed at least in part by one or more hardware logic components. For example, without limitation, exemplary types of hardware logic components that can be used include: field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on a chip (SOCs), complex programmable logic devices (CPLDs), and so on.

[0097] Some embodiments of the present disclosure also provide a computer program product, including a computer program that, when executed by a processor, implements any of the above-described deep learning-based land expansion detection methods.

[0098] The above description is only some preferred embodiments of the present disclosure and an explanation of the applied technical principles. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present disclosure is not limited to the technical solutions formed by the specific combination of the above technical features, and should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above inventive concept. For example, the technical solutions formed by mutually replacing the above features with the (but not limited to) technical features having similar functions disclosed in the embodiments of the present disclosure.

Claims

1. A land expansion detection method based on deep learning, comprising: Obtaining each remote sensing image that meets preset conditions as initial data, where the initial data includes a first-phase remote sensing image group and a second-phase remote sensing image group; Performing preliminary preprocessing on the initial data to obtain preliminarily processed data, where the preliminary preprocessing includes cloud cover removal and size adjustment, and the preliminarily processed data includes a first-phase preliminary remote sensing image group and a second-phase preliminary remote sensing image group; Performing first preprocessing on the first-phase preliminary remote sensing image group to obtain first-phase segmented remote sensing images, where the first preprocessing includes band synthesis and image synthesis; Based on a semantic segmentation model, performing semantic segmentation on the first-phase segmented remote sensing images to obtain a land mask; Performing second preprocessing on the preliminarily processed data to obtain a remote sensing image set, where the second preprocessing includes band synthesis, image synthesis, resolution adjustment, and integration processing, and the remote sensing image set includes first-phase detection remote sensing images and second-phase detection remote sensing images; Based on the land mask, performing mask coverage on each detection remote sensing image in the remote sensing image set to obtain a covered remote sensing image set, where the covered remote sensing image set includes first-phase covered remote sensing images and second-phase covered remote sensing images; Based on a change detection model and the covered remote sensing image set, generating a visualization picture representing land expansion changes.

2. The method according to claim 1, wherein The obtaining each remote sensing image that meets preset conditions as initial data includes: Obtaining remote sensing image data as an initial image set, where the remote sensing image data includes each remote sensing image and the attribute data corresponding to each remote sensing image; According to a preset target area, screening each remote sensing image in the initial image set to obtain a target area image set; According to a preset cloud coverage rate, screening each remote sensing image in the target area image set to obtain a preliminary screening image set; According to a preset time range, screening each remote sensing image in the preliminary screening image set to obtain a target time image set; Based on the attribute data corresponding to each remote sensing image in the target time image set, classifying each remote sensing image in the target time image set to obtain a first-phase remote sensing image group and a second-phase remote sensing image group.

3. The method according to claim 1, wherein The performing preliminary preprocessing on the initial data to obtain preliminarily processed data includes: For each remote sensing image in the initial data, performing the following steps: Obtaining the attribute data of the remote sensing image; Constructing a cloud mask according to the attribute data of the remote sensing image; Based on the cloud mask, generating a preliminarily processed remote sensing image; Adjusting the preliminarily processed remote sensing image to a preset size to obtain a preliminary remote sensing image; Integrating each preliminary remote sensing image into preliminarily processed data.

4. The method according to claim 1, wherein, The performing first preprocessing on the first-phase preliminary remote sensing image group to obtain first-phase segmented remote sensing images includes: According to the band conditions of the segmentation model, perform band synthesis on each first-phase preliminary remote sensing image in the first-phase preliminary remote sensing image group to obtain each first-phase remote sensing image that meets the band conditions of the segmentation model; Perform image synthesis on each first-phase remote sensing image that meets the band conditions of the segmentation model to obtain a first-phase segmented remote sensing image.

5. The method according to claim 1, wherein The second preprocessing of the preliminary processed data to obtain a remote sensing image set includes: According to the band conditions of the detection model, perform band synthesis on each preliminary remote sensing image in the preliminary processed data to obtain each preliminary remote sensing image that meets the band conditions of the detection model; Perform image synthesis on each first-phase remote sensing image that meets the band conditions of the detection model and each second-phase remote sensing image that meets the band conditions of the detection model to obtain a first-phase target remote sensing image and a second-phase detection remote sensing image; According to the resolution condition of the detection model, adjust the resolution of the first-phase target remote sensing image to obtain a first-phase detection remote sensing image; Integrate the first-phase detection remote sensing image and the second-phase detection remote sensing image into a remote sensing image set.

6. The method according to claim 1, wherein, The generation of a visualization picture representing land expansion changes based on the change detection model and the post-covered remote sensing image set includes: According to the resolution condition of the detection model, adjust the resolution of each first-phase remote sensing image that meets the band conditions of the detection model to obtain each first-phase remote sensing image that meets the resolution conditions of the detection model; Based on the land mask, perform mask coverage on each first-phase remote sensing image that meets the resolution conditions of the detection model and each second-phase remote sensing image that meets the band conditions of the detection model to obtain a data set of the change detection model; Train the constructed change detection model with the data set of the change detection model to obtain a change detection model; Input the post-covered remote sensing image set into the change detection model to obtain a change mask of the first-phase post-covered remote sensing image and the second-phase post-covered remote sensing image; Overlay the change mask on the first-phase post-covered remote sensing image and the second-phase post-covered remote sensing image to obtain a visualization picture representing land expansion changes.

7. A land expansion detection device based on deep learning, comprising: An acquisition unit configured to acquire each remote sensing image that meets a preset condition as initial data, where the initial data includes a first-phase remote sensing image group and a second-phase remote sensing image group; A preliminary processing unit configured to perform preliminary preprocessing on the initial data to obtain preliminary processed data, where the preliminary preprocessing includes cloud cover removal and size adjustment, and the preliminary processed data includes a first-phase preliminary remote sensing image group and a second-phase preliminary remote sensing image group; A first processing unit configured to perform first preprocessing on the first-phase preliminary remote sensing image group to obtain a first-phase segmented remote sensing image, where the first preprocessing includes band synthesis and image synthesis; A segmentation unit configured to perform semantic segmentation on the first-phase segmented remote sensing image based on a semantic segmentation model to obtain a land mask; A second processing unit, configured to perform second preprocessing on the preliminary processed data to obtain a remote sensing image set, where the second preprocessing includes band synthesis, image synthesis, resolution adjustment, and integration processing, and the remote sensing image set includes a first-phase detection remote sensing image and a second-phase detection remote sensing image; A covering unit, configured to perform mask covering on each detection remote sensing image in the remote sensing image set based on the land mask to obtain a covered remote sensing image set, where the covered remote sensing image set includes a first-phase covered remote sensing image and a second-phase covered remote sensing image; A generating unit, configured to generate a visualization picture representing land expansion change based on a change detection model and the covered remote sensing image set.

8. An electronic device, comprising: One or more processors; A storage device having stored thereon one or more programs; When the one or more programs are executed by the one or more processors, enabling the one or more processors to implement the method according to any one of claims 1 to 6.

9. A computer-readable medium having a computer program stored thereon, wherein, The program, when executed by the processor, implements the method according to any one of claims 1 to 6.

Images