Mountain area artificial feature change monitoring system based on AI remote sensing

Abstract

The invention discloses a mountain area artificial feature change monitoring system based on AI remote sensing, which comprises a mountain area data enhancement module, a model training module and a dual-model driven artificial feature change monitoring module, and is characterized in that the mountain area data enhancement module comprises remote sensing image super-division and remote sensing image generation; the model training module comprises sample data set making and model training, the dual-model-driven artificial feature change monitoring module comprises dual-model-driven predictive reasoning, data fusion and post-processing, and building change pattern spots occurring in the past are automatically extracted by using a post-time-phase remote sensing image and a territorial change investigation database; automatically extracting newly occurring building change pattern spots in a specific time period from the front and back time phase remote sensing images; and combining the building change pattern spots extracted in the two modes to obtain a building change result. The method can improve the quality of remote sensing images, reduces the missing extraction of change information, and greatly improves the monitoring efficiency and the result precision.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing interpretation, and particularly to a mountainous area artificial ground object change monitoring system based on AI remote sensing. Background Art

[0002] With the rapid development of remote sensing technology and artificial intelligence (AI), ground object change monitoring plays an increasingly important role in fields such as environmental protection, urban planning, and disaster warning. Especially in mountainous areas, due to the complex terrain and changeable climate, the distribution and change of artificial ground objects have uniqueness and complexity. Therefore, monitoring the change of artificial ground objects in these areas is of great significance for regional sustainable development. However, although remote sensing AI technology shows good potential in artificial ground object change monitoring, there are still many challenges in practical applications.

[0003] Firstly, the image quality problem is one of the main bottlenecks in mountainous area artificial ground object change monitoring. The terrain in mountainous areas has large undulations and frequent cloud cover, resulting in low-resolution and noisy remote sensing images. Traditional image processing methods are difficult to effectively improve the image quality, severely limiting the accuracy of change monitoring.

[0004] Secondly, the single monitoring method is another prominent problem in current artificial ground object change monitoring. Most existing artificial ground object change monitoring methods are based on the comparative analysis of two-phase images, and the change information of artificial ground objects is extracted to achieve the monitoring. However, this method has a strong dependence on the time interval between the front and back temporal remote sensing images, and can only extract the change information of newly occurred artificial ground objects within a specific time period. Summary of the Invention

[0005] The purpose of the present invention is to provide a mountainous area artificial ground object change monitoring system based on AI remote sensing that can improve the quality of remote sensing images, reduce the omission of change information extraction, and greatly improve the monitoring efficiency and the accuracy of the results, so as to overcome the above problems.

[0006] To achieve the above purpose, the present invention adopts the following technical solutions.

[0007] A mountainous area artificial ground object change monitoring system based on AI remote sensing of the present invention includes a mountainous area data enhancement module, a model training module, and a dual-model-driven artificial ground object change monitoring module, wherein: The mountainous area data enhancement module described above includes remote sensing image super-resolution and remote sensing image generation; the model training module includes sample dataset production and model training. The sample dataset production uses a semi-automatic annotation tool developed based on the SAM interactive segmentation model to produce sample datasets for building semantic segmentation and building change detection. For model training, U-Net++ is selected as the basic training model for building semantic segmentation. The sample dataset for building semantic segmentation is input, and the training parameters are configured. After iterative training, a high-precision building semantic segmentation model is obtained. The TimeTravelling Pixels (TTP) model is selected as the basic training model for building change detection. The sample dataset for building change detection is input, and the training parameters are configured. After iterative training, a high-precision building change detection model is obtained. The dual-model-driven artificial ground object change monitoring module includes dual-model-driven prediction inference, data fusion, and post-processing. The dual-model-driven prediction inference is based on the high-precision building semantic segmentation model. Using the post-temporal remote sensing image and the land use change survey database, the building change patches that occurred in the past are automatically extracted. Based on the high-precision building change detection model, the building change patches that newly occurred within a specific time period are automatically extracted using the pre- and post-temporal remote sensing images. Data fusion and post-processing involve merging the building change patches that occurred in the past, which are automatically extracted using the post-temporal remote sensing image and the land use change survey database, with the building change patches that newly occurred within a specific time period, which are automatically extracted using the pre- and post-temporal remote sensing images, to obtain the preliminary results of building changes. The preliminary results of building changes are subjected to operations such as removing small patches, filling holes, optimizing graphic topology, and rectifying to obtain the results of building changes.

[0008] For the method of remote sensing image super-resolution in the mountainous area data enhancement module of the above-mentioned AI remote sensing-based mountainous area artificial ground object change monitoring system: Two remote sensing images with a size of 9216×9216, 3 bands, and a resolution of 2 meters are input. One model is selected from the SRGAN and ResShift models, and 4-fold super-resolution is selected to obtain two remote sensing images with a size of 36864×36864, 3 bands, and a resolution of 0.5 meters. One is used as the pre-temporal remote sensing image, and the other is used as the post-temporal remote sensing image.

[0009] The above-mentioned mountainous area artificial ground object change monitoring system based on AI remote sensing, wherein the method for generating remote sensing images of the mountainous area data enhancement module is as follows: Select one of the GeoSynth and DiffusionSat models and input the following content: "Generate a mountainous area remote sensing image with a size of 36864×36864, 3 bands, a resolution of 0.5 meters, and including cultivated land, forest land, and water bodies". The system automatically generates a mountainous area remote sensing image with a size of 36864×36864, 3 bands, a resolution of 0.5 meters, and including cultivated land, forest land, and water bodies as the previous-phase remote sensing image; Through image processing, replace some of the cultivated land and forest land in the previous-phase remote sensing image with houses to obtain a mountainous area remote sensing image with a size of 36864×36864, 3 bands, a resolution of 0.5 meters, and including cultivated land, forest land, houses, and water bodies as the later-phase remote sensing image.

[0010] The above-mentioned mountainous area artificial ground object change monitoring system based on AI remote sensing, wherein the method for making the sample data set of the model training module includes the following steps; Step one: Sample annotation 1) Building semantic segmentation sample annotation: Select a semi-automatic annotation tool, input the later-phase remote sensing image, create a new vector layer, set the geometric type to polygon, and save the new layer as the building semantic segmentation vector layer; Enable the editing mode, select "point hint" or "line hint" or "polygon hint", and semi-automatically annotate the building positions on the later-phase remote sensing image; Save the annotation information and update the building semantic segmentation vector layer; 2) Building change detection sample annotation: Select a semi-automatic annotation tool, input the previous-phase and later-phase remote sensing images respectively, create a new vector layer, set the geometric type to polygon, and save the new layer as the building change detection vector layer; Enable the editing mode, select "point hint" or "line hint" or "polygon hint", and semi-automatically annotate the building change positions according to the previous-phase and later-phase remote sensing images; Save the annotation information and update the building change detection vector layer; Step two: Vector to raster conversion: Input the building semantic segmentation vector layer, set the output pixel size to 0.5 meters to obtain the building semantic segmentation raster label. In the building semantic segmentation raster label, the pixel value of the building is 1, and the pixel values of other ground objects are 0; Input the building change detection vector layer. The method of converting the building change detection vector layer into the building change detection raster label is the same as that of converting the building semantic segmentation vector layer into the building semantic segmentation raster label. In the building change detection raster label, the pixel value of the building change area is 1, and the pixel values of other areas are 0; Step three: Sample slicing 1) Building semantic segmentation sample slices: Input the post-temporal remote sensing image and the building semantic segmentation raster label. Set the cropping size to 1024×1024, the cropping step to 1024, and the validation ratio to 1:9. Obtain a building semantic segmentation sample dataset containing 1292 sample data pairs. Each sample data pair contains a raster label and a post-temporal remote sensing image with the same resolution and size. Among them, the validation set contains 129 sample data pairs, and the training set contains 1163 sample data pairs; 2) Building change detection sample slices: Input the pre- and post-temporal remote sensing images and the building change detection raster label. Set the cropping size to 1024×1024, the cropping step to 1024, and the validation ratio to 1:9. Obtain a building change detection sample dataset containing 1292 sample data pairs. Each sample data pair contains a raster label, a pre-temporal remote sensing image, and a post-temporal remote sensing image with the same resolution and size. Among them, the validation set contains 129 sample data pairs, and the training set contains 1163 sample data pairs.

[0011] In the above mountainous area artificial ground object change monitoring system based on AI remote sensing, the model training method of the model training module includes the following steps: Step 1: Select the basic training model: For the business requirements of artificial ground object change monitoring, select U-Net++ as the basic training model for building semantic segmentation, and select TTP as the basic training model for building change detection; Step 2: Configure the training task data: In the building semantic segmentation model training task, input the building semantic segmentation sample dataset to provide sample data support for training the U-Net++ model; in the building change detection model training task, input the building change detection sample dataset to provide sample data support for training the TTP model; Step 3: Configure the training task parameters: In the building semantic segmentation model and building change detection model training tasks, select Pytorch for the task framework, GPU for the training engine, set the learning rate to 0.001, set the batch size to 4, select Adam as the optimizer, and select the binary cross-entropy function as the loss function; Step 4: Adjust the training data and parameters: According to the training results, perform positive sample augmentation on the missed parts and negative sample augmentation on the misclassified parts, and retrain based on the training results until the model accuracy improves to a certain level and then basically fluctuates around a value. Save the high-precision building semantic segmentation model and building change detection model.

[0012] In the above mountainous area artificial ground object change monitoring system based on AI remote sensing, the dual-model-driven prediction and inference of the dual-model-driven artificial ground object change monitoring module includes the following steps: Step 1 Remote sensing image cropping: Input two remote sensing images with a size of 36864×36864, 3 bands, and a resolution of 0.5 meters, which are used as the former-phase remote sensing image and the latter-phase remote sensing image respectively. Select the method of "drawing a point-shaped area" or "drawing a flow-shaped area" or "drawing a rectangle" to frame the prediction range, and select the overlapping mode with an overlap degree set to 10%. Crop the former and latter-phase remote sensing images within the prediction range into small remote sensing image pieces of 1024×1024 size; Step 2 Feature extraction and prediction: Include building feature extraction and segmentation, and building change feature extraction and prediction; 1) Building feature extraction and segmentation: Select a high-precision building semantic segmentation model, input the small pieces of the latter-phase remote sensing image. After setting the probability threshold to 0.5, extract the building features of each small remote sensing image piece, and segment the buildings in the small remote sensing image piece based on the extracted building features to generate the preliminary building segmentation results of each small remote sensing image piece; 2) Building change feature extraction and prediction: Select a high-precision building change detection model, input the small pieces of the former and latter-phase remote sensing images. After setting the probability threshold to 0.5, extract the building change features of each pair of small remote sensing image pieces, and perform prediction and reasoning on the change information in each pair of remote sensing images based on the extracted building change features to generate the preliminary building change prediction results of each pair of small remote sensing image pieces; Step 3 Prediction result stitching: According to the coordinate information recorded during cropping, seamlessly stitch the preliminary building segmentation results and the preliminary building change prediction results of all small remote sensing image pieces back to the original image size to obtain the building segmentation result and the building change prediction result; For the overlapping area, the average value method is used to ensure the stitching quality; Step 4 Raster to vector: Vectorize the building segmentation result and the building change prediction result respectively to obtain the building segmentation result patches and the preliminary building change result 1; Step 5 Data erasure: First input the building segmentation result patches, and then input the building patches in the land use change survey database within the same range. Erase the overlapping patches to obtain the preliminary building change result 2.

[0013] For the above-mentioned mountainous area artificial ground object change monitoring system based on AI remote sensing, the data fusion and post-processing of the dual-model-driven artificial ground object change monitoring module include the following steps: 1) Data fusion: Calculate the union of the preliminary building change result 1 and the preliminary building change result 2 to obtain the preliminary building change result; 2) Data post-processing: Removing fragmented patches: Input the preliminary results of building changes, set the minimum patch area to 50 square meters, remove fragmented patches smaller than 50 square meters, and obtain the preliminary results after removing fragmented patches; Filling holes: Input the preliminary results after removing fragmented patches, set the threshold to 100 square meters. That is, when the hole area is less than 100 square meters, the system will automatically fill it to obtain the preliminary results without hole areas; Graphic topology optimization: Select the preliminary results without hole areas, create a new topology, and add two topology rules, "cannot self-overlap" and "cannot self-intersect", to obtain the preliminary results that conform to the topology rules; Rectification processing: Input the preliminary results that conform to the topology rules, set the tolerance to 5 meters, and output the building change results.

[0014] Compared with the prior art, the present invention has obvious beneficial effects. From the above technical solutions, it can be seen that the mountain area data enhancement module of the present invention performs super-resolution reconstruction on the original remote sensing image with a resolution lower than 1 meter through the SRGAN and ResShift models for super-resolution utilization of remote sensing images, and obtains a remote sensing image with a resolution better than 1 meter; the remote sensing image generation is based on the GeoSynth and DiffusionSat models, and through inputting text descriptions, automatically generates a remote sensing image with a resolution better than 1 meter, effectively improving the quality of remote sensing images and increasing the number of remote sensing images, providing high-quality data support for mountain area sample collection and change monitoring work; the model training module selects U-Net++ as the basic training model for building semantic segmentation and selects TTP as the basic training model for building change detection, and obtains a high-precision building semantic segmentation model and a building change detection model through iterative training, providing model support for artificial ground object change monitoring; the dual-model-driven artificial ground object change monitoring module, on the basis of the original artificial ground object change monitoring method that uses the building change detection model and compares and analyzes remote sensing images of different time phases, introduces a building semantic segmentation model, and uses the post-time-phase remote sensing image plus the national land change survey database method, breaking the limitation that the traditional method can only extract the information of newly occurred artificial ground object changes within a specific time period, further improving the artificial ground object change discovery mechanism, effectively solving the problem of missing extraction of artificial ground object change information in mountain areas. At the same time, this module integrates data post-processing methods such as removing small patches, filling holes, graphic topology optimization, and rectification processing, which helps to improve the usability of prediction results. Description of the Drawings

[0015] Figure 1 is the schematic diagram of the present invention.

[0016] Figure 2 is the partial effect diagram of the super-resolution reconstruction of the remote sensing image of the present invention.

[0017] Figure 3It is the automatically generated local effect diagram of the remote sensing image of the present invention.

[0018] Figure 4 It is the result display diagram of the present invention. Specific implementation manners

[0019] The following further illustrates the concept, specific structure and technical effects of a mountainous area artificial ground object change monitoring system based on AI remote sensing proposed by the present invention in conjunction with the accompanying drawings, so as to fully understand the purpose, features and effects of the present invention.

[0020] As Figure 1 shown, a mountainous area artificial ground object change monitoring system based on AI remote sensing includes a mountainous area data enhancement module, a model training module and a dual-model-driven artificial ground object change monitoring module, where: The mountainous area data enhancement module includes remote sensing image super-resolution and remote sensing image generation. The quality of remote sensing images is improved through remote sensing image super-resolution, and the number of remote sensing images is increased through remote sensing image generation, providing sufficient high-quality remote sensing image support for ground object monitoring in mountainous areas; The model training module includes sample data set production and model training. The semi-automatic annotation tool developed based on the SAM interactive segmentation model is used to produce the sample data sets for building semantic segmentation and building change detection; U-Net++ is selected as the basic training model for building semantic segmentation, the sample data set for building semantic segmentation is input, and the training parameters are configured. After iterative training, a high-precision building semantic segmentation model is obtained. TTP is selected as the basic training model for building change detection, the sample data set for building change detection is input, and the training parameters are configured. After iterative training, a high-precision building change detection model is obtained; The dual-model-driven artificial ground object change monitoring module includes dual-model-driven prediction inference, data fusion and post-processing. The dual-model-driven prediction inference is based on the high-precision building semantic segmentation model, and the building change patches that occurred in the past are automatically extracted using the later-phase remote sensing image and the land use change survey database. Based on the high-precision building change detection model, the building change patches that newly occurred within a specific time period are automatically extracted using the pre- and post-phase remote sensing images; data fusion and post-processing is to merge the building change patches that occurred in the past automatically extracted using the later-phase remote sensing image and the land use change survey database with the building change patches that newly occurred within a specific time period automatically extracted using the pre- and post-phase remote sensing images to obtain the preliminary results of building changes, and the preliminary results of building changes are subjected to small patch removal, hole filling, graphic topology optimization and right-angled processing to obtain the building change results.

[0021] The method for super-resolution of remote sensing images of the mountainous area data enhancement module is as follows: Input two scenes of remote sensing images with a size of 9216×9216, 3 bands, and a resolution of 2 meters. Select one model from the SRGAN and ResShift models, and select 4-fold super-resolution to obtain two scenes of remote sensing images with a size of 36864×36864, 3 bands, and a resolution of 0.5 meters (see Figure 2 ), one scene as the previous-phase remote sensing image and one scene as the subsequent-phase remote sensing image; The method for generating remote sensing images of the mountainous area data enhancement module is as follows: Select one model from the GeoSynth and DiffusionSat models and input the following content: "Generate a mountainous area remote sensing image with a size of 36864×36864, 3 bands, a resolution of 0.5 meters, and including cultivated land, forest land, and water bodies". The system automatically generates a mountainous area remote sensing image with a size of 36864×36864, 3 bands, a resolution of 0.5 meters, and including cultivated land, forest land, and water bodies as the previous-phase remote sensing image (see Figure 3 ); Through image processing, replace some of the cultivated land and forest land in the previous-phase remote sensing image with houses to obtain a mountainous area remote sensing image with a size of 36864×36864, 3 bands, a resolution of 0.5 meters, and including cultivated land, forest land, houses, and water bodies as the subsequent-phase remote sensing image.

[0022] The method for making the sample data set of the model training module includes the following steps; Step 1 Sample annotation: 1) Building semantic segmentation sample annotation: Select a semi-automatic annotation tool, input the subsequent-phase remote sensing image, create a new vector layer, set the geometric type to polygon, and save the new layer as the building semantic segmentation vector layer; Enable the editing mode, select "point hint" or "line hint" or "polygon hint", and semi-automatically annotate the building positions on the subsequent-phase remote sensing image; Save the annotation information and update the building semantic segmentation vector layer; 2) Building change detection sample annotation: Select a semi-automatic annotation tool, input the previous-phase and subsequent-phase remote sensing images respectively, create a new vector layer, set the geometric type to polygon, and save the new layer as the building change detection vector layer; Enable the editing mode, select "point hint" or "line hint" or "polygon hint", and semi-automatically annotate the building change positions according to the previous-phase and subsequent-phase remote sensing images; Save the annotation information and update the building change detection vector Quantity layer ; Step 2 Vector to Raster: Input the vector layer of building semantic segmentation, set the output pixel size to 0.5 meters, and obtain the raster label of building semantic segmentation. In the raster label of building semantic segmentation, the pixel value of the building is 1, and the pixel values of other features are 0. Input the vector layer of building change detection. The conversion of the vector layer of building change detection into the raster label of building change detection is the same as the method of converting the vector layer of building semantic segmentation into the raster label of building semantic segmentation. In the raster label of building change detection, the pixel value of the building change area is 1, and the pixel values of other areas are 0. Step 3 Sample Slicing: Input the remote sensing image and the raster label, set the cutting size, cutting step, and validation ratio to obtain the sample dataset. 1) Sample Slicing for Building Semantic Segmentation: Input the post-temporal remote sensing image and the raster label of building semantic segmentation. Set the cutting size to 1024×1024, the cutting step to 1024, and the validation ratio to 1:9 (i.e., the number of samples used as the validation set accounts for 10% of the total number of samples). Obtain a sample dataset for building semantic segmentation containing 1292 sample data pairs. Each sample data pair contains a raster label and a post-temporal remote sensing image with the same resolution and size. Among them, the validation set contains 129 sample data pairs, and the training set contains 1163 sample data pairs. 2) Sample Slicing for Building Change Detection: Input the pre-temporal and post-temporal remote sensing images and the raster label of building change detection. Set the cutting size to 1024×1024, the cutting step to 1024, and the validation ratio to 1:9 (i.e., the number of samples used as the validation set accounts for 10% of the total number of samples). Obtain a sample dataset for building change detection containing 1292 sample data pairs. Each sample data pair contains a raster label, a pre-temporal remote sensing image, and a post-temporal remote sensing image with the same resolution and size. Among them, the validation set contains 129 sample data pairs, and the training set contains 1163 sample data pairs.

[0023] The model training method of the model training module includes the following steps: Step 1 Select the Basic Training Model: Facing the business requirements of artificial feature change monitoring, select U-Net++ as the basic training model for building semantic segmentation, and select TTP as the basic training model for building change detection. Step 2 Training Task Data Configuration: In the training task of the building semantic segmentation model, input the sample dataset of building semantic segmentation to provide sample data support for training the U-Net++ model. In the training task of the building change detection model, input the sample dataset of building change detection to provide sample data support for training the TTP model. Step 3 Training task parameter configuration: In the training tasks of the building semantic segmentation model and the building change detection model, the task framework selects Pytorch, the training engine selects GPU, the learning rate is set to 0.001, the batch size is set to 4, the optimizer selects Adam, and the loss function selects the binary cross-entropy function; Step 4 Adjust training data and parameters: According to the training results, expand the positive samples for the parts that are missed in extraction, expand the negative samples for the parts that are mis-extracted, and re-train based on the training results until the model accuracy increases to a certain level and then basically fluctuates around a value. Save the building semantic segmentation model and the building change detection model with high accuracy.

[0024] The dual-model-driven prediction and inference of the artificial ground feature change monitoring module described above is to input the remote sensing images of the previous and current time phases, select the prediction range, and use the high-precision building change detection model and the remote sensing images of the previous and current time phases to automatically extract the preliminary results 1 of building changes; use the high-precision building semantic segmentation model and the remote sensing image of the current time phase to automatically extract the building segmentation result patches, and erase the building segmentation result patches and the building patches in the land use change survey database within the same range to delete the overlapping patches and obtain the preliminary results 2 of building changes. It includes the following steps: Step 1 Remote sensing image cropping: Input two remote sensing images with a size of 36864×36864, 3 bands, and a resolution of 0.5 meters, respectively, as the remote sensing images of the previous and current time phases. Select the method of "drawing a dot-shaped area" or "drawing a flowing area" or "drawing a rectangle" to frame the prediction range, and select the overlapping mode with an overlap degree set to 10%. Crop the remote sensing images of the previous and current time phases within the prediction range into small remote sensing image pieces of 1024×1024 size; Step 2 Feature extraction and prediction: It includes building feature extraction and segmentation and building change feature extraction and prediction; 1) Building feature extraction and segmentation: Select the high-precision building semantic segmentation model, input the small remote sensing image pieces of the current time phase, set the probability threshold to 0.5, extract the building features of each small remote sensing image piece, and segment the buildings in each small remote sensing image piece based on the extracted building features to generate the preliminary building segmentation results of each small remote sensing image piece; 2) Extraction and prediction of building change features: Select a high-precision building change detection model, input small pieces of remote sensing images at different times, set the probability threshold to 0.5, extract the building change features of each pair of small pieces of remote sensing images, and based on the extracted building change features, predict and infer the change information in each pair of small pieces of remote sensing images to generate the preliminary results of building change prediction for each pair of small pieces of remote sensing images; Step 3 Stitching of prediction results: According to the coordinate information recorded during cropping, seamlessly stitch the preliminary results of building segmentation and the preliminary results of building change prediction of all small pieces of remote sensing images back to the original image size to obtain the building segmentation results and the building change prediction results; for the overlapping areas, the average value method is used to ensure the stitching quality; Step 4 Raster to vector: Vectorize the building segmentation results and the building change prediction results respectively to obtain the polygon maps of the building segmentation results and the preliminary results of building change 1 (see Figure 4 ); Step 5 Data erasure: First input the polygon maps of the building segmentation results, and then input the building polygon maps in the national land change survey database within the same range. Erase the overlapping polygons to obtain the preliminary results of building change 2 (see Figure 4 ).

[0025] For the data fusion and post-processing of the dual-model-driven artificial ground object change monitoring module, merge the preliminary results of building change 1 and the preliminary results of building change 2 to obtain the preliminary results of building change; perform operations such as removing small patches, filling holes, optimizing graphic topology, and rectifying to right angles on the preliminary results of building change to obtain the building change results; including the following steps: 1) Data fusion: Calculate the union of the preliminary results of building change 1 and the preliminary results of building change 2 to obtain the preliminary results of building change; 2) Data post-processing: Perform operations such as removing small patches, filling holes, optimizing graphic topology, and rectifying to right angles on the preliminary results of building change to obtain the building change results; the specific method is: Removing small patches: Input the preliminary results of building change, set the minimum patch area to 50 square meters, and remove the small and fragmented patches smaller than 50 square meters to obtain the preliminary results after removing small patches; Filling holes: Input the preliminary results after removing small patches, set the threshold to 100 square meters, that is, when the hole area is less than 100 square meters, the system will automatically fill it to obtain the preliminary results without hole areas; Graphic topology optimization: Select the preliminary results without hole areas, create a new topology, and add two topology rules of "cannot self-overlap" and "cannot self-intersect" to obtain the preliminary results that meet the topology rules; Rectifying to right angles: Input the preliminary results that meet the topology rules, set the tolerance to 5 meters, and output the building change results.

[0026] As described above, it is only the preferred embodiment of the present invention, and it does not impose any form of limitation on the present invention. Any simple modification, equivalent change, and modification made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention still fall within the scope of the technical solution of the present invention.

Claims

1. An artificial ground object change monitoring system for mountainous areas based on AI remote sensing, comprising a mountainous area data enhancement module, a model training module, and a dual-model-driven artificial ground object change monitoring module, characterized in that: The described mountainous area data enhancement module includes remote sensing image super-resolution and remote sensing image generation; the model training module includes sample dataset production and model training. The sample dataset production uses a semi-automatic annotation tool developed based on the SAM interactive segmentation model to produce building semantic segmentation and building change detection sample datasets. For model training, U-Net++ is selected as the basic training model for building semantic segmentation. The building semantic segmentation sample dataset is input, and training parameters are configured. After iterative training, a high-precision building semantic segmentation model is obtained. TTP is selected as the basic training model for building change detection. The building change detection sample dataset is input, and training parameters are configured. After iterative training, a high-precision building change detection model is obtained; the dual-model-driven artificial ground object change monitoring module includes dual-model-driven prediction inference, data fusion, and post-processing. The dual-model-driven prediction inference is based on the high-precision building semantic segmentation model. Using the later-phase remote sensing image and the land use change survey database, the building change patches that occurred in the past are automatically extracted. Based on the high-precision building change detection model, the building change patches that occurred newly within a specific time period are automatically extracted using the pre- and post-phase remote sensing images; data fusion and post-processing is to merge the building change patches that occurred in the past automatically extracted using the later-phase remote sensing image and the land use change survey database with the building change patches that occurred newly within a specific time period automatically extracted using the pre- and post-phase remote sensing images to obtain the preliminary results of building changes. The preliminary results of building changes are subjected to operations such as removing small patches, filling holes, graphic topology optimization, and rectification to obtain the building change results.

2. The mountainous area artificial ground object change monitoring system based on AI remote sensing according to claim 1, characterized in that: The method for remote sensing image super-resolution of the described mountainous area data enhancement module is as follows: Input two remote sensing images with a size of 9216×9216, 3 bands, and a resolution of 2 meters. Select one model from the SRGAN and ResShift models, and select 4-fold super-resolution to obtain two remote sensing images with a size of 36864×36864, 3 bands, and a resolution of 0.5 meters. One is used as the pre-phase remote sensing image, and the other is used as the post-phase remote sensing image.

3. The artificial ground object change monitoring system for mountainous areas based on AI remote sensing according to claim 1 or 2, characterized in that: The method for remote sensing image generation of the described mountainous area data enhancement module is as follows: Select one model from the GeoSynth and DiffusionSat models and input the following content: "Generate a mountainous area remote sensing image with a size of 36864×36864, 3 bands, and a resolution of 0.5 meters, and containing cultivated land, forest land, and water bodies". The system automatically generates a mountainous area remote sensing image with a size of 36864×36864, 3 bands, and a resolution of 0.5 meters, and containing cultivated land, forest land, and water bodies, as the pre-phase remote sensing image; through image processing, part of the cultivated land and forest land in the pre-phase remote sensing image are replaced with houses to obtain a mountainous area remote sensing image with a size of 36864×36864, 3 bands, and a resolution of 0.5 meters, and containing cultivated land, forest land, houses, and water bodies, as the post-phase remote sensing image.

4. The artificial ground object change monitoring system for mountainous areas based on AI remote sensing according to claim 1, wherein: The method for making the sample data set of the described model training module includes the following steps; Step 1: Sample annotation 1) Building semantic segmentation sample annotation: Select a semi-automatic annotation tool, input the post-temporal remote sensing image, create a new vector layer, set the geometric type to polygon, and save the newly created layer as the building semantic segmentation vector layer; Enable the editing mode, select "point hint" or "line hint" or "polygon hint", and semi-automatically annotate the building positions on the post-temporal remote sensing image; Save the annotation information and update the building semantic segmentation vector layer; 2) Building change detection sample annotation: Select a semi-automatic annotation tool, input the pre-temporal and post-temporal remote sensing images respectively, create a new vector layer, set the geometric type to polygon, and save the newly created layer as the building change detection vector layer; Enable the editing mode, select "point hint" or "line hint" or "polygon hint", and semi-automatically annotate the building change positions according to the pre-temporal and post-temporal remote sensing images; Save the annotation information and update the building change detection vector layer; Step 2: Vector to raster conversion: Input the building semantic segmentation vector layer, set the output pixel size to 0.5 meters to obtain the building semantic segmentation raster label. In the building semantic segmentation raster label, the pixel value of the building is 1, and the pixel values of other ground objects are 0; Input the building change detection vector layer. The conversion of the building change detection vector layer to the building change detection raster label is the same as the method of converting the building semantic segmentation vector layer to the building semantic segmentation raster label. In the building change detection raster label, the pixel value of the building change area is 1, and the pixel values of other areas are 0; Step 3: Sample slicing 1) Building semantic segmentation sample slicing: Input the post-temporal remote sensing image and the building semantic segmentation raster label, set the cropping size to 1024×1024, the cropping step to 1024, and the validation ratio to 1:9 to obtain a building semantic segmentation sample data set containing 1292 sample data pairs. Each sample data pair contains 1 raster label and 1 post-temporal remote sensing image with the same resolution and size. Among them, the validation set contains 129 sample data pairs, and the training set contains 1163 sample data pairs; 2) Building change detection sample slicing: Input the pre-temporal and post-temporal remote sensing images and the building change detection raster label, set the cropping size to 1024×1024, the cropping step to 1024, and the validation ratio to 1:9 to obtain a building change detection sample data set containing 1292 sample data pairs. Each sample data pair contains 1 raster label, 1 pre-temporal remote sensing image, and 1 post-temporal remote sensing image with the same resolution and size. Among them, the validation set contains 129 sample data pairs, and the training set contains 1163 sample data pairs.

5. The monitoring system for changes in artificial ground objects in mountainous areas based on AI remote sensing according to claim 1 or 4, characterized in that: The model training method of the described model training module includes the following steps: Step 1: Select the basic training model: Facing the business requirements of artificial ground object change monitoring, select U-Net++ as the basic training model for building semantic segmentation, and select TTP as the basic training model for building change detection; Step 2: Training task data configuration: In the building semantic segmentation model training task, input the building semantic segmentation sample data set to provide sample data support for training the U-Net++ model; in the building change detection model training task, input the building change detection sample data set to provide sample data support for training the TTP model. Step 3: Training task parameter configuration: In the building semantic segmentation model and building change detection model training tasks, select Pytorch for the task framework, GPU for the training engine, set the learning rate to 0.001, set the batch size to 4, select Adam for the optimizer, and select the binary cross-entropy function for the loss function. Step 4: Adjust training data and parameters: According to the training results, perform positive sample augmentation on the missed parts and negative sample augmentation on the misclassified parts, and retrain based on the training results until the model accuracy improves to a certain level and then basically approaches a value with fluctuations. Save the building semantic segmentation model and building change detection model with high accuracy.

6. The mountainous area artificial ground object change monitoring system based on AI remote sensing according to claim 1, characterized in that: The dual-model-driven prediction and inference of the artificial ground object change monitoring module described above includes the following steps: Step 1: Remote sensing image cropping: Input two remote sensing images with a size of 36864×36864, 3 bands, and a resolution of 0.5 meters, which are used as the former-temporal remote sensing image and the latter-temporal remote sensing image respectively. Select the method of "drawing a dot-shaped area" or "drawing a flowing area" or "drawing a rectangle" to frame the prediction range, and select the overlapping mode with an overlap degree set to 10%. Crop the former-temporal and latter-temporal remote sensing images within the prediction range into small remote sensing image patches of 1024×1024 size. Step 2: Feature extraction and prediction: Include building feature extraction and segmentation and building change feature extraction and prediction. 1) Building feature extraction and segmentation: Select a high-precision building semantic segmentation model, input the latter-temporal remote sensing image patch, set the probability threshold to 0.5, extract the building features of each remote sensing image patch, and segment the buildings in the remote sensing image patch based on the extracted building features to generate the preliminary building segmentation results for each remote sensing image patch. 2) Building change feature extraction and prediction: Select a high-precision building change detection model, input the former-temporal and latter-temporal remote sensing image patches, set the probability threshold to 0.5, extract the building change features of each pair of remote sensing image patches, and perform prediction and inference on the change information in each pair of remote sensing image patches based on the extracted building change features to generate the preliminary building change prediction results for each pair of remote sensing image patches. Step 3: Prediction result stitching: According to the coordinate information recorded during cropping, seamlessly stitch the preliminary building segmentation results and preliminary building change prediction results of all remote sensing image patches back to the original image size to obtain the building segmentation result and building change prediction result; for the overlapping area, use the average value method to ensure the stitching quality. Step 4: Raster to vector: Vectorize the building segmentation result and building change prediction result respectively to obtain the building segmentation result map patches and the preliminary building change result 1. Step 5: Data Erasure: First, input the building segmentation result patches, and then input the building patches in the land use change survey database within the same range. The preliminary result 2 of building changes is obtained by erasing the overlapping patches.

7. The monitoring system for changes in artificial ground objects in mountainous areas based on AI remote sensing according to claim 1 or 6, characterized in that: The data fusion and post-processing of the double-model-driven artificial feature change monitoring module include the following steps: 1) Data Fusion: The preliminary result of building changes is obtained by calculating the union of the preliminary result 1 of building changes and the preliminary result 2 of building changes. 2) Data Post-processing: Removing Small Patches: Input the preliminary result of building changes, set the minimum patch area to 50 square meters, and remove the small and fragmented patches smaller than 50 square meters to obtain the preliminary result after removing small patches. Hole Filling: Input the preliminary result after removing small patches, set the threshold to 100 square meters. That is, when the hole area is less than 100 square meters, the system will automatically fill it to obtain the preliminary result without hole areas. Graphic Topology Optimization: Select the preliminary result without hole areas, create a new topology, and add two topology rules: "Cannot self-overlap" and "Cannot self-intersect" to obtain the preliminary result that conforms to the topology rules. Rectangularization Processing: Input the preliminary result that conforms to the topology rules, set the tolerance to 5 meters, and output the building change result.

Images