Specific object remote sensing extraction method and system based on visual Transform

Abstract

The invention discloses a visual Transform-based specific object remote sensing extraction method and system, and belongs to the technical field of remote sensing image processing and deep learning. The method comprises the following steps: firstly, extracting quality parameters such as a cloud shielding rate, AOD (Argon Oxygen Decarburization) and a wetting state from a high-resolution image, calculating a comprehensive quality score for each sub-block, and classifying the sub-blocks into a high-quality region, a medium-quality region and a low-quality region; different processing strategies are adopted for different quality areas; and training a CNN + Transform double-branch model, finally generating a high-precision specific object classification graph, and outputting an uncertainty thermodynamic diagram and a to-be-rechecked label at the same time. The system has the technical advantages of high precision, strong robustness, explainable result and closed-loop optimization.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of remote sensing image processing and deep learning, and in particular to a specific object remote sensing extraction method and system based on visual Transformer. BACKGROUND

[0002] In recent years, remote sensing image semantic segmentation technology has developed rapidly, especially the method based on visual Transformer has made significant progress. Traditional convolutional neural network (CNN) is good at capturing local texture and structure information, but it often has problems such as edge recognition not fine and region misjudgment when dealing with complex target boundary and scattered features (such as abandoned land, road, building fragments). In addition, the existing specific object extraction technology does not fully consider the robustness of the model to environmental interference such as clouds, shadows, haze, which is specifically reflected in:

[0003] Quality parameters are not fully integrated: lack of processing of environmental factors such as aerosol concentration (AOD), terrain slope, and precipitation conditions, which may introduce recognition bias by inputting low-quality subgraphs;

[0004] Rigid block and splicing strategy: usually fixed overlap ratio and probability mean splicing are used, ignoring model confidence and local uncertainty, which may cause splicing edge breakage or artifact problems;

[0005] No compensation mechanism for medium and low quality areas: for lightly occluded or partially shaded areas, existing technologies often use full skipping or rough prediction, which may cause potential target specific objects to be missed;

[0006] Lack of system closed-loop design: lack of manual review, feedback mechanism and dynamic update strategy, which cannot continuously optimize model performance for errors. SUMMARY

[0007] The present application aims to overcome the above-mentioned defects of the prior art and provides a specific object remote sensing extraction method and system based on visual Transformer, which is a complete process from the data quality level, combined with multi-source temporal information, adaptive block strategy and uncertainty-driven splicing mechanism, to realize a specific object remote sensing extraction method with high precision, high robustness and interpretability.

[0008] In a first aspect, the present application provides a specific object remote sensing extraction method based on visual Transformer, comprising the following steps:

[0009] Step S1: obtaining a remote sensing image to be processed, including a main source remote sensing image and an auxiliary source time sequence image;

[0010] Step S2: Segment the main source remote sensing image to obtain sub-blocks, extract multi-source parameter information for each sub-block, and calculate the sub-block comprehensive quality score s based on the multi-source parameter information;

[0011] Step S3: Classify the sub-blocks according to the overall quality score s, dividing them into high-quality areas, medium-quality areas, and low-quality areas;

[0012] Step S4: Apply corresponding processing strategies to the sub-blocks based on the quality classification results;

[0013] Step S5: Feed the sub-blocks of the high-quality region and the medium-quality region into the trained Transformer model in sequence, and output the classification probability map and uncertainty map of the specific object.

[0014] Step S6: In the order of sliding windows, stitch together the classification results of all sub-blocks to restore the complete map, and at the same time generate the full map uncertainty layer and the area to be verified label layer.

[0015] In step S2, the multi-source parameter information includes cloud cover ratio, aerosol optical thickness, and humidity state. In step S4, if the quality classification result is a high-quality region, the latest main-source remote sensing image is used as the input to the Transformer model.

[0016] In step S4, if the quality classification result is a medium quality region, a multi-temporal input mode is adopted, and the latest temporal phase map, the lowest cloud / AOD / no precipitation image, and the latest historical auxiliary source temporal phase map are selected and stitched together as the input of the Transformer model.

[0017] The latest time-phase image is the main source remote sensing image obtained in step S1. The lowest cloud / AOD / no precipitation image is the auxiliary source time-series image with the lowest total cloud cover and aerosol concentration and no precipitation in the set historical time. The latest historical auxiliary source time-phase image is the one with the most recent shooting time among the historical auxiliary source time-series images of the corresponding area of ​​the sub-block.

[0018] The sum of cloud cover and aerosol concentration is calculated by recording the cloud occlusion ratio + AOD value of each image, and selecting the auxiliary source time series image with the lowest cloud occlusion ratio + AOD value and no rainfall.

[0019] If the quality classification result is a low-quality area, obtain the auxiliary source time-series image corresponding to the location of the low-quality area in the historical auxiliary source time-phase image set, treat it as the main source remote sensing image, and perform the operations described in steps S2-S3 to obtain its quality classification result; if it meets the standard of a high-quality area or a medium-quality area, process it according to the corresponding processing strategy. If the quality classification results of all auxiliary source time-series images in the historical auxiliary source time-phase image set fail to meet the standard, skip this area and mark it as "awaiting manual review".

[0020] In a second aspect, the application provides a specific object remote sensing extraction system based on visual Transformer, comprising: a main source image acquisition module for collecting main source remote sensing images;

[0021] An auxiliary source multi-temporal image acquisition module is used for collecting auxiliary source time sequence images and storing them in a system database;

[0022] A data processing module is used for executing steps S2-S6;

[0023] A database is used for storing images, quality classification results in the processing process, and uncertainty maps;

[0024] A display module provides a WebGIS platform and a mobile terminal APP, and is used for displaying result maps and uncertain areas, and supporting manual review input;

[0025] A closed-loop feedback and model updating module automatically adjusts the sub-block quality score weight and the uncertainty threshold after receiving the manual review result.

[0026] Beneficial effects

[0027] In the application, by evaluating and screening environmental parameters such as cloud cover, aerosol optical depth (AOD), and rainfall, the system realizes quality control at the data input end, significantly reduces the misjudgment caused by cloud and rain, improves the reliability of the model input sample, and thus improves the accuracy of the entire target extraction process. Secondly, the application adopts an adaptive blocking mechanism, sets different overlap ratios and multi-temporal input modes according to high, medium and low quality areas, realizes fast and accurate processing in the high quality area, enhances the recognition robustness in the medium quality area through dense overlap and multi-temporal information, and manually reviews the low quality area, effectively avoiding invalid calculation and noise interference, so that the system runs more efficiently and the results are more reliable. In addition, the multi-source time sequence Transformer model fuses the optical and time sequence double branches, combines cross attention to realize comprehensive understanding of spatial texture and time sequence changes, not only enhances the capture of abandoned land edges, forms and dynamic state changes, but also shows strong robustness to light cloud cover or image interference environment. The uncertainty map and classification result output by the decoder can quantize the model confidence in detail, determine the pixel confidence according to the entropy value, and perform subsequent re-inference or manual review according to the uncertainty in the splicing stage, which theoretically ensures the boundary coherence and overall accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0028] Figure 1 A specific object remote sensing extraction method flowchart based on visual Transformer;

[0029] Figure 2Flowchart for implementing corresponding processing strategies for sub-blocks based on quality classification results;

[0030] Figure 3 This is a schematic diagram of the system in Example 2. Detailed Implementation

[0031] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings. It should be noted that these descriptions are for the purpose of aiding understanding the present invention, but do not constitute a limitation thereof. Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0032] Example 1

[0033] like Figure 1 As shown, in one embodiment, taking abandoned land as an example, the steps of the technical solution of the present invention are as follows:

[0034] Step S1: Obtain the remote sensing images to be processed, including the main source remote sensing image and the auxiliary source time series image.

[0035] In remote sensing target extraction, it is necessary to ensure high-quality input data. The data is collected and evaluated from two aspects: optical data quality and environmental impact measurement. The remote sensing images to be processed include main source remote sensing images and auxiliary source time-series images.

[0036] Main source remote sensing imagery: High-resolution satellite imagery (approximately 1 meter / pixel) is used to acquire clear textures and structural details of specific targets, preferably GF-2 imagery. This level of accuracy can be used to precisely identify the edges of abandoned land, the texture of weeds, and small-area land features, but it suffers from high acquisition costs and large storage requirements.

[0037] Secondary source temporal images: Preferably, Sentinel-2 imaging can be used to provide multi-temporal perspectives and enhance the ability to capture long-term change features. Its image size is smaller, it can be acquired frequently, and it is suitable for dynamic monitoring. By introducing secondary source temporal images and utilizing their high frequency of periodic imaging, occasional occlusion can be buffered, making the model more adaptable to sudden environmental changes (such as clouds and shadows) and improving output stability.

[0038] Step S2: Segment the main source remote sensing image to obtain sub-blocks, extract multi-source parameter information for each sub-block, and calculate the sub-block comprehensive quality score s based on the multi-source parameter information.

[0039] In remote sensing processing, a large area image is cut into several sub-blocks, usually 512x512 pixels, which are not limited here. Each sub-block represents a small square in the image, and the model will analyze, score, and segment these sub-blocks respectively. Extract multi-source parameter information from the main source remote sensing image, including cloud cover ratio, aerosol optical depth, and wetness state.

[0040] Cloud cover ratio: refers to the proportion of the number of pixels covered by clouds or cloud shadows in the total number of pixels in the remote sensing sub-block (such as 512x512 pixel blocks).

[0041] Using the main source remote sensing image in step S1, containing multiple spectral bands (such as red, green, blue, near-infrared, etc.), input these bands into a lightweight neural network model specifically designed to detect clouds, such as CD-CTFM, which combines CNN (extracts local texture) and Transformer (learns full-image features). The model will output the probability of each pixel being "cloud" (cloud probability value between 0-1).

[0042] If the cloud probability of a certain pixel is greater, it is considered to be greater probability of being covered by clouds or cloud shadows. Clouds can completely block ground objects, making it impossible to see specific objects below, such as abandoned land texture. By identifying these blocked areas, these areas can be excluded, and missing information can be excluded as a basis for judgment, thereby avoiding false recognition and improving overall accuracy.

[0043] Aerosol Optical Depth (AOD): Aerosol Optical Depth (AOD) is a core atmospheric parameter that measures the degree of blocking and scattering of sunlight by small particles in the atmosphere, such as dust, smoke, and haze. The higher the AOD value, the less light can be transmitted. NASA states that the AOD value represents the following:

[0044] AOD<0.1: Sky is very clear;

[0045] AOD≈1.0: Air is hazy, with low visibility;

[0046] AOD>3.0: Very severe haze, significantly blocking the sun.

[0047] In this embodiment, the AOD image is interpolated to the same grid (consistent with the size of the remote sensing sub-block), and the average AOD within each sub-block (such as 512x512 pixels) is calculated. Haze can reduce image contrast and blur land texture, and excluding haze areas can help the model more accurately identify ground objects and avoid misidentifying texture blur areas as abandoned land.

[0048] Wet condition flag: Analyze weather data to get the rainfall record in the past 2 days. For each sub-block, check the historical rainfall in this area. If the accumulated rainfall in this area (sub-block) in the past 2 days is ≥ 5mm, it is considered as a "wet affected area".

[0049] If a sub-block is marked as a wet affected area, the system will mark this sub-block as "wet" condition. Ground wetness changes light reflection, and excluding or separately processing wet sub-blocks can avoid misidentifying "non-abandoned" water-soaked soil as "abandoned land", thereby improving overall extraction accuracy and reducing interference.

[0050] Map the cloud cover rate, AOD, and wet condition flag to the 0-1 interval in proportion, calculate the comprehensive quality score s, and the calculation formula is as follows:

[0051] s = 1 - (w1 x cloud + w2 x AOD + w3 x rain)

[0052] Wherein, the specific parameters are as follows:

[0053] cloud: cloud cover ratio (e.g. 0.3 means 30% of the area is covered by clouds);

[0054] AOD: average haze value of the sub-block (e.g. 0.25);

[0055] rain: wet condition flag (0 / 1).

[0056] w1, w2, w3 are weights, which can be set as w1 > w2 > w3 according to the influence of each parameter on the result, and their sum is set to 1.

[0057] The closer s is to 1: the less environmental interference the sub-block has, the higher the observation quality, and it is suitable for direct processing. The closer s is to 0: the more environmental factors, the poorer the block quality, and it should be optimized or manually reviewed.

[0058] Step S3: Classify the sub-blocks according to the comprehensive quality score s into high quality area, medium quality area and low quality area.

[0059] The initial overlap between sub-blocks is set to 20% (about 102 pixels); this overlap can ensure that the model can see the context of the adjacent area at the boundary, improving the continuity and accuracy of splicing.

[0060] According to the comprehensive quality score s, the sub-blocks are divided into three categories:

[0061] High quality area: s ≥ 0.8 - clear information, high quality;

[0062] Medium quality area: 0.6 ≤ s < 0.8 - there is some obstruction or interference, but it can still be optimized;

[0063] Low quality area: s < 0.6 - interference is serious, not recommended for direct use, trigger compensation process.

[0064] Step S4: According to the quality classification result, the corresponding processing strategy is taken for the sub-block.

[0065] For the quality classification result of the sub-block, the corresponding processing strategy is different, and the processing strategy of the high quality area is directly used for subsequent processing of the main source remote sensing image. Figure 2 The processing flow according to the quality classification result is shown.

[0066] For the high quality area, the imaging quality is high, and the interference is less. At this time, the main source remote sensing image is directly used for subsequent processing.

[0067] Although the middle quality area has slight interference such as cloud, haze or rainfall, it still contains useful information. If it is directly skipped, important data may be ignored, so additional processing is needed to compensate for the deficiency. The specific operation process is as follows:

[0068] Overlap ratio promotion: from 20% to 50%, that is, about 256 pixels are shared by adjacent sub-blocks;

[0069] Multi-temporal input: select the latest time phase image + the lowest cloud / AOD / non-rainfall image + the latest historical auxiliary source time phase image for a total of three time phase splicing as input.

[0070] The latest time phase image refers to the main source remote sensing image obtained in the step S1, which provides the latest land cover state and texture information, and is the most critical ground feature image.

[0071] The lowest cloud / AOD / non-rainfall image refers to an auxiliary source time sequence image with the least cloud cover, the lowest aerosol concentration and no rainfall in a period of historical time, that is, the clearest image in the historical auxiliary source time sequence image of the region, which ensures the accuracy of model recognition. The historical time can be set according to the database storage condition, for example, it can be set to 7 days, 1 month, etc. The specific selection strategy is to use the cloud detection model to evaluate the cloud cover ratio and calculate the AOD value, record the cloud cover ratio + AOD value of each image, and select the image with the lowest cloud cover ratio + AOD score and no rainfall. If the image with the lowest cloud cover ratio + AOD score has rainfall, select the image with the second lowest cloud cover ratio + AOD score and no rainfall.

[0072] The latest historical auxiliary source time phase image refers to the image with the latest shooting time in the historical auxiliary source time sequence image of the region. In general, its update frequency is 3-7 days, which can serve as a backup to compensate for occasional interference.

[0073] The above operation steps can buffer the sudden cloud cover, poor image quality and other conditions of the latest image, while providing an alternative solution to improve the stability of model judgment.

[0074] Input merging: The multi-temporal images are spliced by channel into a 512x512x9(three temporal x 3 channel RGB) structure. The multi-temporal information improves the fault tolerance to cloud cover and haze, and the overlapping area increases the context, assisting in the positioning of the splicing edge.

[0075] For low-quality areas, trigger the compensation process. Based on the time sequence, obtain the auxiliary source time sequence image of the corresponding position of the low-quality area with the latest acquisition time in the historical auxiliary source phase set, and consider it as the main source remote sensing image. Perform the operations of steps S2-S3 on it to obtain its quality classification result. If it meets the standard of high-quality area or medium-quality area, the corresponding subsequent operation is performed with this image. If it cannot meet the standard, continue to obtain auxiliary source time sequence images of earlier time until the classification result meets the standard of high-quality area or medium-quality area. If the classification results of all auxiliary source time sequence images cannot meet the standard, skip this area and mark it as "to be manually reviewed".

[0076] At the same time, the system marks the sub-block as "to be manually reviewed" and displays it as a warning area on WebGIS or a mobile terminal; the interface allows users to view multi-temporal images and SAR auxiliary images and manually judge the abandoned land attribute to ensure the accuracy of the classification to the greatest extent. The manual input result is encoded into the system and stored as a trusted label. As shown in Table 1, the processing methods of different classification results are shown.

[0077] Table 1 Processing methods of different classification results

[0078] Step S5: The sub-blocks in the high-quality area and the medium-quality area are sequentially input into the trained Transformer model to output the classification probability map and the uncertainty map of the specific object.

[0079] The embodiment adopts a CNN+Transformer hybrid structure: first, use CNN to extract local texture features, process the main source remote sensing image and the auxiliary source time sequence image respectively, and then use Transformer (including ViT or SwinTransformer variant) to further fuse global information and use self-attention mechanism to obtain global context information, thereby forming a comprehensive understanding of the boundaries, structures and textures of the ground objects.

[0080] In the processing of the medium-quality area sub-block, the three images are concatenated to form a multi-temporal input; this can ensure that the model observes the current state, recent changes and clear references at the same time, improving the robustness to slight occlusion or light changes. In the decoding stage, the fused high-dimensional features are decoded and restored to a spatial feature map with a resolution of 512x512, realizing the dual-channel output of the classification probability map and the uncertainty map.

[0081] Classification probability map: the probability of each pixel corresponding to "abandoned" or "non-abandoned", preferably if the probability P of a certain pixel is greater than or equal to 0.5, it is judged as abandoned land.

[0082] Uncertainty map: calculate the pixel-level entropy using MC-Dropout or MC-Frequency Dropout to estimate the confidence level of the model's judgment. The higher the entropy, the less confident the model is in classifying the pixel.

[0083] Step S6: stitch all sub-block classification results into a complete map in sliding window order, while generating a full map uncertainty map layer and a to-be-reviewed area marker layer.

[0084] For each overlapping sub-block area, the system adopts different strategies according to the uncertainty u of the pixels in the area, sets a low uncertainty threshold t1 and a high uncertainty threshold t2, preferably the threshold t1 is 0.2, and the high uncertainty threshold is 0.5.

[0085] If u < t1 (low uncertainty): the model output is reliable, directly mean fusion sub-block overlapping area;

[0086] If t1 ≤ u ≤ t2 (moderate uncertainty): trigger secondary reasoning, select the result with higher confidence for stitching;

[0087] If u > t2 (high uncertainty): the model is uncertain, mark the area as to-be-reviewed and does not participate in automatic stitching.

[0088] Stitch all sub-block classification results into a complete map in sliding window order, while generating a full map uncertainty map layer and a to-be-reviewed area marker layer. The uncertainty map layer displays the confidence level of the model's classification of each pixel in the form of a heat map, and the to-be-reviewed marker layer clearly marks the high uncertainty area for manual review or subsequent resampling. To assist subsequent processing.

[0089] For the "to-be-reviewed" area, the system can automatically perform resampling or prompt manual review, and the review results are fed back into the system to adjust the uncertainty threshold and update the model. All reasoning results (classification, position, uncertainty, review marker) are stored in a spatial database (such as PostGIS) for convenient spatio-temporal query and management. A visualization platform is established to display the classification map, difficult areas and historical review, support on-site positioning, manual verification and data labeling.

[0090] Embodiment two

[0091] Figure 3 is a system schematic diagram provided by the embodiment two of the present application, specifically comprising:

[0092] The main source image acquisition module is used for collecting high-resolution optical images to obtain the latest texture and structure information of a target area. Preferably, GF-2 images are used.

[0093] The auxiliary source multi-temporal image acquisition module is used for collecting multi-temporal remote sensing images for updating and dynamic monitoring, and storing the images in the system database. Preferably, Sentinel-2 platform images are used.

[0094] The data processing module is used for performing the processing steps S2-S6.

[0095] The database module stores the collected remote sensing images, all classification results, uncertainty values and review marks in the spatial database.

[0096] The display module provides a WebGIS platform and a mobile APP interface to display the result map and the uncertain area, and supports manual review input.

[0097] The closed-loop feedback and model updating module automatically adjusts the sub-block quality score weight and the uncertainty threshold after receiving the manual review result, and periodically adds the review mark data to the training set to trigger model fine-tuning and improve future recognition accuracy.

[0098] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the division of the above functional units and modules is exemplified, and in actual application, the above functions can be completed by different functional units and modules according to needs, that is, the internal structure of the device is divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated in one processing unit, or each unit can be physically present separately, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or software. In addition, the specific names of the functional units and modules are only for easy distinction, and do not limit the protection scope of the application. The specific working process of the units and modules in the system can refer to the corresponding process in the foregoing method embodiments, which will not be repeated here.

[0099] In the above embodiments, the description of each embodiment has its own emphasis, and the parts not described or recorded in detail in a certain embodiment can be referred to the related description of other embodiments.

[0100] The embodiments of the application are described in detail above with reference to the drawings, but the application is not limited to the described embodiments. For those skilled in the art, various changes, modifications, replacements and variations of the embodiments can be made without departing from the principles and spirits of the application, and still fall within the protection scope of the application.

Claims

1. A method for specific object remote sensing extraction based on visual Transformer, characterized in that, The method comprises the following steps: Step S1: obtaining a to-be-processed remote sensing image, including a main source remote sensing image and auxiliary source time-series images; Step S2: segmenting the main source remote sensing image to obtain sub-blocks, extracting multi-source parameter information from each sub-block, and calculating a sub-block comprehensive quality score s according to the multi-source parameter information; Step S3: classifying the sub-blocks according to the comprehensive quality score s into high-quality areas, medium-quality areas and low-quality areas; Step S4: adopting a corresponding processing strategy for the sub-blocks according to the quality classification result; Step S5: sequentially inputting the sub-blocks in the high-quality areas and the medium-quality areas into a trained Transformer model to output a classification probability map and an uncertainty map of a specific object; Step S6: splicing the classification results of all the sub-blocks in a sliding window order to restore a complete map, and generating a full-map uncertainty map layer and a to-be-reviewed area marking layer.

2. The method of claim 1, wherein, In the step S2, the multi-source parameter information includes a cloud cover ratio, an aerosol optical depth and a wetness state.

3. The method of claim 1, wherein, In the step S4, if the quality classification result is a high-quality area, the main source remote sensing image in the step S1 is used as the input of the Transformer model.

4. The method of claim 1, wherein, In the step S4, if the quality classification result is a medium-quality area, a multi-time-phase input mode is adopted, three time phases of a latest time-phase image, a lowest cloud / AOD / non-rainfall image and a latest historical auxiliary source time-phase image are selected to splice as the input of the Transformer model.

5. The method of claim 4, wherein, The latest time-phase image is the main source remote sensing image obtained in the step S1, the lowest cloud / AOD / non-rainfall image is an auxiliary source time-series image with the lowest sum of cloud cover and aerosol concentration and without rainfall in the set historical time, and the latest historical auxiliary source time-phase image is an auxiliary source time-series image with the latest shooting time in the historical auxiliary source time-series images corresponding to the region of the sub-block.

6. The method of claim 5, wherein, The calculation method of the sum of the cloud cover and the aerosol concentration is to record the cloud cover ratio + AOD value of each image, and to select an auxiliary source time-series image with the lowest cloud cover ratio + AOD value and without rainfall.

7. The method of claim 1, wherein, If the quality classification result is a low-quality area, an auxiliary source time-series image with the latest collection time in the historical auxiliary source time-phase images corresponding to the position of the low-quality area is obtained, and is regarded as a main source image remote sensing image, and the operations in the steps S2-S3 are performed on the auxiliary source time-series image to obtain a quality classification result thereof; if the quality classification result reaches the standard of a high-quality area or a medium-quality area, the corresponding processing strategy is adopted for processing.

8. The method of claim 1, wherein, If the quality classification results of all the auxiliary source time-series images in the historical auxiliary source time-phase images cannot reach the standard, the region is skipped and is marked as a "to-be-manually-reviewed" state. 9.A visual Transformer-based specific object remote sensing extraction system, characterized in that, The system comprises: a main source image acquisition module, which is used to collect a main source remote sensing image; an auxiliary source multi-time-phase image acquisition module, which is used to collect auxiliary source time-series images and store the auxiliary source time-series images in a system database; a data processing module, which is used to perform the steps S2-S6 in claims 1-8; a database module, which is used to store images, quality classification results in a processing process and uncertainty maps; a display module, which provides a WebGIS platform and a mobile terminal APP, and is used to display a result map, an uncertainty area and support manual review input. Closed loop feedback and model updating module: automatically adjust the sub-block quality score weight and uncertainty threshold after receiving the manual review results.

Images