Large-scale impervious surface abundance extraction method, system, medium and equipment

Abstract

The invention discloses a large-scale impervious surface abundance extraction method, which belongs to the technical field of remote sensing image processing, and is characterized in that impervious surface abundance reference data is generated by combining object-oriented image analysis and grid superposition based on GF-1 2-meter fusion data, and training and test sample areas are selected by referring to GF-6WFV data; a high-quality sample set is generated for training and verification of a machine learning model; and through precision evaluation and visual analysis, an optimal impervious surface abundance extraction model is compared and selected, and high-precision large-scale impervious surface abundance extraction is realized. According to the method, the accuracy and detail identification capability of abundance extraction of the large-scale impervious surface are remarkably improved, impervious surfaces such as roads and buildings in a city can be identified and classified more accurately, and reliable data support is provided for city planning and management.

Description

Technical Field

[0001] The present invention belongs to the technical field of remote sensing image processing, and in particular relates to a method, system, medium and equipment for extracting abundance of large-scale impervious surfaces. Background Art

[0002] With the acceleration of urbanization, the proportion of urban impervious surfaces has increased significantly. Urban impervious surfaces include roads, buildings, parking lots and other surfaces that cannot penetrate water naturally. The increase of these surface covers has a significant impact on the urban environment and ecosystem, such as increasing flood risks, reducing water quality, and enhancing urban heat island effects. Therefore, accurate monitoring and assessment of the abundance of impervious surfaces is of great significance for urban planning and environmental protection.

[0003] Remote sensing data has become increasingly popular for studying impervious surfaces (ISA) due to its wide coverage and easy access. Currently, the main data sources for ISA extraction include SAR images, Landsat series, DMSP-OLS night lights, and MODIS remote sensing data. Due to the diversity of land cover types and the complexity of climate and phenological effects, these data face challenges such as registration errors, low image resolution, and quality affected by weather and lighting conditions during application, making it challenging to accurately map large-scale impervious surfaces. In the past few decades, many techniques have been developed to monitor impervious surfaces in satellite remote sensing images, including index methods, regression methods, and spectral mixture analysis (SMA). The index method is simple to calculate and is mostly used for impervious surface extraction in medium and low spatial resolution images. However, due to the limitation of spatial resolution, there are mixed pixels in the image on the one hand; on the other hand, the spectral similarity between objects can lead to misclassification of objects. Therefore, it is difficult to construct a more applicable index model for different images. The spectral mixture analysis method solves the problem of mixed pixels to a certain extent, but it also has the problem of underestimation and overestimation of impervious surfaces. In addition, due to the complex types of urban impervious surfaces and the limited resolution of remote sensing data, it is difficult to distinguish impervious surfaces from other types of land features, and the extraction steps are relatively complex, the amount of calculation is large, and the extraction results are easily affected by the accuracy of sample pixels. Therefore, the above methods are not reliable in large-scale ISA mapping. In contrast, machine learning and deep learning methods can better distinguish impervious surfaces from other types of surface cover due to their powerful feature learning capabilities and the ability to efficiently use remote sensing data. For example, methods such as support vector machines (SVM), decision trees, and random forests have been widely used and achieved remarkable results, especially the random forest algorithm.

[0004] The random forest algorithm is an integrated learning method that improves the predictive performance and stability of the model by constructing multiple decision trees and combining their results. The construction of a random forest mainly includes the following steps: First, multiple subsets are randomly selected from the original training data set with replacement through Bootstrap sampling. The samples in each subset may be repeated and there may be overlaps between different subsets. Then, for each subset, a decision tree (CART) is constructed. When the node is split, a part of the features is randomly selected to determine the best split. A user-specified number of predictor variables (Mtry) is used for each split to increase the differences between decision trees. A random forest consists of multiple independent decision trees trained on different subsets and different feature subsets. The number of trees in the forest is specified by the user (Ntree). Finally, for classification tasks, random forests decide the final classification result by voting on the prediction results of all decision trees; for regression tasks, the final regression value is obtained by averaging the prediction results of all decision trees, such as Figure 1 As shown in the figure. Random forest has strong noise resistance and efficient feature selection mechanism, can handle high-dimensional data and a large number of features, and has built-in feature importance evaluation function, and has good processing ability for missing values ​​and unbalanced data. These advantages make random forest show strong adaptability and efficiency in large-scale impervious surface abundance mapping, significantly improving the accuracy and reliability of mapping. However, the above methods often rely on traditional classification methods, such as pixel-based classification, when generating impervious surface training and test samples. These methods face many challenges when dealing with complex urban environments. There are various types of surface cover in urban environments, and there are a large number of unevenly distributed features such as high-rise buildings, roads, and green spaces. Traditional methods are difficult to capture these complexities, resulting in the generated samples often lacking in accuracy and representativeness. In particular, in the selection and annotation process of training data, there is a lot of manual intervention, which is easy to introduce subjective errors, further affecting the quality of samples. Low-quality sample data will lead to poor model training results, and the final extraction results will be unstable and inaccurate, which cannot meet the needs of actual applications. Summary of the invention

[0005] The purpose of the present invention is to overcome the deficiencies of the prior art and provide a large-scale impervious surface abundance extraction method, system, medium and equipment. The present invention uses GF-1 2-meter fusion data, object-oriented image analysis (OBIA) method and grid superposition to generate abundance reference data, and combines GF-6 wide-width (GF-6WFV) data to select training sample areas and test sample areas, thereby generating high-quality training and test sample data. Therefore, the present invention can improve the accuracy and representativeness of samples, overcome the deficiencies of traditional methods in sample generation, and ensure that the data foundation in the ISP extraction process is solid and reliable.

[0006] The present invention improves the mapping and classification accuracy through the combination of high-resolution data and multispectral data, overcomes the problem of confusion in the classification of objects, and realizes accurate large-scale ISA mapping. Specifically: Gaofen-1 (GF-1) and Gaofen-6 (GF-6) satellites effectively improve the extraction accuracy of large-scale ISA information by providing rich information content and wide imaging width respectively. The GF-1 satellite carries two 2-meter resolution panchromatic / 8-meter resolution multispectral (PMS) cameras, achieving an imaging width of more than 60 kilometers. PMS data includes four bands: blue, green, red, and near-infrared. Its high spatial resolution 2-meter panchromatic data and 8-meter multispectral data can capture the fine structure inside the city, which helps to distinguish different types of objects and classify land cover. The GF-6 satellite is equipped with a 16-meter multispectral medium-resolution wide-width (WFV) camera with an observation width of 800 kilometers, which is suitable for large-scale surface cover monitoring and change detection. In addition to the four common bands of blue, green, red and near-infrared, WFV data also adds two red-edge bands and purple and yellow bands, a total of eight bands, providing rich spectral information that helps distinguish different types of surface cover. The combination of GF-1 and GF-6 takes full advantage of the rich information content and wide imaging width to provide more accurate large-scale impervious surface information.

[0007] The first object of the present invention is to provide a large-scale impervious surface abundance extraction method, comprising the following steps:

[0008] Step 1, respectively orthorectifying the Gaofen-1 remote sensing satellite panchromatic / multispectral (GF-1PMS) data and the Gaofen-6 remote sensing satellite wide-band (GF-6WFV) data to obtain processed GF-1PMS data and GF-6WFV data;

[0009] Step 2: Register the processed GF-1PMS data and GF-6WFV data respectively to obtain the registered GF-1PMS data and GF-6WFV data, perform full color sharpening on the registered GF-1PMS data to obtain the GF-1 2-meter fused data;

[0010] Step 3: Generate a reference dataset of impervious surface abundance based on the 2-meter fused data of GF-1;

[0011] Step 4: Based on the impervious surface abundance reference dataset and the registered GF-6WFV data, select the training sample area and the test sample area, and randomly generate the training sample set and the test sample set;

[0012] Step 5: Input the training sample sets into multiple machine learning models for model training to obtain multiple trained impervious surface abundance (ISP) models;

[0013] Step 6: Use the test sample set to verify the accuracy of multiple trained ISP models using the accuracy evaluation index to obtain the accuracy evaluation index results of multiple models;

[0014] Step 7: Input the test sample area images into multiple trained ISP models to extract the abundance of impervious surfaces, compare the actual ground objects in the GF-6 true color synthetic image, and perform visual analysis on multiple extraction results;

[0015] Step 8: Compare the accuracy evaluation index results and visual analysis results of multiple ISP models and select the ISP model with the best effect;

[0016] Step 9: Input the target GF-6WFV remote sensing image into the best ISP model to extract the abundance of impervious surfaces and obtain the large-scale impervious surface abundance extraction results.

[0017] Preferably, step 2 comprises the following steps:

[0018] Step 2.1, the panchromatic 2-meter data in the processed Gaofen-1 remote sensing satellite panchromatic / multispectral (GF-1PMS) data and the processed Gaofen-6 remote sensing satellite 16-meter wideband (GF-6WFV) data are aligned to the multispectral 8-meter data in the processed GF-1PMS data using the nearest neighbor resampling method to obtain the aligned GF-1PMS panchromatic 2-meter data and GF-6WFV 16-meter data;

[0019] Step 2.2: Use the nearest neighbor diffusion panchromatic sharpening method to fuse the registered GF-1PMS panchromatic 2-meter data and the processed GF-1PMS multispectral 8-meter data to obtain the GF-1 2-meter fused data.

[0020] Preferably, step 3 comprises the following steps:

[0021] Step 3.1, use the object-oriented image analysis (OBIA) method to perform image segmentation and classification on the GF-1 2-meter fusion data to obtain classified data;

[0022] Step 3.2: Superimpose a grid with a cell side length of 16 meters on the classified data, and calculate the impervious surface abundance of each grid cell to generate impervious surface abundance reference data.

[0023] Further preferably, step 3.2 comprises the following steps:

[0024] 3.2.1. Generate a grid with a side length of 16 meters for each cell and superimpose it on the classified data;

[0025] 3.2.2. Calculate the sum of the areas of all 2-meter impervious surface elements in each grid cell to obtain the total area of ​​the impervious surface elements in each grid cell;

[0026] 3.2.3. Divide the total area of ​​impervious surface elements in each grid cell by the area of ​​each grid cell to generate reference data on impervious surface abundance.

[0027] Preferably, step 4 comprises the following steps:

[0028] Step 4.1, based on the impervious surface abundance reference dataset and the registered GF-6WFV data, select areas with more impervious surfaces as training sample areas and test sample areas;

[0029] Step 4.2: Extract attribute feature values ​​in the selected training sample area and test sample area respectively, and randomly generate training sample sets and test sample sets.

[0030] Preferably, in step 5, the multiple machine learning models are random forest (RF), support vector machine (SVM), decision tree (CART), XGBoost, LightGBM, CatBoost, U-net and Deeplabv3+.

[0031] Preferably, step 6 comprises the following steps:

[0032] Step 6.1: Input the test sample set into the trained ISP model to obtain the prediction results of the test sample;

[0033] Step 6.2: The prediction results of the test samples are verified by the accuracy evaluation index to obtain the accuracy verification results. The calculation formula of the accuracy evaluation index is:

[0034]

[0035]

[0036]

[0037]

[0038] Among them, MSE is the mean square error, RMSE is the root mean square error, MAE is the mean absolute error, R is the determination coefficient, N is the total number of samples, y i and i 'represent the measured value and predicted value of the i-th sample respectively, It represents the average value of the tested samples.

[0039] The second object of the present invention is to provide a large-scale impermeable surface abundance extraction system, which is realized by the above-mentioned extraction method, comprising:

[0040] The preprocessing module is used to perform orthorectification on the GF-1PMS data and the GF-6WFV data, and then perform registration, and perform panchromatic sharpening on the registered GF-1PMS data to generate the GF-1 2-meter fused data.

[0041] Reference data creation module, used to generate a reference dataset of impervious surface abundance based on GF-1 2-meter fusion data;

[0042] The sample creation module is used to combine the impervious surface abundance reference dataset and the registered GF-6WFV data, select the training sample area and the test sample area, and randomly generate the training sample set and the test sample set;

[0043] The model training module is used to input the training sample sets into multiple machine learning models for model training to obtain multiple trained ISP models;

[0044] The accuracy evaluation module is used to verify the accuracy of multiple trained ISP models using the accuracy evaluation index through the test sample set, and obtain the accuracy evaluation index results of multiple ISP models;

[0045] The visual analysis module is used to input the test sample area images into multiple trained ISP models to extract the abundance of impervious surfaces, compare the actual ground objects in the GF-6 true color synthetic images, and perform visual analysis on the extraction results;

[0046] The model selection module is used to compare the accuracy evaluation index results and visual analysis results of multiple ISP models and select the ISP model with the best effect;

[0047] The data extraction module is used to input the target GF-6WFV remote sensing image into the most effective ISP model to extract the abundance of impervious surfaces and obtain large-scale impervious surface abundance extraction results.

[0048] The third object of the present invention is to provide a computer-readable storage medium, which includes a stored computer program; wherein, when the computer program is running, it controls the device where the computer-readable storage medium is located to perform the steps of the large-scale impervious surface abundance extraction method as described above.

[0049] The fourth object of the present invention is to provide a terminal device, comprising a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor implements the steps of the large-scale impervious surface abundance extraction method as described above when executing the computer program.

[0050] Compared with the prior art, the present invention has the following beneficial effects:

[0051] (1) The present invention uses 2-meter fused data from GF-1, object-oriented image analysis (OBIA) and grid overlay to generate ISP reference data, and combines GF-6 wide-width (GF-6WFV) data to select training sample areas and test sample areas, thereby generating high-quality training and test sample data. This method can improve the accuracy and representativeness of samples, overcome the shortcomings of traditional methods in sample generation, and ensure that the data foundation in the ISP extraction process is solid and reliable;

[0052] (2) The present invention significantly improves the accuracy of large-scale impervious surface abundance extraction and detail recognition capabilities by utilizing WFV multispectral data from the Gaofen-6 remote sensing satellite and combining it with an advanced machine learning model RF method. It can more accurately identify and classify impervious surfaces such as roads and buildings in cities, providing more reliable data support for urban planning and management.

[0053] (3) The present invention uses an automated processing flow from sample selection to model training and result extraction, which reduces manual intervention and the possibility of human error, while greatly improving processing efficiency and consistency, making large-scale data processing and analysis more feasible and efficient;

[0054] (4) The present invention provides an advanced machine learning model - random forest method selection. The random forest algorithm performs well in dealing with the complexity and spatial heterogeneity of surface cover types and can capture the complex spatial variation of impervious surface cover, thereby significantly improving the accuracy and reliability of large-scale impervious surface abundance mapping;

[0055] (5) The present invention verifies the accuracy of the trained model by using a test sample set, adopts a variety of evaluation indicators (mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE), and determination coefficient (R-square)) for evaluation, and combines visual analysis to ensure the reliability of the results. This rigorous accuracy verification process ensures the high accuracy of the model extraction results and provides a reliable data foundation for subsequent applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 This is the schematic diagram of the existing random forest algorithm;

[0057] Figure 2A flowchart of a method for extracting large-scale sub-pixel impervious surface abundance using random forest regression and GF-6WFV data provided by an embodiment of the present invention;

[0058] Figure 3 The diagram is a diagram for selecting a training sample area and a test sample area in an embodiment of the present invention.

[0059] Figure 4 It is an ISP image of a test sample area generated by applying eight regression methods in an embodiment of the present invention;

[0060] in, Figure 4 (a) is SVM; Figure 4 (b) CART; Figure 4 (c) RF; Figure 4 (d) is XGBoost; Figure 4 (e) is CatBoost; Figure 4 (f) is LightGBM; Figure 4 (g) is U-net; Figure 4 (h) is Deeplabv3+;

[0061] Figure 5 A visual comparison analysis diagram of the results generated by the RF method and other methods in an embodiment of the present invention;

[0062] in, Figure 5 (a) is the GF-6 true color image; Figure 5 (b) RF; Figure 5 (c) is SVM; Figure 5 (d) CART; Figure 5 (e) is U-net; and Figure 5 (b) Figure 5 (c) Figure 5 (d) and Figure 5 (e) All correspond to Figure 5 (a) GF-6 true color image;

[0063] Figure 5 (f) is the GF-6 true color image; Figure 5 (g) is RF; Figure 5 (h) is XGBoost; Figure 5 (i) CatBoost; Figure 5 (j) is LightGBM; and 5(g), Figure 5 (h) Figure 5 (i) and Figure 5 (j) corresponds to Figure 5 (f) GF-6 true color image;

[0064] Figure 5(k) is the GF-6 true color image; Figure 5 (l) is RF; Figure 5 (m) is Deeplabv3+; and Figure 5 (l) and 5(m) correspond to Figure 5 (k) GF-6 true color image;

[0065] Figure 6 This is a large-scale ISP image generated by the RF method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0066] The following is a description of the embodiments of the present invention. Figures 2 to 6 , the technical solutions in the embodiments of the present invention are clearly and completely described. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0067] Example 1

[0068] like Figure 2 As shown, an embodiment of the present invention provides a method for extracting the abundance of a large-scale impervious surface, comprising the following steps:

[0069] Step 1, orthorectifying the Gaofen-1 remote sensing satellite panchromatic / multispectral (GF-1PMS) data and the Gaofen-6 remote sensing satellite wide-band (GF-6WFV) data respectively to obtain processed GF-1PMS data and GF-6WFV data; specifically, in order to reduce terrain undulation, the GF-6WFV and GF-1PMS images are orthorectified using the nearest neighbor resampling method based on the rational polynomial coefficients in the metadata file;

[0070] Step 2: respectively register the processed GF-1PMS data and GF-6WFV data to obtain the registered GF-1PMS data and GF-6WFV data, perform full color sharpening on the registered GF-1PMS data to obtain the GF-1 2-meter fused data; specifically, the following steps are included:

[0071] Step 2.1, the panchromatic 2-meter data (GF-1PMS panchromatic data (2m)) in the processed Gaofen-1 remote sensing satellite panchromatic / multispectral (GF-1PMS) data and the processed Gaofen-6 remote sensing satellite wide-band data (GF-6WFV data (16m)) are aligned to the multispectral 8-meter data (GF-1PMS multispectral data (8m)) in the processed GF-1PMS data by using the nearest neighbor resampling method to obtain the aligned GF-1PMS panchromatic 2-meter data and GF-6WFV 16-meter data to maintain the position accuracy;

[0072] Step 2.2: The panchromatic 2-meter data (GF-1PMS panchromatic data (2m)) in the registered GF-1PMS data and the multispectral 8-meter data (GF-1PMS multispectral data (8m)) in the processed GF-1PMS data are fused using the nearest neighbor diffusion panchromatic sharpening method to obtain the GF-1 2-meter fused data (GF-1 2m fused data).

[0073] Step 3: Generate an impervious surface abundance reference dataset based on the 2-meter fused data of GF-1, which specifically includes the following steps:

[0074] Step 3.1, using image segmentation software, using the OBIA method to segment the GF-1 2m fusion data map into classification data of six land cover types: impervious surface, woodland, grassland, water body, bare land, and cultivated land; the image classification software in the embodiment of the present invention is an existing image classification software, as long as it can realize the remote sensing image classification function, preferably eCognition software;

[0075] Step 3.2: Superimpose a grid with a cell grid side length of 16 meters on the classified data, and calculate the impervious surface abundance of each grid cell to generate the impervious surface abundance reference data, which specifically includes the following steps:

[0076] 3.2.1. Use image processing software to generate a grid with a cell grid side length of 16 meters according to the spatial resolution of GF-6WFV data, and superimpose it on the classified data; in the embodiment of the present invention, the image processing software is an existing image processing software, preferably ArcGIS software;

[0077] 3.2.2. Calculate the sum of the areas of all 2-meter impervious surface elements in each grid cell to obtain the total area of ​​the impervious surface elements in each grid cell;

[0078] 3.2.3. Divide the total area of ​​impervious surface elements in each grid cell by the area of ​​each grid cell (256 square meters) to generate an impervious surface abundance reference dataset.

[0079] Step 4: Based on the impervious surface abundance reference dataset and the registered GF-6WFV data, select the training sample area and the test sample area, and randomly generate the training sample set and the test sample set, which specifically includes the following steps:

[0080] Step 4.1: Based on the reference dataset of impervious surface abundance and the aligned GF-6WFV data, select areas with more impervious surfaces as training sample areas and test sample areas. The number of sample areas can be set according to the needs. In the embodiment of the present invention, 3 training sample areas and 1 test sample area are preferably selected. ArcGIS tools can be used to assist in the selection, such as Figure 3shown.

[0081] Step 4.2: Extract attribute feature values ​​in the selected training sample area and test sample area respectively, and randomly generate training sample sets and test sample sets.

[0082] Step 5: Input the training sample set into multiple machine learning models for model training to obtain multiple trained ISP models; the multiple machine learning models in this embodiment are random forest (RF), support vector machine (SVM), decision tree (CART), XGBoost, LightGBM, CatBoost, U-net and Deeplabv3+. Some of the above models need to call different numbers of threads according to the CPU performance of their own computers. This function is achieved through the n_jobs parameter;

[0083] Step 6: Use the test sample set to verify the accuracy of multiple trained ISP models using the accuracy evaluation index to obtain the accuracy evaluation index results of multiple ISP models, which specifically includes the following steps:

[0084] Step 6.1: Input the test sample set into the trained ISP model to obtain the prediction results of the test sample;

[0085] Step 6.2: The prediction results of the test samples are verified by the accuracy evaluation index to obtain the accuracy verification results. The calculation formula of the accuracy evaluation index is:

[0086]

[0087]

[0088]

[0089]

[0090] Among them, MSE is the mean square error, RMSE is the root mean square error, MAE is the mean absolute error, R is the determination coefficient, N is the total number of samples, y i and i 'represent the measured value and predicted value of the i-th sample respectively, It represents the average value of the tested samples;

[0091] When R-square is closer to 1, the MSE, RMSE and MAE values ​​are smaller, indicating that the model fit is better and the accuracy is higher.

[0092] Step 7: Input the test sample area images into multiple trained ISP models to extract the abundance of impervious surfaces, compare the actual ground objects in the GF-6 true color synthetic image, and perform visual analysis on multiple extraction results, such as Figure 4 , Figure 5 As shown;

[0093] Step 8: Compare the accuracy evaluation index results and visual analysis results of multiple ISP models to select the best machine learning model. When making a specific selection, the higher the model accuracy and the closer the visual analysis results are to the actual ground objects, the better the machine learning model. Figure 4 and Figure 5 The results show that the best RF model, and the number of trees (Ntree) of the RF model is 500, and the number of features (Mtry) is 3;

[0094] Step 9: Input the target GF-6 WFV remote sensing image into the best ISP model (RF model) to extract the abundance of impervious surfaces, and obtain the large-scale impervious surface abundance extraction result, such as Figure 6 shown.

[0095] Example 2

[0096] The embodiment of the present invention provides a large-scale impermeable surface abundance extraction system, which is implemented by the extraction method provided in Example 1, including:

[0097] The preprocessing module is used to perform orthorectification on the GF-1PMS data and the GF-6WFV data, and then perform registration, and perform panchromatic sharpening on the registered GF-1PMS data to generate the GF-1 2-meter fused data.

[0098] Reference data creation module, used to generate a reference dataset of impervious surface abundance based on GF-1 2-meter fusion data;

[0099] The sample creation module is used to combine the impervious surface abundance reference dataset and the registered GF-6WFV data, select the training sample area and the test sample area, and randomly generate the training sample set and the test sample set;

[0100] The model training module is used to input the training sample sets into multiple machine learning models for model training to obtain multiple trained ISP models;

[0101] The accuracy evaluation module is used to verify the accuracy of multiple trained ISP models using the accuracy evaluation index through the test sample set, and obtain the accuracy evaluation index results of multiple ISP models;

[0102] The visual analysis module is used to input the test sample area images into multiple trained ISP models to extract the abundance of impervious surfaces, compare the actual ground objects in the GF-6 true color synthetic images, and perform visual analysis on the extraction results;

[0103] The model selection module is used to compare the accuracy evaluation index results and visual analysis results of multiple ISP models and select the ISP model with the best effect;

[0104] The data extraction module is used to input the target GF-6WFV remote sensing image into the most effective ISP model to extract the abundance of impervious surfaces and obtain large-scale impervious surface abundance extraction results.

[0105] Example 3

[0106] An embodiment of the present invention provides a computer-readable storage medium, which includes a stored computer program; wherein, when the computer program is running, it controls the device where the computer-readable storage medium is located to execute the steps of the large-scale impervious surface abundance extraction method as described in the above-mentioned embodiment 1.

[0107] Example 4

[0108] An embodiment of the present invention provides a terminal device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, and the processor implements the steps of the large-scale impervious surface abundance extraction method as described in the above-mentioned embodiment 1 when executing the computer program.

[0109] Although the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the claims and their equivalents.

Claims

1. A method for extracting the abundance of large-scale impermeable surfaces, characterized in that: The following steps are involved: Step 1, respectively orthorectifying the Gaofen-1 remote sensing satellite panchromatic / multispectral data and the Gaofen-6 remote sensing satellite wide-band data to obtain processed Gaofen-1 remote sensing satellite panchromatic / multispectral data and Gaofen-6 remote sensing satellite wide-band data; Step 2, respectively align the processed GF-1 remote sensing satellite panchromatic / multispectral data and GF-6 remote sensing satellite wide-band data to obtain the aligned GF-1 remote sensing satellite panchromatic / multispectral data and GF-6 remote sensing satellite wide-band data, perform panchromatic sharpening on the aligned GF-1 remote sensing satellite panchromatic / multispectral data to obtain GF-1 2-meter fused data; Step 3: Generate a reference dataset of impervious surface abundance based on the 2-meter fused data of GF-1; Step 4: Based on the impervious surface abundance reference data set and the aligned GF-6 remote sensing satellite wide-band data, select the training sample area and the test sample area, and randomly generate the training sample set and the test sample set; Step 5: input the training sample sets into multiple machine learning models for model training to obtain multiple trained impervious surface abundance models; Step 6: Use the test sample set to verify the accuracy of multiple trained impervious surface abundance models using the accuracy evaluation index to obtain the accuracy evaluation index results of multiple models; Step 7: Input the test sample area images into multiple trained impervious surface abundance models to extract the impervious surface abundance, compare the actual ground objects in the GF-6 true color synthetic image, and perform visual analysis on multiple extraction results; Step 8, comparing the accuracy assessment index results and visual analysis results of multiple impervious surface abundance models, and selecting the impervious surface abundance model with the best effect; Step 9: Input the target GF-6 remote sensing satellite wide-band image into the best impervious surface abundance model to extract the impervious surface abundance, and obtain the large-scale impervious surface abundance extraction result.

2. The method for extracting the abundance of large-scale impermeable surfaces according to claim 1, characterized in that: Step 2 contains the following steps: Step 2.1, the panchromatic 2-meter data in the processed GF-1 remote sensing satellite panchromatic / multispectral data and the processed GF-6 remote sensing satellite 16-meter wideband data are aligned to the multispectral 8-meter data in the processed GF-1 remote sensing satellite panchromatic / multispectral data by using the nearest neighbor resampling method to obtain the aligned GF-1 panchromatic 2-meter data and GF-6 16-meter wideband data; Step 2.2: Use the nearest neighbor diffusion panchromatic sharpening method to fuse the registered GF-1 panchromatic 2-meter data and the processed GF-1 multispectral 8-meter data to obtain the GF-1 2-meter fused data.

3. The method for extracting the abundance of large-scale impermeable surfaces according to claim 1, characterized in that: Step 3 contains the following steps: Step 3.1, use the object-oriented image analysis method to perform image segmentation and classification on the GF-1 2-meter fusion data to obtain classified data; Step 3.2: Superimpose a grid with a cell side length of 16 meters on the classified data, and calculate the impervious surface abundance of each grid cell to generate impervious surface abundance reference data.

4. The method for extracting large-scale impervious surface abundance according to claim 3, characterized in that: Step 3.2 includes the following steps: 3.2.

1. Generate a grid with a side length of 16 meters for each cell and superimpose it on the classified data; 3.2.

2. Calculate the sum of the areas of all 2-meter impervious surface elements in each grid cell to obtain the total area of ​​the impervious surface elements in each grid cell; 3.2.

3. Divide the total area of ​​impervious surface elements in each grid cell by the area of ​​each grid cell to generate reference data on impervious surface abundance.

5. The method for extracting large-scale impermeable surface abundance according to claim 1, characterized in that: Step 4 includes the following steps: Step 4.1, based on the impervious surface abundance reference dataset and the aligned GF-6 remote sensing satellite wide-band data, select areas with more impervious surfaces as training sample areas and test sample areas; Step 4.2: Extract attribute feature values ​​in the selected training sample area and test sample area respectively, and randomly generate training sample sets and test sample sets.

6. The method for extracting large-scale impermeable surface abundance according to claim 1, characterized in that: In step 5, the multiple machine learning models are random forest, support vector machine, decision tree, XGBoost, LightGBM, CatBoost, U-net and Deeplabv3+.

7. The method for extracting large-scale impervious surface abundance according to claim 1, characterized in that: Step 6 contains the following steps: Step 6.1, input the test sample set into the trained impervious surface abundance model to obtain the prediction results of the test samples; Step 6.2: The prediction results of the test samples are verified by the accuracy evaluation index to obtain the accuracy verification results. The calculation formula of the accuracy evaluation index is: Among them, MSE is the mean square error, RMSE is the root mean square error, MAE is the mean absolute error, R is the determination coefficient, N is the total number of samples, y i and i 'represent the measured value and predicted value of the i-th sample, respectively. It represents the average value of the tested samples.

8. A large-scale impermeable surface abundance extraction system, implemented by the method according to any one of claims 1 to 7, characterized in that: include: The preprocessing module is used to perform orthorectification on the panchromatic / multispectral data of the GF-1 remote sensing satellite and the wide-band data of the GF-6 remote sensing satellite, and then perform registration, and perform panchromatic sharpening on the registered GF-1 remote sensing satellite panchromatic / multispectral data to generate GF-1 2-meter fused data; Reference data creation module, used to generate a reference dataset of impervious surface abundance based on GF-1 2-meter fusion data; The sample creation module is used to combine the impervious surface abundance reference dataset and the aligned GF-6 remote sensing satellite wide-band data, select the training sample area and the test sample area, and randomly generate the training sample set and the test sample set; A model training module is used to input the training sample sets into multiple machine learning models for model training to obtain multiple trained impervious surface abundance models; The accuracy evaluation module is used to verify the accuracy of multiple trained impervious surface abundance models using the accuracy evaluation index through the test sample set, and obtain the accuracy evaluation index results of multiple impervious surface abundance models; The visual analysis module is used to input the test sample area images into multiple trained impervious surface abundance models to extract the impervious surface abundance, compare the actual ground objects in the GF-6 true color synthetic image, and perform visual analysis on the extraction results; Model selection module, used to compare the accuracy assessment index results and visual analysis results of multiple impervious surface abundance models and select the best impervious surface abundance model; The data extraction module is used to input the target Gaofen-6 remote sensing satellite wide-band image into the most effective impervious surface abundance model to extract the impervious surface abundance and obtain the large-scale impervious surface abundance extraction result.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored computer program; wherein, when the computer program is run, it controls the device where the computer-readable storage medium is located to execute the steps of the large-scale impervious surface abundance extraction method as described in any one of claims 1-7.

10. A terminal device, characterized in that: The method comprises a processor, a memory and a computer program stored in the memory and configured to be executed by the processor, wherein the processor implements the steps of the large-scale impervious surface abundance extraction method as described in any one of claims 1 to 7 when executing the computer program.

Images