SAR (Synthetic Aperture Radar) satellite remote sensing surface water extraction method for region with complex African climate

Abstract

The invention discloses an SAR satellite remote sensing surface water extraction method for an African climate complex region, and the method comprises the steps: carrying out the preprocessing of satellite image data, and making sample point label data; making a shadow mask file, and removing the mountain shadow on the satellite image data; calculating a gray level co-occurrence matrix to generate texture features; establishing a multi-dimensional feature space, and screening the feature space; building a semi-supervised cooperative training model by using two random forest classifiers, and introducing label-free data to assist in model training; and predicting the global surface water distribution of the research area. According to the method, the SAR remote sensing image is combined with the semi-supervised collaborative random forest model, the manpower and time cost of actual sample labeling is greatly reduced, the surface water monitoring efficiency is improved, the problem of visible light remote sensing data missing in regions with complex climates can be solved by using microwave remote sensing data, and the method is suitable for popularization and application. And timely and accurate information is provided for water resource management in the African region.

Description

Technical Field

[0001] The present invention relates to the technical field of surface water identification using microwave remote sensing technology, and specifically relates to a method for extracting surface water from SAR satellite remote sensing for complex climate regions in Africa. Background Art

[0002] Due to advantages such as wide coverage, spatial continuity, and short revisit period, satellite remote sensing technology has played an important role in water resource monitoring. However, for optical sensors, they are vulnerable to the influence of clouds, rainfall, fog, etc., and it is difficult to provide continuous high-quality and low cloud fraction image data. On the contrary, synthetic aperture radar (SAR) can obtain high-resolution surface information under various meteorological conditions by actively emitting microwave signals and receiving reflected waves. It has the characteristics of all-weather, all-day, and high penetration, and plays an important role in the research of regular water resource monitoring.

[0003] Common SAR image water body extraction methods include threshold segmentation method, machine learning method, and deep learning method, etc. The threshold segmentation method mainly finds the optimal threshold through the reflection feature differences between the water body and the background. However, due to the complex changes of water bodies in large areas, it is difficult to determine a unified threshold, and this method ignores factors such as ground object texture, shape, and water body proportion, which ultimately easily leads to problems such as incomplete water body extraction results and misclassification. With the rapid development of machine learning and deep learning methods, many related image classification methods have been effectively applied to water body extraction and mapping. Although the deep learning method can effectively extract the semantic features of remote sensing images, it requires a large amount of labeled data. The acquisition of labeled data in large areas requires a lot of time and manpower, resulting in usually difficult to obtain sample data that meets complex deep learning models. The machine learning method can extract surface water according to the characteristics of different objects, through artificially designed classification features, using a small amount of sample data. To solve the above problems, many scholars have proposed semi-supervised learning methods, that is, using a small amount of labeled data and a large amount of unlabeled data for joint training, which can not only make up for the problem of insufficient real sample data, but also effectively utilize unlabeled data to improve the classification accuracy. Among them, the semi-supervised learning method based on disagreement can obtain better results than other semi-supervised classification algorithms by using multiple classifiers and making full use of unlabeled data.

[0004] Combining SAR satellite images with a semi-supervised collaborative random forest model and adding geographical, texture and other environmental features unique to complex climate regions in Africa can more efficiently monitor and analyze surface water resources, providing a scientific basis for local ecological protection, management and sustainable development. Summary of the Invention

[0005] Objective of the present invention: To provide a method for extracting surface water from SAR satellite remote sensing in complex climate regions of Africa. This method combines SAR satellite images and a semi-supervised collaborative random forest model to achieve precise extraction of surface water in complex climate regions of Africa.

[0006] To achieve the above functions, the present invention designs a method for extracting surface water from SAR satellite remote sensing in complex climate regions of Africa. For the target complex climate region in Africa, the following steps A - G are executed to complete the prediction of the surface water distribution:

[0007] Step A: Obtain the GRD image data of the Sentinel-1 satellite in the complex climate region of Africa. Preprocess the image data of the Sentinel-1 satellite and convert the image data of the Sentinel-1 satellite into a backscatter coefficient map with a resolution of 10m, and then proceed to Step B;

[0008] Step B: Use Arcgis software to visually interpret the image data obtained in Step B and create sample point label data for water bodies and non-water bodies, and then proceed to Step C;

[0009] Step C: Use the SRTM DEM software to create a shadow mask file with a slope greater than 5°. Remove the mountain shadows from the image data obtained in Step C, and then proceed to Step D;

[0010] Step D: Calculate the gray-level co-occurrence matrix using the image data obtained in Step D to capture the texture features in the complex climate region of Africa, and then proceed to Step E;

[0011] Step E: Combine the texture features obtained in Step D, the polarization features of the Sentinel-1 satellite image data, the water body index in the complex climate region of Africa, and the terrain features to establish a multi-dimensional feature space. Use the Boruta algorithm to screen the feature space and use shadow features to screen important features, and then proceed to Step F;

[0012] Step F: Divide the screened important features into two different feature subsets and input them into two random forest classifiers respectively. Use the random forest classifiers to build a semi-supervised collaborative training model and introduce unlabeled sample data in the complex climate region of Africa to assist in model training, and then proceed to Step G;

[0013] Step G: Train the semi-supervised collaborative training model built in Step F to obtain a trained semi-supervised collaborative training model. Based on the semi-supervised collaborative training model, predict the surface water distribution in the complex climate region of Africa.

[0014] Beneficial effects: Compared with the prior art, the advantages of the present invention include:

[0015] 1. The present invention combines SAR satellite images and a semi-supervised collaborative random forest model to propose a method for extracting surface water that integrates a multi-dimensional feature space, solving the problem of surface water monitoring in the African region with its complex climate and geographical environment and lack of visible light images.

[0016] 2. The present invention uses Sentinel-1 satellite image data to quickly obtain the distribution of surface water in large-scale areas with complex climates, greatly reducing the manpower and time costs of actual annotation samples, overcoming the problem of a small sample size, and improving the efficiency of surface water monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 is a flowchart of the SAR satellite remote sensing surface water extraction method for African regions with complex climates according to an embodiment of the present invention;

[0018] Figure 2 is a structural diagram of the semi-supervised collaborative random forest model according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0019] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention and should not be used to limit the protection scope of the present invention.

[0020] The SAR satellite remote sensing surface water extraction method for African regions with complex climates provided by the embodiment of the present invention, with reference to Figure 1 , for the target African region with a complex climate, perform the following steps A - G to complete the prediction of the surface water distribution:

[0021] Step A: Obtain the GRD (ground range detected) image data of the IW mode of the Sentinel-1 satellite for the African region with a complex climate, and the polarization mode is VH + V; preprocess the image data of the Sentinel-1 satellite, convert the image data of the Sentinel-1 satellite into a backscattering coefficient map with a resolution of 10 m, and enter Step B;

[0022] In Step A, the improved Lee filter is used for coherent wave filtering of the image data of the Sentinel-1 satellite, further improving the image quality and better suppressing the inherent speckle noise. The preprocessing of the image data of the Sentinel-1 satellite includes: orbit correction, thermal noise removal, radiometric calibration, speckle filtering, terrain correction, and decibel conversion to remove the inherent noise of the image.

[0023] Step B: Use Arcgis software to visually interpret the image data obtained in Step B to produce sample point label data for water bodies and non-water bodies, and enter Step C;

[0024] Step C: Use the SRTM DEM software to create a shadow mask file with a slope greater than 5°. Remove the mountain shadows from the image data obtained in Step C to eliminate the influence of radiation distortion of the SAR satellite image data, and proceed to Step D;

[0025] Step D: Calculate the gray-level co-occurrence matrix using the image data obtained in Step D to capture the texture features of the complex climate regions in Africa, and proceed to Step E;

[0026] The texture features include: Mean, Variance, Correlation, Homogeneity, Contrast, Entropy, Dissimilarity, and Second moment.

[0027] Step E: Combine the texture features obtained in Step D, the polarization features of the Sentinel-1 satellite image data, and the water body index and terrain features of the complex climate regions in Africa to establish a multi-dimensional feature space, enrich the feature space of the SAR satellite image, and use the Boruta algorithm to screen the feature space. Use the shadow features to screen important features, reduce the model training cost, and improve the training accuracy, and proceed to Step F;

[0028] Step F: Divide the screened important features into two different feature subsets, increase the difference between the features in the feature subsets, and obtain a more complete water body distribution; input the two different feature subsets into two random forest classifiers respectively, use the random forest classifiers to build a semi-supervised co-training model, and introduce the unlabeled sample data of the complex climate regions in Africa to assist in model training, and then proceed to Step G;

[0029] Refer to Figure 2 , the semi-supervised co-training model in Step F includes two random forest classifiers. The specific steps for building and training the semi-supervised co-training model are as follows:

[0030] Step F1: Randomly divide the training dataset into two equal-sized labeled sample sets L 1 、L 2 , and each group is used to train an independent random forest classifier, denoted as random forest classifier RF 1 、random forest classifier RF 2 ;

[0031] Step F2: Introduce a randomly generated unlabeled sample set U, and use the trained random forest classifier RF 1 to label the elements in the unlabeled sample set U, and expand the labeled data to the labeled sample set L 2In it, as the new training set L ′ 2 ; Through the trained random forest classifier RF 2 Mark the elements in the unlabeled sample set U, and expand the labeled data to the labeled sample set L 1 In it, as the new training set L ′ 1 ;

[0032] Step F3: Retrain the random forest classifier RF using the constructed new training set 1 、random forest classifier RF 2 , and repeat this process until all unlabeled samples are randomly forest classified by RF 1 、random forest classifier RF 2 Consistently classified as water body or non - water body, then the learning process of unlabeled samples ends;

[0033] Step F4: After the iterative training ends, if there are still differences in the classification results of the random forest classifier RF 1 、random forest classifier RF 2 , according to the weights assigned in the random forest classifier RF 1 、random forest classifier RF 2 , select the result with a larger weight ratio as the final classification result of the sample.

[0034] In step F, the randomly generated unlabeled sample set U is incremented by 1000, the optimal sample threshold of the constructed semi - supervised co - training model is found, and the training is repeated until a certain evaluation index converges.

[0035] In step F, a semi - supervised co - training model is constructed based on the random forest classifier. Each random forest classifier is set with 100 decision trees. The data set of each tree is randomly generated by Bootstrap. During the construction of each tree, the number of features considered at each node is the square root of all the numbers, which helps to reduce the correlation of the model and improve the generalization ability of the model; In one embodiment, the iteration is set to 50 times and the threshold is set to 0.7.

[0036] Step G: Train the semi - supervised co - training model built in step F to obtain a trained semi - supervised co - training model, and based on the semi - supervised co - training model, predict the surface water distribution in the complex climate region of Africa.

[0037] The above has described the embodiments of the present invention in detail with reference to the accompanying drawings. However, the present invention is not limited to the above - described embodiments. Within the scope of knowledge possessed by those of ordinary skill in the art, various changes can be made without departing from the gist of the present invention.

Claims

1. A SAR satellite remote sensing surface water extraction method for African climate complex areas, characterized in that: For the target African climate complex region, perform the following steps A to G to complete the prediction of surface water distribution: Step A: Obtain the GRD image data of the Sentinel-1 satellite in the African climatic complex area, pre-process the image data of the Sentinel-1 satellite, convert the image data of the Sentinel-1 satellite into a 10m resolution backscatter coefficient map, and then proceed to step B; Step B: Use ArcGIS software to visually interpret the image data obtained in step B, generate sample point label data for water bodies and non-water bodies, and proceed to step C; Step C: Use SRTM DEM software to create a shadow mask file with a slope greater than 5°, remove the mountain shadow on the image data obtained in step C, and proceed to step D; Step D: Calculate the gray level co-occurrence matrix using the image data obtained in step D to capture the texture characteristics of the African climate complex area, and then proceed to step E; Step E: Combine the texture features obtained in step D, the polarization features of the image data of the Sentinel-1 satellite, and the water index and terrain features of the African climate complex area to establish a multidimensional feature space, and use the Boruta algorithm to screen the feature space, use the shadow features to screen important features, and then enter step F; Step F: Divide the selected important features into two different feature subsets, input them into two random forest classifiers respectively, use the random forest classifier to build a semi-supervised collaborative training model, and introduce unlabeled sample data from the African climate complex area to assist in model training, and then proceed to step G; Step G: Train the semi-supervised collaborative training model constructed in step F to obtain a trained semi-supervised collaborative training model, and predict the surface water distribution in the African climate complex area based on the semi-supervised collaborative training model.

2. The SAR satellite remote sensing surface water extraction method for African climate complex areas according to claim 1 is characterized in that: In step A, the improved Lee filter is used to perform coherent wave filtering on the image data of the Sentinel-1 satellite.

3. The SAR satellite remote sensing surface water extraction method for African climate complex areas according to claim 1 is characterized in that: The preprocessing of the image data of the Sentinel-1 satellite in step A includes: orbit correction, thermal noise removal, radiation calibration, coherent speckle filtering, terrain correction, and decibel conversion.

4. The SAR satellite remote sensing surface water extraction method for African climate complex areas according to claim 1 is characterized in that: The texture features in step D include: mean, variance, correlation, synergy, contrast, information entropy, dissimilarity and second-order moment.

5. The SAR satellite remote sensing surface water extraction method for African climate complex areas according to claim 1 is characterized in that: The semi-supervised collaborative training model in step F includes two random forest classifiers. The specific steps of building and training the semi-supervised collaborative training model are as follows: Step F1: randomly divide the training data set into two groups of labeled sample sets L1 and L2 of equal size, each of which is used to train an independent random forest classifier, denoted as random forest classifier RF1 and random forest classifier RF2 respectively; Step F2: Introduce a randomly generated unlabeled sample set U, label the elements in the unlabeled sample set U through the trained random forest classifier RF1, and expand the labeled data to the labeled sample set L2 as the new training set L ′ 2. Use the trained random forest classifier RF2 to label the elements in the unlabeled sample set U, and expand the labeled data to the labeled sample set L1 as the new training set L ′ 1; Step F3: retrain the random forest classifier RF1 and the random forest classifier RF2 using the newly constructed training set, and repeat this process until all unlabeled samples are consistently classified as water bodies or non-water bodies by the random forest classifier RF1 and the random forest classifier RF2, and the unlabeled sample learning process ends; Step F4: After the iterative training is completed, if there are still differences in the classification results of the random forest classifier RF1 and the random forest classifier RF2, according to the weights allocated in the random forest classifier RF1 and the random forest classifier RF2, the result with a larger weight ratio is selected as the final classification result of the sample.

6. The SAR satellite remote sensing surface water extraction method for African climate complex areas according to claim 5 is characterized in that: In step F, the randomly generated unlabeled sample set U is incremented by 1000 to find the optimal sample threshold in the constructed semi-supervised collaborative training model.

7. The SAR satellite remote sensing surface water extraction method for African climate complex areas according to claim 5 is characterized in that: In step F, a semi-supervised collaborative training model is constructed based on the random forest classifier. Each random forest classifier is set with 100 decision trees. The data set of each tree is randomly generated by Bootstrap. In the construction process of each tree, the number of features considered for each node is the square root of all numbers.

Images