Satellite remote sensing image block classification water body identification method based on supervised learning algorithm

Abstract

The invention provides a satellite remote sensing image block classification water body recognition method based on a supervised learning algorithm, and the method comprises the steps: taking a Sentinel-1SAR image as basic data, and carrying out the image preprocessing, feature data preparation, sample point preparation, classification model construction and recognition, and image post-processing, thereby achieving the water body recognition based on a satellite remote sensing image. According to the method, VV, VH, angle, VV / VH and sum wavebands of the SAR image are used as training feature input of the classification model, so that the accuracy of a classification result is improved; a hexagon is used as a minimum block unit to carry out block calculation on a large-scale image, so that the edge effect during block calculation can be eliminated, the features of ground features are more accurately captured on a local scale, and the calculation amount of a single local classification model is reduced; noise pixels of non-water bodies in the classified images can be eliminated through the image post-processing step, and the noise influence of vegetation, buildings and the like is weakened. The method provided by the invention has the characteristic of high classification and identification precision, and is especially suitable for identification of water bodies in a large-scale range.

Description

Technical Field

[0001] The invention belongs to the field of remote sensing technology and surface water body identification, and relates to a satellite remote sensing image block classification water body identification method based on a supervised learning algorithm. Background Art

[0002] In recent years, extreme climate has occurred frequently under the influence of climate change and human activities, often causing large-scale floods, droughts and other disasters, which have brought great troubles to social and economic development and ecological and environmental protection. The characteristics of surface water bodies are an important aspect of characterizing water and drought disasters. Mastering the dynamic changes of surface water bodies is of great significance for flood and drought disaster monitoring and risk analysis and social development planning.

[0003] Remote sensing satellite images with high temporal and spatial resolution can be used to achieve rapid and automated water body identification and extraction. At present, commonly used remote sensing images mainly include optical satellite images and radar satellite images. Among them, SAR (Synthetic Aperture Radar) sensors are not affected by factors such as cloud cover and rainy weather, and can capture ground images around the clock. For example, the Sentinel-1 satellite can provide SAR images with a spatial resolution of 10m and a revisit period of 12 days. It can provide a data basis for large-scale surface water body identification and monitoring, and has broad application prospects in analyzing the long-term evolution of surface water bodies, flood inundation monitoring, and drought characteristics.

[0004] In addition, machine learning algorithms have high portability and computational efficiency, and show great potential in the identification of large-scale remote sensing image features. For example, in the analysis of flood disasters using Sentinel-1SAR images, Wang et al. used a support vector machine classification algorithm to explore the flood inundation situation in the middle and lower reaches of the Yangtze River during the flood season in 2020. Dong et al. constructed a transformer-based change detection model and drew a map of the water body evolution and flood inundation in Poyang Lake in 2020. Roth et al. used the TU Wien algorithm to process images and drew a dynamic map of water body changes in Pakistan from August to September 2022, further capturing and analyzing the maximum flood inundation range. Although the use of machine learning algorithms to process Sentinel-1SAR images is a common means of identifying large-scale water bodies, there is still a lack of water body identification methods that are both universal and highly accurate, especially in large-scale areas, where it is even more difficult to effectively identify water bodies. Summary of the invention

[0005] Since there is currently a lack of water body identification methods that are both universal and highly classified, in order to achieve accurate identification of water bodies over a large scale, the present invention provides a water body identification method based on satellite remote sensing image block classification based on a supervised learning algorithm to improve the accuracy of water body identification, especially to improve the accuracy of water body identification over a large scale.

[0006] In order to achieve the above-mentioned invention object, the technical solution adopted by the present invention is as follows:

[0007] The method for identifying water bodies by block classification of satellite remote sensing images based on supervised learning algorithm includes the following steps:

[0008] (1) Image preprocessing

[0009] Acquire the Sentinel-1SAR Level-1 GRD dataset of the area to be classified within the target period, recorded as the target SAR dataset, which includes several SAR images. Preprocess each SAR image in the target SAR dataset to obtain several preprocessed SAR images; the preprocessing includes band screening, edge noise correction, terrain correction, noise filtering and HAND index mask;

[0010] Band screening refers to removing the HH and HV band information in the SAR image and retaining the VV, VH and angle band information in the SAR image; edge noise correction refers to performing edge noise correction on the SAR image after the band screening; terrain correction refers to using the DEM (Digital Elevation Model) data of the area to be classified to mask the overlap and shadow areas in the SAR image after edge noise correction; noise filtering refers to using the Refined Lee filter processing algorithm to process the SAR image after terrain correction to reduce the clutter in the SAR image; HAND index masking refers to calculating the HAND index based on the SAR image after noise filtering, and masking out the area with a HAND index greater than 20;

[0011] (2) Feature data preparation

[0012] Based on each preprocessed SAR image, VV / VH and sum are calculated, VV / VH is the ratio of VV to VH, and sum is the sum of VV and VH; VV, VH, angle, VV / VH and sum of each preprocessed SAR image are used as feature data;

[0013] (3) Sample point preparation

[0014] The RGB channels were set with VV, VH, and VV / VH of the preprocessed SAR images, and each preprocessed SAR image was visualized. The visualized SAR images were mosaicked to obtain a base map. Several water body sample points and non-water body sample points were randomly selected from the base map, and all sample points were divided into a training data set and a validation data set.

[0015] (4) Construction and identification of classification models

[0016] The base map of step (3) is divided into blocks by using block calculation to generate a plurality of block units that completely cover the area to be classified, and a local classification model is constructed in each block unit, wherein the local classification model is constructed using a random forest algorithm as a classifier, and each local classification model is trained in its own block unit and its adjacent block units; during training, the feature data is input as feature data of the local classification model, and the data in the training data set is input as training data; each local classification model after training is used to identify water bodies and non-water bodies in the block unit where it is located, and obtain a preliminary recognition result image;

[0017] (5) Image post-processing

[0018] Acquire all Sentinel-2MSI optical images of the area to be classified in the target time period in which the percentage of cloud pixels is less than 10%, use a declouding function to remove the cloud pixels in each Sentinel-2MSI optical image, calculate the MNDWI index and EVI index of each Sentinel-2MSI optical image with cloud pixels removed, construct an MNDWI maximum value image and an EVI minimum value image, pixels with a maximum MNDWI value < 0 in the MNDWI maximum value image or pixels with a minimum EVI value > 0 in the EVI minimum value image are non-water body pixels, and the non-water body pixels in the preliminary recognition result image obtained in step (4) are removed in combination with the spatial coordinates of the non-water body pixels to obtain an image after non-water pixel filtering processing; then remove the water body pixels with the number of adjacent pixels less than a preset value in the image after non-water pixel filtering processing to obtain a water body distribution image of the area to be classified, thereby completing water body recognition.

[0019] In step (4) of the above technical solution, the number of decision trees of each local classification model is preferably 60 to 100.

[0020] In step (4) of the above technical solution, the block unit is preferably a regular hexagonal block unit. Further, the side length of the regular hexagonal block unit in step (4) is preferably 0.4° to 0.5°, where the unit "°" refers to the unit of the geographic coordinate system.

[0021] In step (3) of the above technical solution, all sample points can usually be divided into a training data set and a validation data set according to the division ratio of (7-8):(3-2).

[0022] In step (5) of the above technical solution, when removing water pixels whose number of adjacent pixels in the image after non-water pixel filtering is less than a preset value, the preset value of the number of adjacent pixels is determined according to actual application requirements. When it is necessary to obtain larger-scale water body distribution information in the area to be classified, a relatively large preset value can be selected. For example, when it is necessary to identify water bodies with an area greater than 0.3 hectares in the classified area, the preset value can be set to 30, that is, it is necessary to remove images of water bodies with an area not exceeding 0.3 hectares in the image after non-water pixel filtering.

[0023] The method of the present invention is particularly suitable for identifying water bodies in a large scale range and can produce a higher recognition accuracy. The large scale here is usually an area of ​​more than 3,000 square kilometers.

[0024] Compared with the prior art, the technical solution of the present invention produces the following beneficial technical effects:

[0025] The present invention provides a satellite remote sensing image block classification water body identification method based on a supervised learning algorithm. The method uses Sentinel-1SAR images as basic data, and after image preprocessing, feature data preparation, sample point preparation, classification model construction and identification, and image post-processing steps, water body identification based on satellite remote sensing images is realized. The present invention is beneficial to improving the accuracy of classification results by optimizing the VV, VH, angle, VV / VH and sum bands of SAR images as training feature inputs of the classification model. The high accuracy of the classification results also confirms the effectiveness of the training features, and can provide a reference for SAR image water body identification technology. The present invention uses hexagons as the minimum block unit to perform block calculation on large-scale images. The hexagonal block unit is beneficial to eliminating the edge effect that is easy to appear during block calculation, and can improve the consistency of classification and identification results; at the same time, the block calculation can more accurately capture the characteristics of the ground object at the local scale, thereby overcoming the problem of large-scale spatial heterogeneity; in addition, the block calculation can significantly reduce the calculation amount of a single local classification model and improve the speed of classification calculation. In the image post-processing step, the present invention uses the MNDWI maximum value or EVI minimum value image to effectively eliminate the noise pixels of non-water bodies in the classified image, which can effectively reduce the noise impact of vegetation, buildings, etc. Compared with the prior art, the method of the present invention has the characteristics of high classification and recognition accuracy. It is suitable for the identification of large-scale water bodies in areas with an area of ​​more than 3,000 square kilometers, and provides a reliable technical solution for disaster monitoring such as floods and droughts at the basin scale, national scale or intercontinental scale, which is of great significance to the development of disaster prevention and mitigation and policy formulation. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is a schematic diagram of the identification process of the method described in the present invention.

[0027] Figure 2 It is a flowchart of image preprocessing.

[0028] Figure 3 It is a flowchart of local classification model training and recognition using the trained local classification model.

[0029] Figure 4 It is the spatial position and terrain distribution map of the area to be classified in Example 1.

[0030] Figure 5 This is the distribution of sample points in Example 1, where blue sample points are water bodies and red sample points are non-water bodies.

[0031] Figure 6 This is the block division of the area to be classified in Example 1.

[0032] Figure 7 This is the classification and recognition result of Example 1, where blue represents water bodies and gray represents non-water bodies. DETAILED DESCRIPTION

[0033] The following examples further illustrate the satellite remote sensing image block classification water body identification method based on the supervised learning algorithm provided by the present invention. It is necessary to point out that the following examples are only used to further illustrate the present invention and cannot be understood as limiting the scope of protection of the present invention. The technical personnel in the relevant field make some non-essential improvements and adjustments to the present invention according to the above invention content for specific implementation, which still falls within the scope of protection of the invention.

[0034] Example 1

[0035] Take a certain area in Beijing, Tianjin and Hebei as the area to be classified. Its spatial location and terrain distribution are as follows: Figure 4 As shown in the figure, there are a large number of surface water bodies such as reservoirs and rivers in the area. Analyzing the distribution of water bodies in the area is of great significance for monitoring flood and drought disasters and water resource changes in the area. Specifically, this embodiment uses the method of the present invention to identify the surface water bodies in the area between June 13 and June 18, 2023. The identification process diagram is shown in the figure. Figure 1 shown.

[0036] 1. Image preprocessing

[0037] The Sentinel-1SAR Level-1 GRD dataset is used as the image source (access URL: https: / / developers.google.com / earth-engine / datasets / catalog / COPERNICUS_S1_GRD). Based on the original Sentinel-1SAR image, this dataset has been subjected to orbit calibration, GRD boundary noise removal, thermal noise removal, radiation calibration, terrain correction and other operations. The image band parameters of the Sentinel-1SAR Level-1 GRD dataset are shown in Table 1. The data in the Sentinel-1SAR Level-1 GRD dataset all have spatial coordinate information, and the SAR images obtained in the following processing process also have spatial coordinate information.

[0038] Table 1 Band parameter information of Sentinel-1SAR Level-1 GRD dataset images

[0039]

[0040]

[0041] The Sentinel-1SARLevel-1GRD dataset of the area to be classified in the target period (from June 13 to June 18, 2023) is obtained, which is recorded as the target SAR dataset. The target SAR dataset includes several SAR images. Each SAR image in the target SAR dataset is preprocessed on the Google Earth Engine cloud platform to obtain several preprocessed SAR images. The flowchart of image preprocessing is shown in the figure. Figure 2 As shown, the specific operations are as follows:

[0042] (1) Filter bands

[0043] The HH and HV band information in each SAR image is removed, and the VV, VH and angle band information in each SAR image is retained.

[0044] (2) Edge noise correction

[0045] Each SAR image processed in step (1) is subjected to edge noise correction to remove noise pixels at the edge of the image.

[0046] (3) Terrain correction

[0047] The DEM data of the area to be classified (available at: https: / / developers.google.com / earth-engine / datasets / catalog / CGIAR_SRTM90_V4) is used to mask the overlapped and shadowed areas in each SAR image processed in step (2) to avoid misidentification of water bodies caused by these areas.

[0048] (4) Noise filtering

[0049] The Refined Lee filter processing algorithm is used to perform noise filtering on each SAR image processed in step (3) to reduce the clutter in the SAR image and weaken the uneven brightness distribution in the SAR image caused by interference and scattering effects.

[0050] (5)HAND mask

[0051] The HAND (height above nearest drainage) index characterizes the relative height of the pixel from the local water system and can be applied to disaster mapping, landform classification and other fields. The HAND index is calculated based on each SAR image processed in step (4). Pixels with a HAND index greater than 20 are considered to be pixels that cannot be flooded, and the area is masked to obtain several pre-processed SAR images.

[0052] 2. Feature data preparation

[0053] Based on each preprocessed SAR image, VV / VH and sum are calculated. VV / VH is the ratio of VV to VH, and sum is the sum of VV and VH. The two bands of VV / VH and sum can clearly reflect the difference between water pixels and other pixels. VV, VH, angle, VV / VH and sum of each preprocessed SAR image are used as feature data for subsequent classification models.

[0054] 3. Sample point preparation

[0055] Since the surface of the water body is smooth, the backscatter coefficient of the water pixel in the SAR image is low. When visualizing the SAR image, the color of the water pixel will be darker than that of the non-water pixel. Based on this feature, the RGB channels are set with VV, VH, and VV / VH of the preprocessed SAR image, that is, VV is assigned to the red channel, VH is assigned to the green channel, and VV / VH is assigned to the blue channel. Each preprocessed SAR image is visualized, and the visualized SAR images are mosaicked to obtain the base map. 895 water body sample points and 1429 non-water body sample points are randomly selected from the base map. These 2324 sample points are divided into a training data set (1787 sample points) and a validation data set (537 sample points) according to a division ratio of 8:2. The distribution of the sample points is shown in the figure. Figure 5 As shown, Figure 5 In the figure, blue sample points are water bodies and red sample points are non-water bodies.

[0056] 4. Construction and identification of classification models

[0057] In the field of large-scale classification, block calculation (referring to dividing the target area into a number of small blocks and then calculating) can quickly and more accurately obtain classification results. On the one hand, block calculation can avoid the spatial heterogeneity problem in large-scale overall classification. The classification model trained by local features can more accurately capture the difference between water bodies and non-water bodies in the area and enhance the accuracy of the classification model; on the other hand, after the target area is divided into blocks, the calculation speed can be accelerated by parallel calculation and loop, thereby improving the efficiency of the overall calculation. The operation in this step is completed in Google EarthEngine. The parameters of each local classification model also involve: the minimum number of samples of leaf nodes (using the default value 1), the sample ratio used in each tree training (using the default value 0.5), and the random seed (using the default value 0). These parameters do not affect the accuracy of the classification model. Generally, the default values ​​are used and no special settings are required. In this step, the area to be classified is first divided into blocks for calculation, and then a local classification model is established. The local classification model is trained and the trained local classification model is used to identify water bodies and non-water bodies. The specific operations are as follows:

[0058] A regular hexagon with a side length of 0.5° (geographic coordinate system unit) is used as a block unit to generate 23 block units that completely cover the area to be classified, such as Figure 6As shown. A local classification model is constructed in each block unit. The local classification model is constructed using the random forest algorithm, which is a supervised classification algorithm, as a classifier. Each local classification model is trained in its own block unit and its adjacent block units. During training, the feature data in step 2 is input as the feature data of the local classification model, and the data in the training data set in step 3 is input as the training data. The number of decision trees for each local classification model is set to 60. Each trained local classification model is used to identify the water body and non-water body in the block unit where it is located, that is, 23 trained local classification models are used to identify the water body and non-water body in the block unit where each local classification model is located, and all the recognition results are mosaicked to obtain a preliminary recognition result image.

[0059] Figure 3 The process of training each local classification model and using the trained local classification model for recognition is shown. In each block unit ( Figure 3 Red hexagon) to build a local classification model, each local model is in its own block unit and the adjacent block unit of the block unit ( Figure 3 The dotted hexagon in the figure is used for training. This increases the sample size of the training and enhances the classification ability of the local classification model. The local classification model after training is used to train the block unit ( Figure 3 water bodies and non-water bodies (red hexagons) are identified.

[0060] 5. Image post-processing

[0061] Usually, adjacent objects are prone to the phenomenon of "different objects with the same spectrum", and different types of objects show the same spectral characteristics in remote sensing images. This phenomenon is reflected in the classification results, that is, non-water bodies are also identified as water bodies. In order to eliminate this problem of misidentification, optimize the classification results, and eliminate small patch noise, the preliminary recognition result image obtained in step 4 is post-processed, and the post-processing includes two steps: non-water body pixel filtering and smoothing.

[0062] All Sentinel-2MSI optical images with a resolution of 10m in the area to be classified within the target period are obtained, and all Sentinel-2MSI optical images with a cloud pixel percentage of less than 10% are screened out, and the cloud pixels in each Sentinel-2MSI optical image are removed by using a cloud removal function. Then the MNDWI index (water body index) and EVI index (enhanced vegetation index) of each Sentinel-2MSI optical image with cloud pixels removed are calculated to construct the MNDWI maximum value image and the EVI minimum value image. The data in the Sentinel-2MSI optical image are all with spatial coordinate information, and the constructed MNDWI maximum value image and EVI minimum value image are also with spatial coordinate information. The pixels with MNDWI maximum value <0 in the MNDWI maximum value image or the pixels with EVI minimum value >0 in the EVI minimum value image are non-water pixels. Combined with the spatial coordinate information of the non-water pixels, the non-water pixels in the preliminary recognition result image obtained in step 4 are removed to obtain the image after non-water pixel filtering.

[0063] Since this embodiment hopes to identify water bodies of larger scale in the area to be classified, the area is 0.3 hectares as the boundary, and water pixels with less than 30 adjacent pixels in the image obtained after non-water pixel filtering are removed, that is, only water bodies with an area larger than 0.3 hectares are retained, and small water patches with an area of ​​no more than 0.3 hectares in the identification results are removed to achieve a smoothing effect. In this way, the water distribution image of the area to be classified is obtained, that is, the water body identification is completed, and the results are as follows: Figure 7 As shown, Figure 7 Medium blue represents water bodies and gray represents non-water bodies.

[0064] 6. Accuracy Verification

[0065] In order to verify the accuracy and reliability of the classification and recognition results, the confusion matrix of the classification results is calculated using the sample points (537 sample points) in the validation data set. The calculation results are shown in Table 2.

[0066] Table 2 Confusion matrix

[0067]

[0068]

[0069] As shown in Table 2, among the 537 sample points in the validation data set, there are 190 water body sample points and 347 non-water body sample points. Among the 190 water body sample points, 24 were identified as non-water bodies after the classification and identification steps above, and among the 347 non-water body sample points, only 1 was identified as a water body after the classification and identification steps above. The overall classification accuracy (OA) of the classification and identification results is 0.95, achieving a high classification accuracy. At the same time, the Kappa index for testing the consistency of the classification results reached 0.90, indicating that the prediction results of the classification model are almost completely consistent with the actual classification results. In addition, the mapping accuracy (PA) of water bodies reached 0.87, indicating that 87% of the test samples were correctly classified, and the user accuracy (UA) of water bodies reached 0.94, indicating that 94% of the samples classified as water bodies were correctly classified. In general, the method of the present invention has shown a high accuracy in water body identification based on satellite remote sensing images.

Claims

1. A water body identification method based on satellite remote sensing image block classification based on supervised learning algorithm, characterized in that: The following steps are involved: (1) Image preprocessing Acquire the Sentinel-1SAR Level-1 GRD dataset of the area to be classified within the target period, recorded as the target SAR dataset, which includes several SAR images. Preprocess each SAR image in the target SAR dataset to obtain several preprocessed SAR images; the preprocessing includes band screening, edge noise correction, terrain correction, noise filtering and HAND index mask; Band screening refers to removing the HH and HV band information in the SAR image and retaining the VV, VH and angle band information in the SAR image; edge noise correction refers to performing edge noise correction on the SAR image after the band screening; terrain correction refers to using the DEM data of the area to be classified to mask the overlap and shadow areas in the SAR image after edge noise correction; noise filtering refers to using the Refined Lee filter processing algorithm to process the SAR image after terrain correction to reduce the clutter in the SAR image; HAND index masking refers to calculating the HAND index based on the SAR image after noise filtering, and masking out the area with a HAND index greater than 20; (2) Feature data preparation Based on each preprocessed SAR image, VV / VH and sum are calculated. VV / VH is the ratio of VV to VH, and sum is the sum of VV and VH. The VV, VH, angle, VV / VH and sum of each preprocessed SAR image are used as feature data; (3) Sample point preparation The RGB channels were set with VV, VH, and VV / VH of the preprocessed SAR images, and each preprocessed SAR image was visualized. The visualized SAR images were mosaicked to obtain a base map. Several water body sample points and non-water body sample points were randomly selected from the base map, and all sample points were divided into a training data set and a validation data set. (4) Construction and identification of classification models The base map of step (3) is divided into blocks by using block calculation to generate a plurality of block units that completely cover the area to be classified, and a local classification model is constructed in each block unit, wherein the local classification model is constructed using a random forest algorithm as a classifier, and each local classification model is trained in its own block unit and its adjacent block units; during training, the feature data is input as feature data of the local classification model, and the data in the training data set is input as training data; The trained local classification models are used to identify the water and non-water bodies in the block units to obtain the preliminary recognition result images. (5) Image post-processing Acquire all Sentinel-2MSI optical images of the area to be classified in the target time period in which the percentage of cloud pixels is less than 10%, use a declouding function to remove the cloud pixels in each Sentinel-2MSI optical image, calculate the MNDWI index and EVI index of each Sentinel-2MSI optical image from which the cloud pixels have been removed, construct an MNDWI maximum value image and an EVI minimum value image, pixels with a maximum MNDWI value of less than 0 in the MNDWI maximum value image or pixels with a minimum EVI value of more than 0 in the EVI minimum value image are non-water body pixels, and the non-water body pixels in the preliminary recognition result image obtained in step (4) are removed in combination with the spatial coordinates of the non-water body pixels to obtain an image after non-water pixel filtering processing; then remove the water body pixels with a number of adjacent pixels less than a preset value in the image after non-water pixel filtering processing to obtain a water body distribution image of the area to be classified, thereby completing water body recognition.

2. The method for identifying water bodies by block classification of satellite remote sensing images based on a supervised learning algorithm according to claim 1 is characterized in that: In step (4), the number of decision trees for each local classification model is 60 to 100.

3. The method for identifying water bodies by block classification of satellite remote sensing images based on a supervised learning algorithm according to claim 1 is characterized in that: In step (4), the block unit is a regular hexagonal block unit.

4. The method for identifying water bodies by block classification of satellite remote sensing images based on a supervised learning algorithm according to claim 3 is characterized in that: In step (4), the side length of the regular hexagonal block unit is 0.4° to 0.5°.

5. The method for identifying water bodies by block classification of satellite remote sensing images based on a supervised learning algorithm according to any one of claims 1 to 4, characterized in that: In step (3), all sample points are divided into a training data set and a validation data set according to the division ratio of (7-8):(3-2).

Images