Water body remote sensing recognition method based on hyperspectral data simulation and U-net

Abstract

The invention discloses a water body remote sensing recognition method based on hyperspectral data simulation and U-net. The water body remote sensing recognition method comprises the following steps: A, data collection and preprocessing; b, selecting and manufacturing a sample area; c, data simulation; d, derived parameters are calculated; e, manufacturing a training sample and a verification sample; f, training a U-net network model; g, multispectral water body detection; and H, precision evaluation. According to the method, a high-quality training sample is obtained through a hyperspectral remote sensing image, multispectral data of different sensors are simulated, the problem that the reusability of water body classification samples among different multispectral satellites and at different geographic distances and time is not sufficient is solved, the multispectral data Landsat8 and Sentinel2 are taken as examples, and the reusability of the water body classification samples is improved. It is proved that by means of the method, easily-confused ground objects such as building shadows and cloud shadows in the water body recognition process can be effectively removed, and on the basis of accurately extracting large-area rivers and lakes, the recognition precision of small lakes and ponds is improved.

Description

Technical Field

[0001] The invention relates to a method for extracting water bodies using satellite remote sensing images, and in particular to a method for performing remote sensing identification of water bodies using multispectral data using hyperspectral data simulation and U-net. Background Art

[0002] Surface water is an important factor affecting human health, climate change and biodiversity. It is particularly important to accurately and timely understand the extent and spatial distribution of water bodies. The detection of the spatial distribution of surface water is an important and valuable task that can be used for water resources monitoring and management, natural disaster and climate change analysis, and urban environmental management. With the continuous advancement of satellite technology, a variety of remote sensing data provided by spaceborne sensors have brought new opportunities for multi-temporal and spatial resolution water mapping on a global scale. However, automated water mapping is not an easy task. On the one hand, different components such as phytoplankton, suspended sediments, and other organic matter lead to the diversification of the spectral characteristics of water bodies, increasing the difficulty of accurately identifying water bodies. On the other hand, the confusion error caused by the similarity of water bodies to shadows and low-reflectivity surfaces has always been one of the main challenges facing water mapping.

[0003] Although some progress has been made in water body extraction methods, there are still many shortcomings. First, traditional threshold-based methods often rely on environmental conditions, such as illumination changes and climate factors, which leads to blurred water body boundaries and low extraction accuracy. In complex terrain and diverse vegetation coverage areas, water bodies have similar spectral characteristics to wet soil, shadows and other objects, which are easy to confuse and increase the difficulty of extraction. In addition, the spatial resolution of multispectral and hyperspectral images limits the detection of small water bodies, especially at fine-grained scales, making it difficult to accurately extract water bodies. The monitoring of small water bodies such as lakes and ponds is particularly affected. Secondly, many existing methods still rely on manual selection and adjustment of parameters, which not only increases the complexity of the operation, but also reduces the extraction efficiency and affects the consistency and repeatability of the results. In large-scale regional monitoring, manual intervention is particularly cumbersome and time-consuming.

[0004] Although advanced algorithms such as deep learning have performed well in water extraction, their need for a large number of high-quality samples remains a key challenge. The training of deep learning models usually relies on a large amount of accurately labeled training data. However, in the field of water extraction, it is not easy to obtain comprehensive and accurate water body labeled samples, especially under diverse environmental conditions and complex terrain. Data collection and labeling are time-consuming and costly. In addition, the water body characteristics in different regions vary greatly, resulting in the problem of insufficient generalization ability of the model when applied across regions. Retraining or tuning is required for specific regions, which further increases the difficulty of technical implementation.

[0005] Therefore, the prior art needs to be further improved and developed. Summary of the invention

[0006] In view of the above-mentioned deficiencies in the prior art, the present invention provides a water body remote sensing identification method based on hyperspectral data simulation and U-net. The method obtains high-quality training samples through hyperspectral remote sensing images, and simulates multispectral data of different sensors through data simulation, thereby solving the problem that the reusability of water body classification samples between different multispectral satellites, different geographical distances and times is still insufficient. Taking multispectral data Landsat8 and Sentinel2 as examples, it is proved that the method can effectively eliminate easily confusing objects in water body identification, such as building shadows and cloud shadows, and accurately extract large rivers and lakes, while also improving the accuracy of identifying small lakes and ponds. In order to solve the above-mentioned technical problems, the scheme of the present invention includes:

[0007] A water body remote sensing identification method based on hyperspectral data simulation and U-net comprises the following steps:

[0008] A. Data collection and preprocessing: This solution uses hyperspectral AVIRIS data and multispectral Landsat8 and Sentinel-2 data. Download AVIRIS-NG L2 surface reflectance data (380-2510nm, 425 band) from 2014 to 2022 through the AVIRIS Data Portal. Download Landsat 8-9 OLI / TIRS C2L2 (surface reflectance) data from the USGS Earth Explorer website, and download Sentinel-2 Level-2A data through the Copernicus Data Space Ecosystem. After data collection, remove invalid bands and areas from AVIRIS data, resample Landsat 8 and Sentinel-2, and perform reflectance calculations to obtain reflectance datasets.

[0009] B. Selection and production of sample areas: Based on the preprocessed AVIRIS hyperspectral data, first select images with rich land cover types, and preferably crop areas containing water bodies to ensure that there are both water bodies and non-water bodies (such as vegetation, bare soil, buildings, etc.) in the image. Subsequently, ArcGIS software is used to assist in the identification of water and non-water areas through visual interpretation combined with spectral characteristics, texture, color and other information. The identified water areas are labeled, classified as water bodies and non-water bodies, and saved as Shapefile format vector files. Finally, a binary raster image (1 represents water bodies and 0 represents non-water bodies) is generated through a vector-to-raster operation as a subsequent input label.

[0010] C. Data simulation: Data simulation is the process of synthesizing narrow-band hyperspectral data into wide-band multispectral data. By weighted combination of multiple hyperspectral bands, the band data of the multispectral sensor is generated. The present invention is based on a sample library constructed based on AVIRIS hyperspectral images. Through data simulation, simulated data of multispectral sensors (such as Landsat and Sentinel) in the visible to short-wave infrared range can be generated. Data simulation includes two parts: spectral interpolation and convolution: first, the central wavelength of the multispectral band is extracted, and the spectral response of the AVIRIS data is matched through interpolation processing; then, convolution calculation is performed to generate the multispectral image of the corresponding sensor.

[0011] D. Calculation of derived parameters: Five main derived parameters are used in this invention: Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), Modified Normalized Difference Water Index (mNDWI), Enhanced Vegetation Index (EVI), Automatic Water Extraction Index (AWEI), and nsh ).

[0012] E. Preparation of training samples and validation samples: Use the binary raster images of 0 and 1 (water body is 1, non-water body is 0) generated in step B as input labels, and the multispectral images simulated in step C and the images of the derived parameter combination obtained in step D as input data. Generate a sufficient number of training samples through image block processing and remote sensing data enhancement technology (such as rotation, mirroring, flipping, etc.) to enhance sample diversity and model generalization ability.

[0013] F. U-net network model training: The U-net network is used. The network contains 23 convolutional layers, which is essentially an encoder-decoder. Cross-validation is used in the network training process. In order to prevent overfitting of the network, zero padding and regularization strategies are used in the training process of the U-net network. After the network training is completed, the training model is saved.

[0014] G. Multispectral water body detection: Use the training model to predict the water body distribution of multispectral images, and finally obtain the water body detection results of multispectral images.

[0015] H. Accuracy evaluation. Qualitative and quantitative methods were used to verify the water body extraction effect. The binary water body image extraction results were superimposed and analyzed with the color composite image of the blue, red, and near-infrared bands. The results were compared with the manual interpretation results by visual inspection. The results show that water body features such as rivers and lakes can be completely extracted, and the shadows of mountains, buildings, clouds, and farmland interference can be effectively eliminated. This method is suitable for small-scale water areas with complex terrain. It is easy to operate and has strong versatility, and achieves high-precision water body extraction.

[0016] The method for remote sensing identification of water bodies, wherein the step A specifically includes: downloading the hyperspectral AVIRIS data through the AVIRIS Data Portal website, selecting AVIRIS-NG L2 Surface Reflectance, Facility Instrument Collection, V1 data set, and downloading data from 2014 to 2022. AVIRIS can record reflected radiation in the visible to short-wave infrared spectrum range from 380-2510nm, every 5nm interval. The AVIRIS-NG sensor has an instantaneous field of view of 1 milliradian and provides a highly correlated ground sampling distance from 20 meters to sub-meter range. After the data is downloaded, it is directly the surface reflectance value, and the value range is 0-1.

[0017] The water body remote sensing identification method, wherein the step A specifically further comprises:

[0018] For multispectral data, we take Landsat8 and Sentinel2 data as examples. First, download Landsat8 data from the USGS Earth Explorer website, select the date, range, and cloud cover to download, and select Landsat 8-9OLI / TIRS C2 L2 type in the data set to download the required data. Landsat8-9OLI / TIRS C2 L2 provides multispectral band surface reflectance and thermal infrared band surface temperature products, that is, multispectral data has been processed with atmospheric correction. The reflectance data is calculated by using the XML format metadata file provided in the data compression package to obtain the surface reflectance value. This solution chooses to download Sentinel2 from the Copernicus Data Space Ecosystem website, select the download date, range and cloud cover, and choose Level-2A data download. Level-2A data is the bottom-of-Atmosphere corrected reflectance data that has been atmospherically corrected and radiometrically calibrated. However, the pixel value of the image needs to be multiplied by 0.0001 (scale factor) to be converted into the surface reflectance value.

[0019] The method for remote sensing identification of water bodies, wherein the step B specifically includes: in order to produce sample data, first select AVIRIS-NG hyperspectral images with rich ground coverage types and containing a certain area of ​​water bodies, and then pre-process the images to remove the surrounding invalid data areas and invalid bands to ensure data validity. Then, the images are analyzed by visual interpretation methods with the aid of ArcGIS software, and the water bodies and non-water body areas are identified and classified in combination with spectral features. Finally, the interpretation results are saved as vector files (Shapefile format). For the water body vector spots within the sample range, use the vector to raster conversion tool to convert them into raster format, generate binary images with water body area values ​​of 1 and other area values ​​of 0, as input labels for training.

[0020] In the water body remote sensing identification method, the step C specifically includes: calculating and generating a spectral response function matching the multispectral sensor band according to the central wavelength of the AVIRIS hyperspectral data and the response function of the multispectral sensor. Specifically, by interpolating the spectral values ​​of multiple narrow bands of AVIRIS according to the response range of the multispectral band, integrating the spectral response corresponding to the multispectral band, the spectral response function of the multispectral data supported by the hyperspectral data is obtained.

[0021] In the water body remote sensing identification method, the step C specifically further includes: based on the obtained interpolated spectral response function, combining the hyperspectral remote sensing data, performing convolution operation on it to generate simulated multispectral remote sensing data. By performing band-by-band convolution processing on the interpolated spectral response function and the hyperspectral data, the hyperspectral data can be effectively converted into remote sensing data of a specified multispectral sensor. In this way, based on the constructed AVIRIS sample data, the number of samples that meet the target multispectral sensor in the visible light to short-wave infrared wavelength range is obtained through data simulation.

[0022] The water body remote sensing identification method, wherein the step E specifically includes: calculating the derived parameters NDVI, NDWI, mNDWI, EVI, AWEI for the simulated multispectral data nsh , and combine them with the simulated multispectral reflectance data through band fusion as the input images for training.

[0023] The water body remote sensing identification method, wherein the step E specifically also includes: for the multispectral remote sensing image generated by the hyperspectral data simulation and the derived parameter combination and the corresponding water body binarization result, a sufficient number of training samples are obtained by processing the image in blocks, and remote sensing data enhancement technology is applied at the same time. Specific operations include rotating, mirroring, flipping, etc. of the image to ensure the diversity of the samples, thereby enhancing the generalization ability of the model.

[0024] The water body remote sensing identification method, wherein the step F specifically includes: using a U-net network for model training, the network consists of 23 convolutional layers, which is essentially a coder-decoder structure. During the training process, a cross-validation method is used to divide the data set into 80% for training and 20% for validation. The learning rate of the network is set to 0.0001, and the Adam optimizer is used for optimization. In order to prevent overfitting, regularization and zero-filling strategies are applied during the U-net training process. After the training is completed, the trained model is finally saved for subsequent use.

[0025] The method for remote sensing identification of water bodies, wherein the step G specifically includes: calculating the derived parameters of the normalized vegetation index (NDVI), the normalized water index (NDWI), the modified normalized water index (mNDWI), the enhanced vegetation index (EVI), the automatic water extraction index (AWEI) by band operation on the pre-processed multispectral image. nsh ), through band fusion, the reflectance band data and derived parameters are combined together and input into the trained model as parameters to estimate the water body range.

[0026] In the water body remote sensing identification method, the step H specifically includes: using qualitative and quantitative methods to verify the water body extraction effect. The qualitative method is to overlay and analyze the binary water body image with the color composite image of the blue, red and near-infrared band combination, and visually check the extraction effect. The quantitative evaluation is evaluated by comparing with the manual interpretation results, and the indicators are evaluated by user accuracy, producer accuracy, overall accuracy and F1 score.

[0027] The present invention has the following advantages:

[0028] Satellite images, especially multispectral images, have been widely used in surface water range mapping. For different multispectral images, sufficient and high-quality training samples are particularly important for accurate mapping of surface water. However, the reusability of water classification samples between different multispectral satellites, different geographical distances and times is not sufficient, which indicates that different water body identification satellite images need to establish corresponding water sample databases. The present invention provides a water body remote sensing identification based on hyperspectral data simulation and U-net, which solves this limitation by using hyperspectral images to obtain high-quality training samples, and these samples can be used for water detection on different multispectral sensors. At the same time, U-Net is a convolutional neural network structure commonly used in image segmentation, and its encoder and decoder can accurately extract and restore the details in the image, and can more accurately segment and identify complex water body boundaries and morphology. While accurately identifying large-scale water body areas, it has good recognition effects on small-scale details, such as rivers, small lakes, etc. In general, the water body remote sensing identification method proposed in the present invention, combined with advanced sample acquisition methods and deep learning technology, can better make full use of the spectral information and spatial information of remote sensing images, and achieve high-precision and high-efficiency water body monitoring and classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is a schematic diagram of the flow of the remote sensing identification method in the present invention;

[0030] Figure 2 This is an example of hyperspectral AVIRIS raw data, cropped sample area image, and visual interpretation results and label files;

[0031] Figure 3 is an example of multispectral satellite remote sensing data from different sensors simulated by hyperspectral data;

[0032] Figure 4 is an example of derived parameters of a simulated Sentinel2 calculation;

[0033] Figure 5 This is an example of water detection results for multispectral data Sentinel2 using the trained model;

[0034] Figure 6 This is an example of water body detection results for multispectral data Landsat8 using the trained model;

[0035] Figure 7 This is an example of the effect of water body detection in the case of cloud shadow interference and small water bodies. DETAILED DESCRIPTION

[0036] The present invention provides a water body remote sensing identification method based on hyperspectral data simulation and U-net. In order to make the purpose, technical solution and effect of the present invention clearer and more specific, the present invention is further described in detail below. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0037] The present invention provides a water body remote sensing identification method based on hyperspectral data simulation and U-net, such as Figure 1 As shown, it includes the following steps:

[0038] Step 1: Data collection and preprocessing: This solution uses hyperspectral AVIRIS data and multispectral Landsat 8 and Sentinel-2 data. Download AVIRIS-NG L2 surface reflectance data (380-2510nm, 425 band) from 2014 to 2022 through the AVIRIS Data Portal. Download Landsat 8-9 OLI / TIRS C2L2 (surface reflectance) data from the USGS Earth Explorer website, and download Sentinel-2 Level-2A data through the Copernicus Data Space Ecosystem. After data collection, remove invalid bands and areas from AVIRIS data, resample Landsat 8 and Sentinel-2, and perform reflectance calculations to obtain reflectance datasets.

[0039] The second step is the selection and production of sample areas: Based on the preprocessed AVIRIS hyperspectral data, first select images with rich land cover types, and preferably crop areas containing water bodies to ensure that there are both water bodies and non-water bodies (such as vegetation, bare soil, buildings, etc.) in the image. Subsequently, ArcGIS software is used to assist in the identification of water and non-water areas through visual interpretation combined with spectral characteristics, texture, color and other information. The identified water areas are labeled, classified as water bodies and non-water bodies, and saved as Shapefile format vector files. Finally, a binary raster image (1 represents water bodies and 0 represents non-water bodies) is generated through a vector-to-raster operation as a subsequent input label.

[0040] The third step is data simulation: data simulation is the process of synthesizing narrow-band hyperspectral data into wide-band multispectral data. By weighted combination of multiple hyperspectral bands, the band data of the multispectral sensor is generated. The sample library constructed based on the AVIRIS hyperspectral image of the present invention can generate simulated data of multispectral sensors (such as Landsat, Sentinel) in the visible light to short-wave infrared range through data simulation. Data simulation includes two parts: spectral interpolation and convolution: first, the central wavelength of the multispectral band is extracted, and the spectral response of the AVIRIS data is matched by interpolation processing; then, convolution calculation is performed to generate the multispectral image of the corresponding sensor.

[0041] Step 4: Calculation of derived parameters: Five main derived parameters are used in this invention: Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), Modified Normalized Difference Water Index (mNDWI), Enhanced Vegetation Index (EVI), Automatic Water Extraction Index (AWEI), and nsh ).

[0042] Step 5: Preparation of training samples and validation samples: Use the binary raster images of 0 and 1 (water body is 1, non-water body is 0) generated in the second step as input labels, and the multispectral images and derived parameter combination images simulated in the third step as input data. Generate a sufficient number of training samples through image block processing and remote sensing data enhancement technology (such as rotation, mirroring, flipping, etc.) to enhance sample diversity and model generalization ability.

[0043] Step 6: U-net network model training: The U-net network is used. The network contains 23 convolutional layers, which is essentially an encoder-decoder. Cross-validation is used in the network training process. In order to prevent overfitting of the network, zero padding and regularization strategies are used in the training process of the U-net network. After the network training is completed, the training model is saved.

[0044] Step 7: Multispectral water body detection: Use the training model to predict the water body distribution of the multispectral image, and finally obtain the water body detection results of the multispectral image.

[0045] Step 8: Accuracy evaluation: Qualitative and quantitative methods are used to verify the water extraction effect. The binary water image is superimposed and analyzed with the color composite image of the blue, red, and near-infrared bands, and the evaluation is performed by visual inspection and comparison with the manual interpretation results. The results show that water features such as rivers and lakes can be completely extracted, and the shadows of mountains, buildings, clouds, and farmland can be effectively eliminated. This method is suitable for small-scale water areas with complex terrain. It is easy to operate and has strong versatility, and achieves high-precision water extraction.

[0046] In order to further describe the solution of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0047] The first step is data collection and preprocessing. The entire solution requires three types of remote sensing data, hyperspectral AVIRIS data and multispectral Landsat8 and Sentinel2 data. Data collection is mainly done through data download websites. (1) AVIRIS data download: Download through the AVIRIS Data Portal website, select AVIRIS-NG L2 Surface Reflectance (secondary reflectance data), Facility Instrument Collection, V1 dataset, and you can download data from 2014 to 2022. AVIRIS can record reflected radiation in the visible to shortwave infrared spectrum range from 380-2510nm, every 5nm interval, for a total of 425 bands. The AVIRIS-NG sensor has an instantaneous field of view of 1 milliradian, providing a highly correlated ground sampling distance from 20 meters to sub-meter range. (2) Download Landsat8 data. First, download Landsat8 data from the USGS Earth Explorer website. Landsat8 provides images since February 11, 2013, including "LandsatCollection 2 Level-1" (original images after geometric correction) or "Landsat Collection 2 Level-2" (surface reflectance images after atmospheric correction). This solution selects Landsat 8-9OLI / TIRS C2 L2 type ("Landsat Collection 2Level-2"). Select the download date, range, and cloud cover to download the required data. (3) Download Sentinel2 data. This solution chooses to download Sentinel2 from the Copernicus DataSpace Ecosystem website. It provides two types of data: ①Level-1C: including ground reflectance images, providing geographic correction but not atmospheric correction; ②Level-2A: surface reflectance images after atmospheric correction. Select the download date, range, and cloud cover. This solution chooses to download Level-2A data.

[0048] After the data is collected, it needs to be preprocessed. (1) After the hyperspectral AVIRIS data is downloaded, the surrounding invalid data areas are removed to ensure the validity of the data. At the same time, the invalid bands are removed, mainly the water vapor absorption bands. No processing is required for the image values. The image values ​​of the AVIRIS-NG L2 Surface Reflectance dataset are surface reflectance values ​​between 0 and 1. (2) After the multispectral Landsat8 dataset is downloaded, for the six bands we use, Band2 (blue), Band3 (green), Band4 (red), Band5 (near red NIR), Band6 (short wave infrared SWIR-1), and Band7 (short wave infrared SWIR-2), the spatial resolution of all bands is 30 meters. At the same time, the pixel value of the image needs to be multiplied by the coefficient minus the fixed value to convert the image value into a surface reflectance of 0-1. The specific conversion formula is shown in formula (1).

[0049] ρ i =0.000275×Val i -0.2 Formula (1)

[0050] Among them, ρ i is the reflectance value of each band, Val i is the image value of each band.

[0051] (3) After downloading the multispectral Sentinel2, we use the six bands Band 2 (blue), Band 3 (green), Band 4 (red), Band 8 (near-infrared NIR), Band 11 (short-wave infrared SWIR) and Band 12 (short-wave infrared SWIR).

[0052] Bands 2 to 3 to 4 to 8 have a spatial resolution of 10 meters, and Bands 11 to 12 have a spatial resolution of 20 meters. Bands 11 to 12 need to be resampled to 10 meters. At the same time, the pixel value of the image needs to be multiplied by 0.0001 (scale factor) to be converted into a surface reflectance value of 0-1.

[0053] The second step is the selection and production of sample areas. This step is mainly based on the pre-processed hyperspectral data AVIRIS. First, we need to select images: select AVIRIS-NG images with rich coverage types of objects. It is preferred to crop water areas with a certain area to ensure that the image contains both water bodies and non-water body types of objects (such as vegetation, bare soil, buildings, etc.), providing more diverse samples for subsequent classification and identification. Through this step, several images with rich coverage of land object types are obtained. Through visual interpretation, the objects in the image are identified, combined with the spectral characteristics, texture, color and other information of the image, and ArcGIS software is used to assist in the interpretation, focusing on identifying water and non-water areas, marking the identified water areas, and clearly marking the water and non-water classifications. After the visual interpretation is completed, the identified water bodies and non-water body areas are saved as vector format files, usually in Shapefile format (.shp). Finally, through the vector-to-raster operation, the data is converted into a binary raster image containing only 0 and 1, with water bodies as 1 and non-water bodies as 0, as subsequent input labels. Figure 2 , where green represents 1 water body and black represents 0 non-water body.

[0054] The third step is data simulation. Data simulation is the process of synthesizing narrow-band hyperspectral data into wide-band multispectral data. In this process, multiple hyperspectral data bands are weighted and synthesized into a multispectral data band. The present invention is based on the sample library constructed by AVIRIS hyperspectral image construction. Through the data simulation process, the sample library data of the multispectral sensor we want to use in the visible light to short-wave infrared band can be obtained. The data simulation expression is shown in formula (2).

[0055]

[0056] in, represents the spectral radiance of the synthetic multispectral image band, Indicates the spectral radiation of hyperspectral radiation. N MSI is the number of bands of the multispectral instrument (MSI), N HSI is the number of bands of the hyperspectral instrument (HSI), Δ j is the channel width of the hyperspectral instrument, c ij It is the spectral response function corresponding to the central wavelength of the multispectral sensor and the hyperspectral sensor.

[0057] The entire data simulation work is divided into two parts, spectral interpolation and convolution. The process of spectral interpolation is to convert AVIRIS hyperspectral image data into spectral responses of different multispectral data to match the band settings of different sensors. The specific steps are as follows: First, the central wavelengths of these bands need to be extracted for subsequent interpolation processing. According to the specific multispectral sensor (such as Landsat, Sentinel, etc.), the response function of each band is obtained. Use weighted average or other interpolation methods to make AVIRIS's hyperspectral data match the spectral response of the multispectral band. After interpolation processing, the generated spectral data will simulate the response of the multispectral sensor. The purpose of convolution is to obtain a multispectral image corresponding to the hyperspectral data. The principle is to convolve the Value (the present invention refers to the surface reflectance) on the hyperspectral image with the spectral response function obtained by spectral interpolation to obtain a multispectral image of the specified sensor. For examples of simulation of different sensor data, see Figure 3 .

[0058] The fourth step is to calculate the derived parameters. Derived indices are widely used in the field of remote sensing to extract information about land objects such as water bodies, vegetation, and soil. In this invention, five derived parameters are mainly used: Normalized Difference Vegetation Index (NDVI), Normalized Water Index (NDWI), Modified Normalized Water Index (mNDWI), Enhanced Vegetation Index (EVI), Automatic Water Extraction Index (AWEI), and the NDVI. nsh ), the calculation formulas of the derived parameters are as follows:

[0059] NDWI=(Green-NIR) / (Green+NIR) Formula (3)

[0060] mNDWI=(Green-SWIR1) / (Green+SWIR1) Formula (4)

[0061] NDVI=(NIR-Red) / (NIR+Red) Formula (5)

[0062] EVI=2.5*(NIR-Red) / (NIR+6*Red-7.5*Blue+1) Formula (6)

[0063]

[0064] Figure 4 is an example of derived parameters of a simulated Sentinel2 calculation;

[0065] Step 5: Production of training samples and validation samples. The binary raster image containing only 0 and 1 obtained in the second step is used as the input label, with water body as 1 and non-water body as 0. The multispectral image obtained by data simulation in the third step and the new multispectral remote sensing image generated by the combination of derived parameters are used as input images. By processing the image in blocks and applying remote sensing data enhancement technology, the specific operations include rotating, mirroring, flipping, etc., to obtain a sufficient number of training samples to ensure the diversity of samples and enhance the generalization ability of the model. Finally, 203 multispectral images and labels were obtained as training data, and 51 multispectral images and labels were used as validation data. The input sample images and sample labels were cropped according to the size of 256dpi×256dpi and data enhancement was performed. Finally, a total of 15,856 training samples and 3,964 validation samples were obtained.

[0066] Step 6, U-net network model training: The U-net network is used, which contains 23 convolutional layers, which are divided into upsampling and downsampling parts. The downsampling part contains five convolutional blocks. Each block consists of two convolutional layers, the convolution kernel size is 3×3, the stride is 1, and the activation function of the convolutional layer is the Relu function. A 2×2 maximum pooling operation with a stride of 2 is used to downsample at the end of each block except the last block. It is essentially an encoder-decoder. Cross-validation is used during network training, and the data set is divided into 80% for training and 20% for validation. In order to prevent overfitting of the network, zero padding and regularization strategies are used during the training of the U-net network. After the network training is completed, the training model is saved.

[0067] Step 7: Multispectral water body detection: Use the trained model to predict the water body in the area to be studied and obtain the water body results in the area to be studied. First, calculate the derived indices of Landsat 8 and Sentinel-2 multispectral data (such as NDWI, mNDWI, NDVI, EVI, AWEI nsh ), and combined with the reflectance band information for data fusion to generate a new image. This image will be used as the input data of the model. Subsequently, the generated image is input into the pre-trained water body detection model. The model predicts the water body distribution in the area to be studied by analyzing and learning the input data, and finally obtains the water body detection results of the multispectral image. Figure 5 and Figure 6 Demonstrated the water body detection effect of multispectral data Sentinel2 and Landsat8 data.

[0068] Step 8: Accuracy evaluation. Qualitative and quantitative methods are used to verify the water body extraction effect. The qualitative method is to overlay and analyze the binary water body image with the color composite image of the blue, red, and near-infrared bands, and visually check the extraction effect. The quantitative evaluation is carried out by comparing with the manual interpretation results. The indicators are user accuracy, producer accuracy, overall accuracy, and F1 score.

[0069]

[0070] Where TP, TN, FP and FN represent the number of true positives, true negatives, false positives and false negatives in the prediction, respectively. user is the user accuracy, A producer is the producer precision, A overall is the overall accuracy and F1 class is the F1 score.

[0071] The results show that water features such as rivers, lakes, and ponds can be completely extracted, and the interference of mountain shadows, buildings and their shadows, cloud shadows, and farmland in the flooding period can be effectively eliminated, avoiding the situation where some water bodies are mistakenly deleted during the elimination process. This method effectively solves the problem of high error rate in the remote sensing interpretation of water bodies, and is particularly suitable for areas with many small-scale lakes and ponds, complex objects, and images of flooding periods. It is easy to operate and has strong versatility. It can more accurately eliminate the influence of non-water objects and shadows, and achieve high-precision water extraction.

[0072] This is an example of qualitative evaluation of the accuracy of water body detection results using this method. It can be seen that this method effectively eliminates the influence of cloud shadows and has high extraction accuracy for small water bodies.

[0073] The accuracy evaluation of the water body detection experimental results shows that the water body extracted by the method of the present invention can achieve high accuracy and has high practicality.

[0074] Of course, the above description is only a preferred embodiment of the present invention, and the present invention is not limited to the above embodiments. It should be noted that all equivalent substitutions and obvious deformation forms made by any technician familiar with the field under the guidance of this specification fall within the essential scope of this specification and should be protected by the present invention.

Claims

1. A water body remote sensing identification method based on hyperspectral data simulation and U-net, characterized in that The steps include: A. Data collection and preprocessing: This scheme uses hyperspectral AVIRIS data and multispectral Landsat8 and Sentinel-2 data. After data collection, invalid bands and regions are removed from AVIRIS data, and calculation operations such as resampling and reflectance calculation are performed on Landsat 8 and Sentinel-2 to obtain reflectance datasets. B. Selection and production of sample areas: Based on the preprocessed AVIRIS hyperspectral data, first select images with rich land cover types, and preferably crop areas containing water bodies to ensure that there are both water bodies and non-water bodies (such as vegetation, bare soil, buildings, etc.) in the images; then, use ArcGIS software to assist in identifying water and non-water areas through visual interpretation combined with spectral characteristics, texture, color and other information; the identified water areas are labeled, classified into water bodies and non-water bodies, and saved as Shapefile format vector files; finally, through the vector-to-raster operation, a binary raster image (1 represents water bodies, 0 represents non-water bodies) is generated as the subsequent input label; C. Data simulation: Data simulation is the process of synthesizing narrow-band hyperspectral data into wide-band multispectral data, and generating band data of multispectral sensors by weighted combination of multiple hyperspectral bands. The present invention is based on the sample library constructed by the hyperspectral images of the AVIRIS sample library. Through data simulation, simulated data of multispectral sensors (such as Landsat and Sentinel) in the visible to short-wave infrared range can be generated. Data simulation includes two parts: spectral interpolation and convolution: first, the central wavelength of the multispectral band is extracted, and the spectral response of the AVIRIS data is matched by interpolation processing; then, convolution calculation is performed to generate the multispectral image of the corresponding sensor; D. Calculation of derived parameters: Five main derived parameters are used in this invention: Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), Modified Normalized Difference Water Index (mNDWI), Enhanced Vegetation Index (EVI), Automatic Water Extraction Index (AWEI), and nsh ); E. Preparation of training samples and validation samples: Use the binary raster images of 0 and 1 (water body is 1, non-water body is 0) generated in step B as input labels, the multispectral images simulated in step C and the images of the derived parameter combination obtained in step D as input data, and generate a sufficient number of training samples through image block processing and remote sensing data enhancement technology (such as rotation, mirroring, flipping, etc.) to enhance sample diversity and model generalization ability; F. U-net network model training: The U-net network is used. The network contains 23 convolutional layers, which is essentially an encoder-decoder. Cross-validation is used in the network training process. In order to prevent overfitting of the network, zero padding and regularization strategies are used in the training process of the U-net network. After the network training is completed, the training model is saved. G. Multispectral water body detection: Use the training model to predict the water body distribution of multispectral images, and finally obtain the water body detection results of multispectral images; H. Accuracy evaluation: Qualitative and quantitative methods are used to verify the water body extraction effect. The binary water body image is superimposed and analyzed with the color composite image composed of blue, red and near-infrared bands. The evaluation is carried out by comparing the visual inspection with the manual interpretation results. The results show that water body features such as rivers and lakes can be completely extracted, and the shadows of mountains, buildings, clouds and farmland can be effectively eliminated. This method is suitable for small-scale water areas with complex terrain. It is easy to operate and has strong versatility, and can achieve high-precision water body extraction.

2. A water body remote sensing identification method based on hyperspectral data simulation and U-net according to claim 1, characterized in that: In the step B, the specific method is as follows: In order to prepare sample data, we first select AVIRIS-NG hyperspectral images with rich land cover types and a certain area of ​​water bodies. Then, we preprocess the images to remove the surrounding invalid data areas and invalid bands to ensure data validity. Then, we analyze the images through visual interpretation methods with the help of ArcGIS software, and identify and classify water and non-water areas based on spectral features. Finally, we save the interpretation results as vector files (Shapefile format). For water body vector patches within the sample range, we use the vector to raster conversion tool to convert them into raster format, and generate binary images with water area values ​​of 1 and other area values ​​of 0 as input labels for training.

3. A water body remote sensing identification method based on hyperspectral data simulation and U-net according to claim 1, characterized in that: In the step C, the specific method is as follows: According to the central wavelength of AVIRIS hyperspectral data and the response function of the multispectral sensor, the spectral response function matching the multispectral sensor band is calculated and generated; specifically, by interpolating the spectral values ​​of multiple narrow bands of AVIRIS according to the response range of the multispectral band, the spectral response corresponding to the multispectral band is integrated to obtain the spectral response function of the multispectral data supported by the hyperspectral data. On the basis of the interpolated spectral response function, the hyperspectral remote sensing data is combined and convolution operation is performed on it to generate simulated multispectral remote sensing data. By performing band-by-band convolution processing on the interpolated spectral response function and the hyperspectral data, the hyperspectral data can be effectively converted into the remote sensing data of the specified multispectral sensor. In this way, based on the constructed AVIRIS sample data, the number of samples that meet the target multispectral sensor in the visible light to short-wave infrared wavelength range is obtained through data simulation.

4. A water body remote sensing identification method based on hyperspectral data simulation and U-net according to claim 1, characterized in that: In the step E, the specific method is as follows: For the simulated multispectral data, the derived parameters NDVI, NDWI, mNDWI, EVI, AWEI AWEI are calculated. nsh And through band fusion, they are combined with the simulated multispectral reflectance data as the input image for training. For the multispectral remote sensing images generated by hyperspectral data simulation and derived parameter combination and the corresponding water body binarization results, a sufficient number of training samples are obtained by dividing the images into blocks. At the same time, remote sensing data enhancement technology is applied. Specific operations include rotating, mirroring, flipping, etc. of the image to ensure the diversity of samples, thereby enhancing the generalization ability of the model.

5. A water body remote sensing identification method based on hyperspectral data simulation and U-net according to claim 1, characterized in that: In the step F, the specific method is as follows: The U-net network is used for model training. The network consists of 23 convolutional layers and is essentially an encoder-decoder structure. During the training process, the cross-validation method is used to divide the data set into 80% for training and 20% for validation. The learning rate of the network is set to 0.0001, and the Adam optimizer is used for optimization. In order to prevent overfitting, regularization and zero padding strategies are applied during the U-net training process. After the training is completed, the trained model is finally saved for subsequent use.

6. A water body remote sensing identification method based on hyperspectral data simulation and U-net according to claim 1, characterized in that: In step G, the specific method is as follows: The pre-processed multispectral images are used to calculate the derived parameters of the normalized vegetation index (NDVI), normalized water index (NDWI), modified normalized water index (mNDWI), enhanced vegetation index (EVI), automatic water extraction index (AWEI) through band operation. nsh ), through band fusion, the reflectance band data and derived parameters are combined together and input into the trained model as parameters to estimate the water body range.

7. A water body remote sensing identification method based on hyperspectral data simulation and U-net according to claim 1, characterized in that: In the step H, the specific method is as follows: Qualitative and quantitative methods were used to verify the water body extraction effect. The qualitative method was to overlay and analyze the binary water body image with the color composite image composed of blue, red and near-infrared bands, and visually check the extraction effect. The quantitative evaluation was carried out by comparing with the manual interpretation results, and the indicators were evaluated by user accuracy, producer accuracy, overall accuracy and F1 score.

Images