Terrain correction method considering influence of multiple time phases and complex terrains

Abstract

The invention relates to a terrain correction method considering influence of multiple time phases and complex terrain, and the method comprises the steps: obtaining original remote sensing images and DEM data of a target terrain region under multiple time phases, and carrying out the preprocessing of the original remote sensing images and DEM data, and obtaining the preprocessed original remote sensing images and DEM data; performing terrain correction on the preprocessed original remote sensing image based on the preprocessed DEM data to obtain a reference image; inputting the preprocessed original remote sensing image and the corresponding DEM data into the initial terrain correction model, and performing iterative training based on a reference image corresponding to the original remote sensing image to obtain a target terrain correction model; the initial terrain correction model is constructed based on a generative adversarial network; and obtaining a to-be-corrected remote sensing image and the corresponding DEM data, performing preprocessing, and inputting the preprocessed to-be-corrected image and the corresponding DEM data into the target terrain correction model to obtain a target remote sensing image. The terrain correction effect and the processing efficiency of the remote sensing image can be improved.

Description

Technical Field

[0001] This application relates to the technical field of remote sensing image processing, and particularly to a terrain correction method considering the influence of multi-temporal and complex terrain. Background Art

[0002] Currently, traditional terrain correction methods usually rely on complex physical models, such as C-correction and Minnaert models. When calculating the terrain effects in complex terrain areas, these methods rely on accurate terrain parameters (such as slope, aspect, etc.) and require a large amount of calculations. Due to these cumbersome calculation processes and high calculation costs, especially when dealing with large-scale and complex terrain areas, the efficiency is low. In addition, traditional methods are usually designed for single-temporal and single-sensor images and cannot effectively meet the correction requirements of multi-temporal and multi-sensor images, especially in dynamic scenarios (such as seasonal changes and spatio-temporal changes), and the correction effect is not good.

[0003] Especially in the application of remote sensing images, traditional methods show poor adaptability when dealing with the illumination angle changes of multi-temporal images and the radiation differences between different sensor data. These limitations seriously restrict the wide application of remote sensing images in the fields of ecological monitoring, environmental assessment, etc. Summary of the Invention

[0004] Based on the above description, this application provides a terrain correction method considering the influence of multi-temporal and complex terrain, which can improve the terrain correction effect and processing efficiency of remote sensing images.

[0005] This application provides a terrain correction method considering the influence of multi-temporal and complex terrain, including:

[0006] Obtain the original remote sensing images and DEM data of the target terrain area at multiple time phases and perform preprocessing to obtain the preprocessed original remote sensing images and DEM data;

[0007] Based on the preprocessed DEM data, perform terrain correction on the preprocessed original remote sensing images to obtain reference images;

[0008] Input the preprocessed original remote sensing images and the corresponding DEM data into an initial terrain correction model, and perform iterative training based on the reference images corresponding to the original remote sensing images to obtain a target terrain correction model; the initial terrain correction model is constructed based on a generative adversarial network;

[0009] Obtain the remote sensing image to be corrected and the corresponding DEM data and perform preprocessing, and input the preprocessed image to be corrected and the corresponding DEM data into the target terrain correction model to obtain the target remote sensing image.

[0010] In one or more embodiments, the initial terrain correction model includes a generator and a discriminator. The generator adopts a U-Net structure, and the discriminator adopts a convolutional neural network;

[0011] The training process of the initial terrain correction model includes:

[0012] Fuse the preprocessed original remote sensing image and the corresponding DEM data and input them into the generator to output a generated image, and input the generated image and the corresponding reference image into the discriminator;

[0013] Based on the generated image and the corresponding reference image, the discriminator is optimized using a first cross-entropy loss function, and the generator is optimized using a second cross-entropy loss function, a pixel loss function, a structural similarity loss function, and a visual perception loss function.

[0014] In one or more embodiments, the first cross-entropy loss function determines the cross-entropy loss of the discriminator based on the predicted probability values output by the discriminator for the generated image and the corresponding reference image respectively.

[0015] In one or more embodiments, the second cross-entropy loss function determines the cross-entropy loss of the generator based on the predicted probability value output by the discriminator for the generated image;

[0016] The pixel loss function uses the L1 norm to calculate the pixel difference between the generated image and the corresponding reference image at each same position to determine the pixel loss of the generator;

[0017] The structural similarity loss function uses a sliding window method to calculate the SSIM value of the generated image and the corresponding reference image in each local area to determine the structural similarity loss of the generator;

[0018] The visual perception loss function uses the L1 norm and determines the visual perception loss of the generator based on the feature maps extracted from the generated image and the corresponding reference image in the same convolutional layer by a pre-trained VGG network.

[0019] In one or more embodiments, obtaining the original remote sensing images and DEM data of the target terrain area at multiple time phases and performing preprocessing includes:

[0020] Under the condition that the cloud coverage rate of the target terrain area is less than a preset coverage rate threshold, obtain the original remote sensing images and the corresponding DEM data at multiple time phases; the preset coverage rate threshold is between 4% and 6%;

[0021] Perform radiometric calibration and geometric correction on the original remote sensing image and the corresponding DEM data, and perform georegistration on the processed original remote sensing image and the corresponding DEM data.

[0022] In one or more embodiments, based on the preprocessed DEM data, perform topographic correction on the preprocessed original remote sensing image to obtain a reference image, including:

[0023] Reproject the preprocessed DEM data so that it is in the same coordinate system as the preprocessed original remote sensing image;

[0024] Obtain the solar zenith angle and solar azimuth angle of each pixel from the preprocessed original remote sensing image, and calculate the slope and aspect of each pixel based on the corresponding reprojected DEM data;

[0025] Calculate the illumination condition value of each pixel based on the solar zenith angle, solar azimuth angle, slope, and aspect of each pixel;

[0026] Use the linear fitting method to linearly fit the illumination condition value and reflectivity of each pixel to obtain a correction coefficient;

[0027] Based on the correction coefficient, the average illumination condition value and reflectivity of each pixel, and based on the correction function, perform correction processing on the preprocessed original remote sensing image;

[0028] Perform selection and evaluation on the corrected original remote sensing image to obtain the reference image.

[0029] In one or more embodiments, calculate the illumination condition value of each pixel according to the following formula:

[0030]

[0031] where, IC n , θ An , θ sn respectively represent the illumination condition value, solar azimuth angle, solar zenith angle, aspect, and slope of the nth pixel;

[0032] Perform correction processing on the preprocessed original remote sensing image according to the following correction function:

[0033] Image’ n = Image n -(a*IC n +b)+p;

[0034] where, Image’ n represents the reference image; Image ndenote the preprocessed original remote sensing image; p denotes the average reflectance; both a and b denote correction coefficients.

[0035] Compared with the prior art, the technical solution of this application has the following beneficial technical effects:

[0036] The terrain correction method proposed in this embodiment realizes automatic elimination of terrain effects through deep learning technology, avoids complex terrain parameter calculations, reduces manual intervention, significantly improves processing efficiency, and can adaptively correct multi-temporal and multi-sensor images, automatically adapting to the spatio-temporal changes of the images, avoiding the dependence on illumination changes and sensor differences in traditional methods, and having stronger adaptability and robustness. Especially in complex terrains such as mountains and hills, the terrain correction method of this embodiment significantly improves the radiometric consistency of the images and can overcome the limitations of traditional methods in these terrains. Description of the Drawings

[0037] Figure 1 is a schematic flowchart of a terrain correction method considering multi-temporal and complex terrain effects provided by an embodiment of this application;

[0038] Figure 2A is a schematic diagram of a target terrain area selected in the world map and the corresponding original remote sensing image in this embodiment;

[0039] Figure 2B is a schematic diagram of another target terrain area selected in the world map and the corresponding original remote sensing image in this embodiment;

[0040] Figure 3 is a schematic flowchart of step S102 in an embodiment of this application;

[0041] Figure 4 is a schematic diagram of the principle of terrain correction involved in step S102 in an embodiment of this application;

[0042] Figure 5 is a schematic diagram of the architecture principle of the initial terrain correction model in an embodiment of this application;

[0043] Figures 6A - 6D is a schematic diagram before terrain correction of Sentinel-2 satellite images of a mountainous area in Anhui, China in spring, summer, autumn, and winter in this embodiment;

[0044] Figures 6a - 6d is Figures 6A - 6D a schematic diagram after terrain correction of Sentinel-2 satellite images of a mountainous area in Anhui, China in spring, summer, autumn, and winter;

[0045] Figures 7a - 7dIt is a relationship diagram of the regression slope and correlation coefficient between the illumination condition and the radiance of each spectral band after terrain correction of remote sensing images in a mountainous area of Anhui Province in spring, summer, autumn, and winter using different terrain correction methods;

[0046] Figures 8a - 8d It is a schematic diagram of the peak signal-to-noise ratio of each band before and after correcting the remote sensing images of a flat area in Anhui Province in spring, summer, autumn, and winter using different terrain correction methods;

[0047] Figures 9a - 9d It is a schematic diagram of the structural similarity index of each band before and after correcting the remote sensing images of a flat area in Anhui Province in spring, summer, autumn, and winter using different terrain correction methods;

[0048] Figures 10a - 10d It is a schematic diagram of the error evaluation of the normalized difference vegetation index for the matching of sunny slopes and shady slopes when performing terrain correction on the remote sensing images of a target terrain area in Anhui Province in spring, summer, autumn, and winter using different terrain correction methods;

[0049] Figure 11A and Figure 11a It is a comparison diagram before and after terrain correction of Sentinel-2 satellite images of a mountainous area in Nanning, Guangxi using different terrain correction methods in this embodiment;

[0050] Figure 12 It is a schematic diagram of the regression slope and correlation coefficient between the illumination condition and the radiance of each spectral band after terrain correction of the remote sensing images of a mountainous area in Nanning, Guangxi using different terrain correction methods in this embodiment;

[0051] Figure 13 It is a schematic diagram of the peak signal-to-noise ratio of each band before and after terrain correction of the remote sensing images of a flat area in Nanning, Guangxi using different terrain correction methods in this embodiment;

[0052] Figure 14 It is a schematic diagram of the structural similarity index of each band before and after terrain correction of the remote sensing images of a flat area in Nanning, Guangxi using different terrain correction methods in this embodiment;

[0053] Figure 15 It is a schematic diagram of the error evaluation of the normalized difference vegetation index for the matching of sunny slopes and shady slopes when performing terrain correction on the remote sensing images of a target terrain area in Nanning, Guangxi using different terrain correction methods in this embodiment. Detailed implementation manner

[0054] To facilitate the understanding of this application, the following will provide a more comprehensive description of this application with reference to the relevant drawings. Embodiments of this application are shown in the drawings. However, this application can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, these embodiments are provided to make the disclosure of this application more thorough and comprehensive.

[0055] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the technical field to which this application belongs. The terms used in the specification of this application herein are only for the purpose of describing specific embodiments and are not intended to limit this application.

[0056] Currently, traditional terrain correction methods (such as C-correction, Minnaert model, etc.) rely on complex geometric models and accurate calculations of terrain parameters (such as slope, aspect, solar altitude angle, etc.). When facing large-scale and highly complex regions, these methods have a huge amount of calculation, low efficiency, and poor adaptability to multi-temporal and multi-sensor images. Especially in dynamic scenarios (such as seasonal changes, comparative analysis of multi-temporal data), traditional methods are difficult to effectively handle the impact of changes in illumination angle on images. Relevant scholars analyzed the relevant literature on terrain correction in the past decade (Hu & Li, 2024), and they believe that current terrain correction research is mainly carried out in low-altitude areas and mainly concentrated in months with good solar illumination conditions, which may overestimate the performance of correction methods. At the same time, Nugraha et al. used 10 commonly used terrain correction methods (Nugraha et al., 2024), and their research showed that all 10 terrain correction methods used performed poorly when the solar elevation angle was small.

[0057] There are great difficulties in solving the above technical problems: the problem of uneven illumination in complex terrain is difficult to accurately correct through traditional geometric models. Especially when dealing with multi-temporal images, the changes in illumination angle and seasonal differences greatly reduce the accuracy and efficiency of traditional methods. Therefore, it is necessary to develop a correction method with strong adaptive ability that can automatically adapt to the spatio-temporal changes of images, especially in the case of seasonal changes and changes in illumination angle.

[0058] Based on this, the embodiments of this application provide a terrain correction method that takes into account the influence of multi-temporal and complex terrain, which can improve the terrain correction effect and processing efficiency of remote sensing images.

[0059] In the embodiments of this application, the terrain correction method based on deep learning algorithms still faces multiple challenges in design. Specifically, the following aspects need to be considered:

[0060] In terms of data processing: Deep learning relies on a large amount of high-quality labeled data for training. In remote sensing data, it is very difficult to obtain labeled data, especially for terrain data and the reflection characteristics of ground objects. Image data from different seasons and different sensors often have different lighting conditions and radiation characteristics. How to unify the spectral range and geometric features of these data is the primary difficulty. Therefore, in the embodiments of this application, the existing labeled data is amplified, and the diversity of training data is extended through methods such as rotation, flipping, cropping, scaling, hue adjustment, and noise addition. Convolutional neural networks (CNNs) and generative adversarial networks (GANs) are introduced to use various loss functions (such as perceptual loss, structural similarity loss, etc.) to automatically adjust the lighting and radiation differences of images during the training process to improve the correction effect.

[0061] In terms of model training: Deep learning models, especially training on large-scale datasets, usually require high-performance computing resources such as GPUs, which makes model training very difficult when computing resources are limited. At the same time, during the training process, the amount of remote sensing data is huge and complex. When training a deep learning model, the optimization algorithm may face difficulties in convergence, and the training process may be stuck at a local optimal solution for a long time, resulting in unsatisfactory model performance. To overcome the problems of model training, in the embodiments of this application, the images are standardized, the model is accelerated using GPUs, an adaptive Adam optimizer is used, and a combination of various loss functions is used to improve the quality of the generated images and avoid convergence difficulties during the training process. The BCE loss and the L1 loss help to constrain the generator to generate more realistic images, thereby reducing the risk of overfitting.

[0062] Feature extraction: Feature extraction of remote sensing images is crucial for calibration accuracy. The terrain features in remote sensing data are very complex, with a variety of terrain types (such as mountains, hills, plains, water areas, etc.), and the performance of each terrain in remote sensing images may vary greatly. The reflection characteristics of the same ground object in remote sensing images may change significantly between different seasons. Moreover, different remote sensing sensors (such as Landsat, Sentinel-2, WorldView, etc.) have different resolutions, spectral bands, and radiation characteristics. The deep learning model needs to learn to extract consistent terrain features under different seasonal conditions to ensure the stability and robustness of the model. And it should be able to automatically adapt to the possible spectral deviations and spatial resolution differences between the image data obtained by different sensors, so as to extract reliable terrain features. At the same time, remote sensing data usually contains multiple spectral bands (such as red, green, blue, near-infrared, short-wave infrared, etc.), and the dimension of these data is very high. How to effectively extract terrain information features from these high-dimensional data and avoid the "curse of dimensionality" (that is, the information is too sparse and difficult to model) is a challenge. To overcome the problems faced in feature extraction, in the training process of the generator in the embodiments of this application, a perceptual loss based on the pre-trained VGG16 network is adopted. The intermediate layer features extracted can capture high-level semantic information in the image, helping the generator better restore the image details and structure, and avoiding the limitations of relying only on pixel-level loss. In the architecture of the generator, an encoder-decoder structure is used, and multi-layer feature maps are fused through skip connections in the decoding stage. The feature map of each layer can be restored to a higher resolution in the decoding stage, thus retaining the multi-scale information of the image. By introducing a discriminator, a generative adversarial network (GAN) is constructed. The discriminator guides the generator to generate more realistic images by judging whether the generated image is realistic. The loss of the discriminator and the loss of the generator are jointly optimized to increase the diversity and quality of the generated images. And the image data and elevation data are combined into a multi-channel input, enabling the generator to process these multi-source data simultaneously during the learning process and using the elevation information to help generate more realistic scene images. During the training process, each loss is weighted through weight settings, enabling the generator to comprehensively optimize the adversarial loss, pixel loss, perceptual loss, and SSIM loss, ensuring that the generated images reach a high level in terms of visual quality, structural similarity, and details, thus effectively removing the terrain effect.

[0063] The topographic correction method proposed in the embodiments of this application can not only effectively remove the topographic effects in multi-temporal remote sensing images, but also quickly restore the true spectral information of the remote sensing images. Through batch processing, manual intervention is reduced, and the availability of remote sensing image data is improved. Compared with traditional methods, the embodiments of this application have stronger adaptability and robustness, can maintain high-precision correction in multi-temporal and multi-sensor images, and provide a new topographic effect elimination technology for related fields. The wide application of this technology will greatly promote the application of remote sensing images in related fields and provide more reliable data support for related research.

[0064] Refer to Figure 1 , Figure 1 The flowchart of the topographic correction method considering multi-temporal and complex terrain effects provided by an embodiment of this application is shown. A topographic correction method considering multi-temporal and complex terrain effects provided by an embodiment of this application includes the following steps:

[0065] S101. Obtain the original remote sensing images and DEM data of the target topographic area at multiple time phases and perform preprocessing to obtain the preprocessed original remote sensing images and DEM data;

[0066] S102. Based on the preprocessed DEM data, perform topographic correction on the preprocessed original remote sensing images to obtain reference images;

[0067] S103. Input the preprocessed original remote sensing images and the corresponding DEM data into the initial topographic correction model, and perform iterative training based on the reference images corresponding to the original remote sensing images to obtain the target topographic correction model; the initial topographic correction model is constructed based on a generative adversarial network;

[0068] S104. Obtain the remote sensing images to be corrected and the corresponding DEM data and perform preprocessing, and input the preprocessed images to be corrected and the corresponding DEM data into the target topographic correction model to obtain the target remote sensing images.

[0069] Specifically, obtain optical remote sensing images output by different sensors at different time points (such as the four seasons of spring, summer, autumn, and winter) in multiple complex terrain areas as the original remote sensing images, and obtain the digital elevation model (DEM) data corresponding to the original remote sensing images of the same target topographic area to form a data pair of the original remote sensing images and DEM data, and then perform preprocessing such as radiometric calibration and geometric correction on the data pair. Then, use a classic topographic correction method or other topographic correction methods, and perform topographic correction on the preprocessed original remote sensing images according to the preprocessed DEM data to obtain reference images. In this way, it is convenient to construct a training dataset.

[0070] An initial terrain correction model is constructed based on a generative adversarial network. Then, the preprocessed original remote sensing image and the corresponding DEM data are input into the initial terrain correction model, and the reference image corresponding to the original remote sensing image is used as the real image to perform back-iteration training on the initial terrain correction model to obtain the trained target terrain correction model.

[0071] When subsequent terrain correction of any complex terrain and multi-temporal remote sensing images is required, the image to be corrected and the corresponding DEM data are obtained and preprocessed, and the preprocessed image to be corrected and the corresponding DEM data are input into the target terrain correction model to obtain the target remote sensing image. In this way, not only can the terrain effect in multi-temporal remote sensing images be effectively removed, but also the true spectral information of the remote sensing image can be quickly restored. Through batch processing, manual intervention is reduced, and the availability of remote sensing image data is improved. Compared with traditional methods, the terrain correction method of this embodiment has stronger adaptability and robustness, can maintain high-precision correction in multi-temporal and multi-sensor images, and provides more reliable data support for research in related fields.

[0072] Compared with traditional methods, the terrain correction method proposed in this embodiment realizes automatic elimination of terrain effects through deep learning technology, avoids complex terrain parameter calculations, reduces manual intervention, and significantly improves processing efficiency. The terrain correction method of this embodiment can adaptively correct multi-temporal and multi-sensor images, automatically adapt to the spatio-temporal changes of the images, and avoids the dependence on illumination changes and sensor differences in traditional methods, with stronger adaptability and robustness. Especially in complex terrains such as mountains and hills, the terrain correction method of this embodiment significantly improves the radiometric consistency of the images and can overcome the limitations of traditional methods in these terrains.

[0073] In addition, the terrain correction method of this embodiment reduces the dependence on high-precision external terrain data, can directly extract key information from the images, and can maintain excellent correction effects even in the absence of accurate terrain data. This enables the terrain correction method of this embodiment to provide accurate and stable remote sensing image support under various environmental conditions, promotes the application of remote sensing images in fields such as ecological monitoring and environmental assessment, and provides a reliable image data basis for related research.

[0074] Refer to Figure 2A and Figure 2B , Figure 2A shows a schematic diagram of a target terrain area selected in the world map and the corresponding original remote sensing image in this embodiment, Figure 2B shows a schematic diagram of another target terrain area selected in the world map and the corresponding original remote sensing image in this embodiment.

[0075] In some embodiments, step S101 is to obtain the original remote sensing images and DEM data of the target terrain area at multiple time phases and perform preprocessing, including:

[0076] Under the condition that the cloud coverage rate of the target terrain area is less than the preset coverage rate threshold, obtain the original remote sensing images and corresponding DEM data at multiple time phases; the preset coverage rate threshold is between 4% and 6%; perform radiometric calibration and geometric correction on the original remote sensing images and corresponding DEM data, and perform georegistration on the processed original remote sensing images and corresponding DEM data.

[0077] Specifically, in this embodiment, obtain the original remote sensing images at different time points covering the target terrain area (such as the four seasons of spring, summer, autumn, and winter), and require the cloud coverage rate to be less than 5%. In other embodiments, the cloud coverage rate can be required to be less than 4% or 6%, which can be specifically set according to the actual situation. Obtain the DEM data of the original remote sensing images of the same target terrain area, perform radiometric calibration and geometric orthorectification on all the original remote sensing images and DEM data, and then perform spatial registration. In this way, through radiometric calibration, the digital value (i.e., DN value) of the original remote sensing image can be converted into the radiance at the top of the atmosphere to ensure the consistency of subsequent processing, as shown in Figure 2 and Figure 2B as shown. Through geometric orthorectification, the stability and accuracy of subsequent data processing can be ensured; through spatial registration, the original remote sensing images and DEM data can be adjusted to a unified earth surface grid to ensure the consistency of subsequent processing.

[0078] Refer to Figure 3 and Figure 4 , Figure 3 shows the schematic flow chart of step S102 in an embodiment of the present application, Figure 4 shows the schematic principle diagram of terrain correction involved in step S102 in an embodiment of the present application.

[0079] In some embodiments, step S102 is to perform terrain correction on the preprocessed original remote sensing image based on the preprocessed DEM data to obtain a reference image, including:

[0080] S301. Reproject the preprocessed DEM data so that it is in the same coordinate system as the preprocessed original remote sensing image.

[0081] S302. Obtain the solar zenith angle and solar azimuth angle of each pixel from the preprocessed original remote sensing image, and calculate the slope and aspect of each pixel based on the corresponding reprojected DEM data.

[0082] In this embodiment, the solar zenith angle and solar azimuth angle of each pixel are obtained from the metadata of the preprocessed original remote sensing image. The slope and aspect of each pixel are calculated through DEM (Digital Elevation Model) data to achieve terrain analysis. Specifically, the slope calculation is based on the elevation values of each pixel and its neighborhood in the DEM data, and the aspect calculation is based on the maximum downhill direction of each pixel relative to the slope surface in the DEM data.

[0083] S303. Calculate the illumination condition value of each pixel based on the solar zenith angle, solar azimuth angle, slope, and aspect of each pixel.

[0084] Furthermore, the illumination condition value of each pixel can be calculated according to the following formula:

[0085]

[0086] where IC n , θ An , θ sn successively represent the illumination condition value, solar azimuth angle, solar zenith angle, aspect, and slope of the nth pixel. In this way, through multiple terrain parameters, the illumination situation of each pixel in the original remote sensing image can be accurately calculated and analyzed.

[0087] S304. Use the regression analysis method to perform linear fitting on the illumination condition value and reflectivity of each pixel to obtain the correction coefficient. Specifically, use the regression analysis method of linear fitting to perform linear fitting on the illumination condition value and reflectivity of each pixel to obtain the correction coefficient.

[0088] S305. Based on the correction coefficient, the illumination condition value and reflectivity average value of each pixel, and based on the correction function, perform correction processing on the preprocessed original remote sensing image.

[0089] Furthermore, perform correction processing on the preprocessed original remote sensing image according to the following correction function: Image’ n = Image n -(a * IC n + b)+ p;

[0090] where Image’ n represents the reference image; Image n represents the preprocessed original remote sensing image; p represents the reflectivity average value; both a and b represent correction coefficients.

[0091] Specifically, perform terrain correction on the preprocessed original remote sensing image according to the above expression to eliminate the terrain effect and obtain the terrain-corrected reference image.

[0092] S306. Select and evaluate the original remotely sensed image after calibration processing to obtain the reference image.

[0093] Specifically, perform radiometric consistency and geometric accuracy checks on the original remotely sensed image after calibration processing, and compare it with the uncalibrated original remotely sensed image to screen out the effectively calibrated reference image, thereby providing effective data support for subsequent training of the terrain correction model.

[0094] Refer to Figure 5 , Figure 5 which shows a schematic diagram of the architecture principle of the initial terrain correction model in an embodiment of the present application. In some embodiments, the initial terrain correction model includes a generator and a discriminator. The generator adopts a U-Net structure, and the discriminator adopts a convolutional neural network.

[0095] Specifically, design a generative adversarial network (GAN) architecture, including a generator (Generator) and a discriminator (Discriminator), as Figure 5 shown. The generator adopts a U-Net structure, which is an encoder-decoder structure. The encoder processes the multi-channel original remotely sensed image with terrain effect influence and DEM data to generate a terrain-corrected image. The decoder restores the image resolution through a transposed convolution layer and retains the details of the input image by combining skip connections. The discriminator adopts a convolutional neural network, which is used to judge whether the input image is a real target image or a corrected image generated by the generator. Higher-level feature information is extracted through layer-by-layer convolution and reduction of spatial resolution, and finally the true or false judgment of the image is output.

[0096] Furthermore, the training process of the initial terrain correction model includes:

[0097] Fuse the preprocessed original remotely sensed image and the corresponding DEM data and input them into the generator to output a generated image, and input the generated image and the corresponding reference image into the discriminator; based on the generated image and the corresponding reference image, the generator is optimized using a second cross-entropy loss function, a pixel loss function, a structural similarity loss function, and a visual perception loss function, and the discriminator is optimized using a first cross-entropy loss function.

[0098] It should be noted that the input size of the image data is set to 512×512 pixels, and the number of channels is adjusted as needed to ensure the combined processing of multi-temporal image data and DEM data. In this embodiment, the input data needs to include remotely sensed images of different seasons or time points to ensure that the model can handle the spatio-temporal changes of multi-temporal data.

[0099] Specifically, first, the preprocessed original remote sensing images and the corresponding DEM data are read from the image data folder. Then, the rasterio library is used to load these original remote sensing images and DEM data and convert them into a suitable Tensor format. By resizing the image size to 512x512 and performing normalization operations, it is ensured that the original image data and DEM data can be effectively input into the initial terrain correction model. The original image data and DEM data are merged into a single tensor Tensor to ensure that the generator can simultaneously learn the image content and terrain information.

[0100] Furthermore, the generated image output by the generator and the corresponding reference image are stitched together and input into the discriminator; the discriminator makes predictions on the stitched image and outputs the probability value of the image authenticity. Ideally, the discriminator will consider the reference image as true and output a probability value close to 1, and the generated image output by the generator as false and output a probability value close to 0. It should be noted that the second cross-entropy loss function and the first cross-entropy loss function are constructed based on binary cross-entropy (BCE), and the loss errors of the generator and the discriminator are calculated respectively to optimize the output of the model. In addition, the generator also uses a pixel loss function, a structural similarity loss function, and a visual perception loss function for joint optimization, enabling the model to learn terrain effect correction and improve its correction level.

[0101] In some embodiments, the first cross-entropy loss function determines the cross-entropy loss of the discriminator based on the predicted probability values output by the discriminator for the generated image and the corresponding reference image respectively.

[0102] Specifically, the loss is calculated and averaged using cross-entropy to obtain the final discriminator loss value. The formula for the first cross-entropy loss function is:

[0103]

[0104] where L D_BCE represents the cross-entropy loss of the discriminator; L real_BCE represents the loss of the discriminator's prediction for the reference image; L fake_BCE represents the loss of the discriminator's prediction for the generated image; N real represents the total number of pixels in the reference image; D(x real ) represents the predicted probability value of the discriminator for the reference image x real ; N fake represents the total number of pixels in the generated image; D(x fake ) represents the predicted probability value of the discriminator for the generated image x fake .

[0105] In this way, by training the discriminator through reverse iteration, the discriminator can distinguish whether the input image is a reference image or an image generated by the generator.

[0106] In some embodiments, the second cross-entropy loss function determines the cross-entropy loss of the generator based on the predicted probability values output by the discriminator for the generated image.

[0107] Specifically, the binary cross-entropy is used to calculate the cross-entropy loss of the generator. The formula of the second cross-entropy loss function is:

[0108]

[0109] where L G_BCE represents the cross-entropy loss of the generator; x fake represents the generated image; N fake represents the total number of pixels of the generated image, and i represents the pixel index; D(x fake ) represents the predicted probability value of the discriminator for the generated image; it should be noted that during the first round of training, both the generator and the discriminator are randomly initialized, so the output is also random.

[0110] Furthermore, the pixel loss function adopts the L1 norm to calculate the pixel differences at the same positions between the generated image and the corresponding reference image, so as to determine the pixel loss of the generator.

[0111] Specifically, the L1 norm is used to measure the pixel differences between the generated image and the reference image, ensuring that the generated image is as close as possible to the corresponding reference image. In this embodiment, the formula of the pixel loss function is:

[0112]

[0113] where N represents the total number of pixels in the image; y i represents the spectral reflectance value of the i-th pixel in the reference image; represents the spectral reflectance value of the i-th pixel in the generated image. It should be noted that the absolute errors of all pixels are calculated to ensure the precise alignment of the generated image and the reference image at the pixel level.

[0114] Furthermore, the structural similarity loss function uses the sliding window method to calculate the SSIM value of the generated image and the corresponding reference image in each local region, so as to determine the structural similarity loss of the generator.

[0115] Specifically, the sliding window method is used to calculate the SSIM value (i.e., structural similarity) of the generated image and the reference image in the local region. The SSIM value is used to measure the differences in brightness, contrast, and structure of the images. In this embodiment, the SSIM value ranges from 0 to 1, where 1 represents exactly the same and 0 represents no similarity. In this embodiment, the formula of the structural similarity loss function is:

[0116]

[0117] Where: μ x and μ y respectively represent the average brightness of the generated image x and the reference image y; and respectively represent the brightness variances of the generated image x and the reference image y; 2σ xy represents the brightness covariance between the generated image x and the reference image y; C1 and C2 represent stabilization constants. It should be noted that the structural similarity of the generated image and the reference image in each local region is calculated and the global average value is taken to measure the overall structural consistency.

[0118] Furthermore, the visual perception loss function adopts the L1 norm, and based on the feature maps extracted from the generated image and the corresponding reference image in the same convolutional layer through a pre-trained VGG network, the visual perception loss of the generator is determined.

[0119] Specifically, the multi-channel remote sensing image with terrain influence is input and the generated image is output by the generator (Generator), i.e., the image after terrain correction. The pre-trained VGG network is used to extract the feature maps of the generated image

[0120]

[0121] and the corresponding reference image y in different convolutional layers. Then, the L1 norm difference between the feature map of the generated image and the feature map of the reference image is calculated to measure the similarity between the two in the high-level feature space, so as to ensure that the high-level perception features of the generated image are close to the reference image. In this embodiment, the formula of the visual perception loss function is: i where x i represents the spectral reflectance value of the i-th pixel in the generated image, and y i represents the spectral reflectance value of the i-th pixel in the reference image; Ψ(x i ) represents the feature map extracted from the generated image through the VGG network; Ψ(y represents the L1 norm.

[0122] In this embodiment, based on the cross-entropy loss, pixel loss, structural similarity loss, and visual perception loss, and calculated according to different weight coefficients, the final generator loss value is obtained. The total loss of the generator is: L G = L G_BCE + 200L L1 + 50LSSIM +50L Perceptual In this way, during the training process, each loss is weighted through weight setting, and the parameters of the generator are updated, enabling the generator to comprehensively optimize the cross-entropy loss, pixel loss, structural similarity loss, and visual perception loss, ensuring that the generated image reaches a high level in terms of visual quality, structural similarity, and details, thereby effectively removing the terrain effect. During the testing process, the new remote sensing image and DEM data are preprocessed and used as input data, and then the input data is input into the trained generator model to obtain the terrain-corrected image.

[0123] In this embodiment, through the continuous cyclic competition between the generator and the discriminator, the generated image output by the generator gradually approaches the real reference image, enabling the model to effectively learn the variation law of the terrain effect of remote sensing images at different times. In addition, in the model, the generator uses visual analysis, pixel analysis, and structural similarity index to perform quality control on all corrected images, automatically identifying images with poor correction effects and performing post-processing; and through the error analysis and post-processing steps, the radiometric consistency, geometric accuracy, and visual quality of the final output image are ensured. It should be noted that the remote sensing image after the model correction process is adjusted to the original image size, and a complete corrected image is output to ensure consistency with the original remote sensing image in terms of spatial resolution and physical size. In this way, the trained target terrain correction model can generate a corrected image without the influence of the terrain effect, obtain a high-quality corrected image, eliminate the interference of complex terrain conditions on the remote sensing image, and provide reliable image data for subsequent remote sensing data analysis.

[0124] Considerations on algorithm complexity and computational efficiency:

[0125] (1) The terrain correction method of this embodiment does not require complex geometric calculations and terrain parameter adjustments, and completely relies on the training results of the generative adversarial model for correction, greatly reducing manual intervention and calculation time. It supports batch processing of a large amount of image data and can quickly process thousands of images after training, significantly improving the efficiency of data processing.

[0126] (2) The training and inference processes of the terrain correction method in this embodiment have certain hardware requirements, especially when running on large-scale datasets. GPU acceleration is used to move the generator and discriminator models to the GPU. In the training loop, the training data is migrated to the GPU to improve processing efficiency. For large-scale remote sensing image data, it is recommended to use a GPU with a large memory to process high-resolution images and multi-channel data. For hardware with memory limitations, techniques such as model pruning and mixed-precision training can be used to reduce memory consumption. For ultra-large-scale datasets, it is recommended to perform distributed training in a cluster environment. In terms of storage requirements, storing high-resolution remote sensing images and DEM data requires a large amount of disk space. Especially when processing datasets in the order of several terabytes, a high-speed storage system (such as an SSD) needs to be equipped to ensure fast data reading and writing. Depending on the data volume, it is recommended to configure at least 1TB of available storage space.

[0127] Model correction effect verification:

[0128] (1) Visual analysis

[0129] Refer to Figures 6A - 6D 、 Figures 6a - 6d , Figures 6A - 6D shows a schematic diagram before terrain correction of Sentinel-2 satellite images of a mountainous area in Anhui, China, for the four seasons of spring, summer, autumn, and winter in this embodiment. Figures 6a - 6d shows Figures 6A - 6D a schematic diagram after terrain correction of Sentinel-2 satellite images of a mountainous area in Anhui, China, for the four seasons of spring, summer, autumn, and winter; a false-color (near-infrared, red, green) band combination image is used in the figure; among them, Figure 6A 、 Figure 6B 、 Figure 6C 、 Figure 6D show the original remote sensing images of spring, summer, autumn, and winter respectively; Figure 6a 、 Figure 6b 、 Figure 6c 、 Figure 6d show the images after terrain correction processing of the original remote sensing images of spring, summer, autumn, and winter using the terrain correction method proposed in this embodiment respectively. Visual comparison shows that the terrain correction method proposed in this embodiment effectively removes the terrain effect in the original remote sensing image and visually maintains the color consistency of each land cover.

[0130] (2) Correlation analysis

[0131] Refer to Figures 7a - 7d , Figures 7a - 7dThe figure respectively shows the relationship diagrams of the regression slope (Slope) and correlation coefficient (R) between the illumination condition (IC) and the radiance of each spectral band after terrain correction of remote sensing images in spring, summer, autumn, and winter in a mountainous area of Anhui Province using different terrain correction methods in this embodiment, which is used to measure the improvement effect of terrain correction on the image. In each figure, band1-band6 respectively represent blue, green, red, near infrared (nir), shortwave infrared 1 (swir1), and shortwave infrared 2 (swir2). In an ideal state, terrain correction should make Slope and R as close to 0 as possible to indicate that the relationship between the image and the solar incidence angle is fully corrected and the radiation error caused by the terrain is eliminated. In the image without terrain correction (UNTOC), the Slope and R values of each band in the four seasons are significantly high. Especially in the autumn of Figure 7c and the winter of Figure 7d , in the near-infrared spectral band band4 and shortwave infrared 1 spectral band band5, Slope reaches 0.184 and 0.198 respectively, and R is as high as 0.643 and 0.698 respectively, indicating that the original remote sensing image is seriously affected by terrain radiation error. Against this background, different correction methods attempt to reduce the image radiation error caused by the terrain through their respective algorithms.

[0132] The terrain correction method (Deep-TC) of this embodiment performs excellently in most bands and seasons, significantly reducing the Slope value and R value. For example, Figure 7d in the near-infrared spectral band band4 in winter, Slope is reduced from 0.198 of UNTOC to 0.025, and R is reduced from 0.698 of UNTOC to 0.113, and the correction effect is significant; Figure 7c in the shortwave infrared 1 spectral band band5 in autumn, Slope is reduced from 0.145 of UNTOC to -0.011, and R is reduced from 0.662 of UNTOC to -0.073, further verifying its advantage in high-error bands. In addition, compared with traditional methods such as C, EMIN, MIN, and SCS+C, the terrain correction method (Deep-TC) of this embodiment performs more excellently in most cases, not only effectively reducing the radiation error in high-error bands, but also maintaining stable correction ability in low-error bands. Especially in winter and autumn, the correction effect of the Deep-TC method is particularly prominent, showing its applicability and robustness under complex terrain conditions. Thus, it can be seen that the terrain correction method of this embodiment has a powerful correction ability for the radiation error of different bands.

[0133] (3) Structural similarity analysis

[0134] Since the areas with different slopes and aspects in the mountainous region are affected by sunlight differently. In this embodiment, to clarify the influence of terrain on the image reflectance, the regions are classified based on slope and aspect. The flat region refers to the region with a slope less than 5°. Due to the small terrain undulation, the lighting conditions in these regions are relatively uniform, and the influence of terrain on the image reflectance is weak. In this study, by calculating the peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) of the images before and after terrain correction in the flat region, the improvement effect of the terrain correction method on spectral consistency is evaluated. In an ideal state, the corrected images of different bands should have high and consistent PSNR and SSIM values, indicating that the corrected images are highly consistent with the original images in spectral characteristics and successfully eliminate the interference of terrain on spectrum and structure.

[0135] The results are as Figures 8a - 8d and Figures 9a - 9d shown in Figure 8a 、 Figure 8b 、 Figure 8c 、 Figure 8d respectively show the schematic diagrams of the peak signal-to-noise ratio (PSNR) of each band before and after correcting the remote sensing images of spring, summer, autumn, and winter in a flat region of Anhui Province using different terrain correction methods in this embodiment. Figure 9a 、 Figure 9b 、 Figure 9c 、 Figure 9d respectively show the schematic diagrams of the structural similarity index (SSIM) of each band before and after correcting the remote sensing images of spring, summer, autumn, and winter in a flat region of Anhui Province using different terrain correction methods in this embodiment. It can be seen from the figures that both PSNR and SSIM are at a relatively high level. For example, in the blue spectral band band1 in spring, the PSNR of the remote sensing image of the flat region before and after terrain correction based on the terrain correction method of this embodiment reaches 55.7, and the SSIM is 0.9998, approaching the ideal state; in the short-wave infrared 1 band5 in summer, the PSNR is 45.26 and the SSIM is 0.9986, indicating that even in the high band with large interference, the terrain correction method of this embodiment can maintain high spectral consistency. In addition, under the complex lighting conditions in autumn and winter, the terrain correction method of this embodiment still shows a high correction effect, and the stability of PSNR and SSIM is better than that of most traditional correction methods. Generally speaking, the terrain correction method of this embodiment shows good spectral consistency and structural similarity in the flat region, indicating that the method of the present invention can maintain the original spectral characteristics and avoid introducing noise or distorting the true ground object information (such as farmland, water body) due to overcorrection, verifying its correction ability and stability under uniform lighting conditions.

[0136] (4) Error assessment

[0137] When evaluating errors, in this embodiment, the effects of different slopes and aspects on NDVI (Normalized Difference Vegetation Index) are considered simultaneously. The matching pairs of sunny slopes and shady slopes are compared according to similar aspect intervals, so as to evaluate the influence of different lighting conditions on NDVI. In this embodiment, the sunny slope and the shady slope are respectively divided into 6 sub-intervals. The sunny slope (the included angle between the orientation and the solar azimuth angle is less than 90°) is subdivided. According to the aspect angle, it is divided into six intervals, and the slope is limited to be greater than the flat threshold (5°): SunSlope1-6, each angle is 15°. The shady slope (the included angle between the orientation and the solar azimuth angle is greater than 90°) is subdivided. Similarly, according to the aspect angle, it is divided into six intervals, and the slope is limited to be greater than the flat threshold (5°): ShadeSlope1-6, each angle is 15°. According to the aspect angle composition: matching pair Pair1 (SunSlope1 and ShadeSlope1), matching pair Pair2 (SunSlope2 and ShadeSlope2), matching pair Pair3 (SunSlope3 and ShadeSlope3), matching pair Pair4 (SunSlope4 and ShadeSlope4), matching pair Pair5 (SunSlope5 and ShadeSlope5), matching pair Pair6 (SunSlope6 and ShadeSlope6). The aspect angle ranges of each pair are similar, but the orientations are opposite, so as to more reasonably reflect the effectiveness of terrain correction under different slope conditions.

[0138] See Figures 10a - 10d , Figures 10a - 10d respectively show the schematic diagrams of the error evaluation of the normalized difference vegetation index (NDVI) of the matching pairs of sunny slopes and shady slopes when remote sensing images of a target terrain area in spring, summer, autumn, and winter in Anhui Province are subjected to terrain correction using different terrain correction methods. It can be seen from the figure that after terrain correction, the vegetation index values of sunny slopes and shady slopes should be closer. If the prediction error (MRE) and the normalized root mean square error value (NRMSE) of NDVI are reduced, it indicates that the reflectance of different aspect regions tends to be consistent after correction, and the correction effect is good. It can be seen from the figure that the terrain correction method of this embodiment shows significant error control advantages under different aspect conditions, especially more obvious for the matching pairs with large slopes and significant differences between sunny slopes and shady slopes. First, there are seasonal differences in errors. As Figure 10a and Figure 10b show, for the remote sensing images in spring and summer, the Deep-TC method generally performs best, with the lowest MRE and NRMSE, and a smaller error range. The SCS+C and C methods perform sub-optimally for the remote sensing images in spring and summer, especially for matching pairs Pair1 to Pair4. The MRE and NRMSE of UNTOC are higher than other methods in most matching pairs. As Figure 10c and Figure 10dAs shown, in autumn and winter, the error in winter is significantly higher than that in spring and summer. In particular, the MRE of UNTOC reaches 0.7319 and the NRMSE reaches 0.7231 in Pair1. The performances of VECA and Deep-TC are relatively stable in autumn and winter. In particular, the MRE and NRMSE of the VECA method are relatively low in Pair5 and Pair6. The errors of the EMIN and MIN methods gradually increase, but they are still superior to the traditional methods in winter. The terrain correction method of this embodiment performs optimally in all seasons. Especially in Pair1 to Pair4, the MRE and NRMSE are the lowest, indicating its advantages in reducing errors and matching stability, and having good cross-band stability and versatility.

[0139] Consider the model calculation efficiency:

[0140] The training and inference processes of the terrain correction method of this embodiment have certain hardware requirements, especially when running on large-scale datasets. The processor used in the terrain correction method of this embodiment is an Intel(R) Core(TM) i7-10700KF CPU@3.80GHz, which has 8 physical cores and 16 threads, a base main frequency of 3.80GHz, and a maximum turbo frequency of 5.10GHz. It can effectively process multiple tasks in parallel, especially in the early data preprocessing and feature extraction processes of the model. The terrain correction method of this embodiment uses an NVIDIA GeForce RTX 3080 GPU, equipped with 8704 CUDA cores and 10GB GDDR6X video memory. In the training and inference stages of the computationally intensive model, it can significantly reduce the training time. The hardware acceleration of RTX3080 enables the terrain correction method of this embodiment to process a large amount of pixel data within a limited time in the large-scale training task of multi-temporal data, thereby greatly improving the model training efficiency. This system is equipped with 32GB of memory, thereby reducing the data reading delay caused by memory bottlenecks and ensuring that the model has a high operating efficiency when processing a large number of high-resolution remote sensing images. This system is equipped with 3TB of storage space for storing training data, model parameters, and intermediate calculation results. In addition, the storage configuration of the system also supports the rapid backup and recovery of large-scale datasets, ensuring the data security during the calculation process. When performing data training, the computing platform used can complete the training process of the entire multi-temporal dataset in about 4 hours, including image data preprocessing, feature extraction, model training, and parameter optimization, etc. In the model inference stage, the correction time for a single piece of data is less than 30 seconds, and the time for outputting the complete corrected image (i.e., the consistency of the spatial resolution and physical size of the uniformly corrected image) is less than 1 minute.

[0141] Model generalization test:

[0142] To verify the adaptability and generalization of the target terrain correction model provided in this embodiment in different geographical regions, another test area (Nanning, Guangxi) with significant geographical feature differences from the above data (Xuancheng, Anhui) was selected for verification. The remote sensing image data of these test areas cover different terrain, climate, seasonal and other characteristics, ensuring a comprehensive evaluation of the model's performance in diverse environments. The remote sensing images and DEM data were preprocessed and used as test data. The test data was corrected using the target terrain correction model of this embodiment to evaluate the overall performance of the model in the new test area.

[0143] Specifically, Nanning City, Guangxi was selected as the new test area. During the test, the coverage area of each piece of data was 12,000 square kilometers, and the resolution was 120 million pixels. The test data contains rich ground object information and has relatively complex terrain features, such as a mixed landform of hills, rivers, and urbanized areas. The same preprocessing process as in the training stage was performed on the remote sensing image data of Nanning, including radiometric correction, geometric correction, noise removal, etc., to ensure the consistency and accuracy of the input data. The target terrain correction model obtained by training was used to infer the multi-temporal data of Nanning, and the correction results were output. Refer to Figure 11A and Figure 11a , Figure 11A and Figure 11a show the comparison diagrams before and after terrain correction of the Sentinel-2 satellite image of a mountainous area in Nanning, Guangxi using different terrain correction methods in this embodiment. The figure shows the false color (near-infrared, red, green) band combination image; among them, Figure 11A shows the original remote sensing image; Figure 11a shows the image after terrain correction using the terrain correction method of this embodiment. Visual comparison shows that the terrain correction method of this embodiment can effectively remove the terrain effect in the image and visually maintain the color consistency of each land cover.

[0144] Refer to Figure 12 , Figure 12 show the schematic diagrams of the regression slope (Slope) and correlation coefficient (R value) between the illumination condition and the radiance of each spectral band after terrain correction of the remote sensing image of a mountainous area in Nanning, Guangxi using different terrain correction methods in this embodiment. In the uncorrected image (UNTOC), the Slope and R values of each band are relatively high, and the highest Slope and R reach 0.422 and 0.226 respectively. After correction using the terrain correction method of this embodiment, the Slope and R values are reduced, and the Slope of each band is less than 0.06, and the R is less than 0.13.

[0145] Refer to Figure 13 and Figure 14 ,Figure 13 It shows a schematic diagram of the peak signal-to-noise ratio (PSNR) of each band before and after terrain correction of remote sensing images of a flat area in Nanning, Guangxi, using different terrain correction methods in this embodiment. Figure 14 It shows a schematic diagram of the structural similarity index (SSIM) of each band before and after terrain correction of remote sensing images of a flat area in Nanning, Guangxi, using different terrain correction methods in this embodiment. It can be seen from the two figures that the PSNR and SSIM of each band are at a relatively high level. After correcting with the terrain correction method of this embodiment, the PSNR reaches up to 56.6, and the SSIM is 0.999, approaching the ideal state.

[0146] Refer to Figure 15 , Figure 15 It shows a schematic diagram of the error evaluation of the normalized difference vegetation index (NDVI) for the matching of sunny slopes and shady slopes when performing terrain correction on remote sensing images of a target terrain area in Nanning, Guangxi, using different terrain correction methods in this embodiment; after using the terrain correction method of this embodiment, the errors are all reduced. For example, the mean relative error (MRE) of the prediction for the matching pair pair1 is reduced from 0.16 to 0.10, a decrease of approximately 37.5%, and the MRE of the prediction for the matching pair pair2 is reduced from 0.11 to 0.07, a decrease of approximately 36.4%. The other matching pairs are reduced by 33.3%, 40.0%, 37.5% and 16.7% respectively.

[0147] Generally speaking, the terrain correction method of this embodiment has achieved remarkable results in the generalization test in Nanning, Guangxi, indicating that it has strong adaptability under different regional and terrain conditions, can effectively improve the quality of remote sensing images, and has broad application prospects in practical applications.

[0148] The above are only the preferred embodiments of the present application, and are not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present application shall be included within the protection scope of the present application.

Claims

1. A terrain correction method considering the influence of multi-temporal and complex terrain, characterized in that, Including: Obtain the original remote sensing images and DEM data of the target terrain area at multiple time phases and perform preprocessing to obtain the preprocessed original remote sensing images and DEM data; Based on the preprocessed DEM data, perform terrain correction on the preprocessed original remote sensing image to obtain a reference image; Input the preprocessed original remote sensing image and the corresponding DEM data into an initial terrain correction model, and perform iterative training based on the reference image corresponding to the original remote sensing image to obtain a target terrain correction model; the initial terrain correction model is constructed based on a generative adversarial network; Obtain the remote sensing image to be corrected and the corresponding DEM data and perform preprocessing, and input the preprocessed image to be corrected and the corresponding DEM data into the target terrain correction model to obtain a target remote sensing image.

2. The terrain correction method considering the influence of multi-temporal and complex terrain according to claim 1, characterized in that, The initial terrain correction model includes a generator and a discriminator. The generator adopts a U-Net structure, and the discriminator adopts a convolutional neural network; The training process of the initial terrain correction model includes: Fuse the preprocessed original remote sensing image and the corresponding DEM data and input them into the generator, output a generated image, and input the generated image and the corresponding reference image into the discriminator; Based on the generated image and the corresponding reference image, the discriminator is optimized using a first cross-entropy loss function, and the generator is optimized using a second cross-entropy loss function, a pixel loss function, a structural similarity loss function, and a visual perception loss function.

3. The terrain correction method considering the influence of multi-temporal and complex terrain according to claim 2, characterized in that The first cross-entropy loss function determines the cross-entropy loss of the discriminator based on the predicted probability values output by the discriminator for the generated image and the corresponding reference image respectively.

4. The terrain correction method considering the influence of multi-temporal and complex terrain according to claim 2, characterized in that The second cross-entropy loss function determines the cross-entropy loss of the generator based on the predicted probability value output by the discriminator for the generated image; The pixel loss function uses the L1 norm to calculate the pixel difference between the generated image and the corresponding reference image at each same position to determine the pixel loss of the generator; The structural similarity loss function uses a sliding window method to calculate the SSIM value of the generated image and the corresponding reference image in each local area to determine the structural similarity loss of the generator; The visual perception loss function uses the L1 norm and determines the visual perception loss of the generator based on the feature maps extracted by the pre-trained VGG network for the generated image and the corresponding reference image in the same convolutional layer.

5. The terrain correction method considering the influence of multi-temporal and complex terrain according to any one of claims 1-4, characterized in that Obtain the original remote sensing images and DEM data of the target terrain area at multiple time phases and perform preprocessing, including: Under the condition that the cloud coverage rate of the target terrain area is less than a preset coverage rate threshold, obtain the original remote sensing images and the corresponding DEM data at multiple time phases; the preset coverage rate threshold is between 4% and 6%; Perform radiometric calibration and geometric correction processing on the original remote sensing image and the corresponding DEM data, and perform georegistration on the processed original remote sensing image and the corresponding DEM data.

6. The terrain correction method considering the influence of multi-temporal and complex terrain according to any one of claims 1-4, characterized in that Based on the preprocessed DEM data, perform terrain correction on the preprocessed original remote sensing image to obtain a reference image, including: Reproject the preprocessed DEM data so that it is in the same coordinate system as the preprocessed original remote sensing image; Obtain the solar zenith angle and solar azimuth angle of each pixel from the preprocessed original remote sensing image, and calculate the slope and aspect of each pixel based on the corresponding reprojected DEM data; Calculate the illumination condition value of each pixel based on the solar zenith angle, solar azimuth angle, slope and aspect of each pixel; Use the linear fitting method to linearly fit the illumination condition value and reflectivity of each pixel to obtain the correction coefficient; Based on the correction coefficient, the illumination condition value and reflectivity average value of each pixel, and based on the correction function, perform correction processing on the preprocessed original remote sensing image; Select and evaluate the preprocessed original remote sensing image to obtain the reference image.

7. The terrain correction method considering the influence of multi-temporal and complex terrain according to claim 6, characterized in that, Calculate the illumination condition value of each pixel according to the following formula: Among them, IC n , θ An , θ sn successively represent the illumination condition value, solar azimuth angle, solar zenith angle, slope aspect and slope of the nth pixel; Perform correction processing on the preprocessed original remote sensing image according to the following correction function: Image’ n = Image n -(a * IC n + b) + p; Among them, Image’ n represents the reference image; Image n represents the preprocessed original remote sensing image; p represents the average reflectance; both a and b represent calibration coefficients.

Images