Remote sensing image terrain correction optimization method based on deep learning

Abstract

The invention discloses a remote-sensing image terrain correction optimization method based on deep learning, and the method comprises the steps: carrying out the preprocessing of a remote-sensing image sample, inputting the preprocessed remote-sensing image sample and an SE correction image sample into an initial terrain correction model, outputting a terrain correction image sample, and obtaining a terrain correction model; the initial terrain correction model is constructed by a codec and a discriminator; respectively calculating a pixel-level loss function value and a terrain consistency loss function value according to the terrain correction image sample and the SE correction image sample, and calculating an antagonism loss function value through a discriminator; and optimizing the initial terrain correction model based on the adversarial loss function value, the pixel-level loss function value and the terrain consistency loss function value, so that the optimized terrain correction model performs terrain correction on the to-be-processed remote sensing image. The method does not need to depend on specific parameters of the image, automatically extracts the image features only through the terrain correction model of deep learning, and achieves efficient terrain correction based on the image features.

Description

Technical Field

[0001] The present invention relates to the technical field of terrain correction, and particularly to an optimization method for terrain correction of remote sensing images based on deep learning. Background Art

[0002] Terrain correction is of great significance in eliminating the interference of terrain effects on the reflectance of remote sensing images and improving the image interpretation accuracy. Due to the differences in slope, aspect, and solar illumination angle, the reflectance of ground objects in the same image varies greatly, seriously affecting the unity and accuracy of the image.

[0003] Existing terrain correction methods include empirical models, semi-empirical models, and physical models. These methods perform well under specific conditions, but there are still limitations in complex terrain or cross-regional applications. Therefore, how to achieve more stable and efficient terrain correction under complex terrain or cross-regional conditions has become an urgent problem to be solved.

[0004] The above content is only used to assist in understanding the technical solution of the present invention, and does not represent an admission that the above content is prior art. Summary of the Invention

[0005] The main purpose of the present invention is to provide an optimization method for terrain correction of remote sensing images based on deep learning, aiming to solve the technical problem of how to achieve more stable and efficient terrain correction under complex terrain or cross-regional conditions.

[0006] To achieve the above purpose, the present invention provides an optimization method for terrain correction of remote sensing images based on deep learning, and the optimization method for terrain correction of remote sensing images based on deep learning includes:

[0007] Preprocess the remote sensing image samples;

[0008] Input the preprocessed remote sensing image samples into the SE model to obtain SE corrected image samples;

[0009] Respectively use the preprocessed remote sensing image samples and the SE corrected image samples as inputs of the initial terrain correction model, and the initial terrain correction model is constructed based on an encoder-decoder and a discriminator of a generative adversarial network;

[0010] Generate terrain correction image samples through the encoder-decoder according to the preprocessed remote sensing image samples;

[0011] Calculate the pixel-level loss function value and the terrain consistency loss function value according to the terrain correction image samples and the SE corrected image samples respectively, and calculate the adversarial loss function value through the discriminator;

[0012] Optimize the initial terrain correction model based on the adversarial loss function value, the pixel-level loss function value, and the terrain consistency loss function value to obtain a remote sensing terrain correction model;

[0013] Perform terrain correction on the remote sensing image to be processed based on the remote sensing terrain correction model.

[0014] Optionally, the step of preprocessing the remote sensing image samples includes:

[0015] Perform spectral information processing on the remote sensing image samples through radiometric calibration and atmospheric correction to obtain standard spectral image samples;

[0016] Perform resampling and coordinate projection according to the DEM data corresponding to the standard spectral image samples.

[0017] Optionally, the step of generating terrain correction image samples through the codec according to the preprocessed remote sensing image samples includes:

[0018] Extract terrain local features and terrain global features from the preprocessed remote sensing image samples through an encoder based on the SE correction image samples;

[0019] Perform image reconstruction on the terrain local features and the terrain global features through a decoder to generate terrain correction image samples.

[0020] Optionally, the step of calculating the pixel-level loss function value according to the terrain correction image samples and the SE correction image samples includes:

[0021] Determine the pixel values corresponding to the terrain correction image samples;

[0022] Calculate the pixel-level loss function value through the pixel-level loss function according to the pixel values corresponding to the terrain correction image samples, the pixel values corresponding to the SE correction image samples, and the terrain correction image samples and the SE correction image samples.

[0023] Optionally, the step of calculating the terrain consistency loss function value according to the terrain correction image samples and the SE correction image samples includes:

[0024] Respectively determine the pixel slope values and pixel aspect values corresponding to the terrain correction image samples and the SE correction image samples;

[0025] Calculate the terrain consistency loss function value through the terrain consistency loss function according to the pixel slope values, the pixel aspect values, and the number of pixels between the terrain correction image samples and the SE correction image samples.

[0026] Optionally, the step of calculating the adversarial loss function value according to the terrain-corrected image sample and the SE-corrected image sample includes:

[0027] Obtain the pixel values corresponding to the SE-corrected image sample;

[0028] Based on the pixel values corresponding to the SE-corrected image sample, determine the pixel prediction probability and pixel supplementation probability corresponding to the terrain-corrected image sample, and determine the number of pixels between the terrain-corrected image sample and the SE-corrected image sample;

[0029] Based on the discriminator, calculate the adversarial loss function value through the adversarial loss function according to the pixel values corresponding to the SE-corrected image sample, the pixel prediction probability, the pixel supplementation probability, and the number of pixels.

[0030] In addition, to achieve the above object, the present invention also proposes a remote sensing image terrain correction optimization system based on deep learning. The remote sensing image terrain correction optimization system based on deep learning includes:

[0031] A processing module for preprocessing the remote sensing image sample;

[0032] The processing module is further configured to input the preprocessed remote sensing image sample into the SE model to obtain the SE-corrected image sample;

[0033] A correction module for respectively using the preprocessed remote sensing image sample and the SE-corrected image sample as inputs to an initial terrain correction model, where the initial terrain correction model is constructed based on an encoder-decoder and a discriminator of a generative adversarial network;

[0034] The correction module is further configured to generate a terrain-corrected image sample through the encoder-decoder according to the preprocessed remote sensing image sample;

[0035] An optimization module is further configured to calculate the pixel-level loss function value and the terrain consistency loss function value according to the terrain-corrected image sample and the SE-corrected image sample respectively, and calculate the adversarial loss function value through the discriminator;

[0036] The optimization module is further configured to optimize the initial terrain correction model based on the adversarial loss function value, the pixel-level loss function value, and the terrain consistency loss function value to obtain a remote sensing terrain correction model;

[0037] The correction module is further configured to perform terrain correction on the to-be-processed remote sensing image based on the remote sensing terrain correction model.

[0038] The present invention first preprocesses remote sensing image samples, and inputs the preprocessed remote sensing image samples into the SE model to enable the SE model to output SE-corrected image samples. Then, the preprocessed remote sensing image samples and the SE-corrected image samples are respectively used as inputs to the initial terrain correction model, and terrain-corrected image samples are generated through an encoder-decoder. After that, pixel-level loss function values and terrain consistency loss function values are respectively calculated based on the terrain-corrected image samples and the SE-corrected image samples, and adversarial loss function values are calculated through a discriminator. Finally, the initial terrain correction model is optimized based on the adversarial loss function values, pixel-level loss function values, and terrain consistency loss function values, so that the optimized terrain correction model performs terrain correction on the remote sensing image to be processed. Compared with the prior art, the terrain effect in complex terrain areas is highly non-linear, and it is difficult for traditional correction methods to effectively eliminate this effect through linear or quasi-linear models. In the present invention, without relying on specific parameters of the image (such as solar zenith angle, slope, aspect, etc.), the terrain correction model based on deep learning automatically extracts image features, thereby realizing efficient terrain correction based on the image features. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 FIG. is a schematic structural diagram of a remote sensing image terrain correction optimization device based on deep learning in the hardware operating environment involved in the embodiment solution of the present invention;

[0040] Figure 2 FIG. is a schematic flowchart of the first embodiment of the remote sensing image terrain correction optimization method based on deep learning of the present invention;

[0041] Figure 3 FIG. is a schematic working principle diagram of the deep learning network in the first embodiment of the remote sensing image terrain correction optimization method based on deep learning of the present invention;

[0042] Figure 4 FIG. is a comparison diagram of false color results of correction based on a deep learning model and a traditional terrain correction method in the first embodiment of the remote sensing image terrain correction optimization method based on deep learning of the present invention;

[0043] Figure 5 FIG. is a comparison diagram of true color results of correction based on a deep learning model and a traditional terrain correction method in the first embodiment of the remote sensing image terrain correction optimization method based on deep learning of the present invention;

[0044] Figure 6 FIG. is a structural block diagram of the first embodiment of the remote sensing image terrain correction optimization system based on deep learning of the present invention.

[0045] The realization, functional features, and advantages of the object of the present invention will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0046] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0047] Refer to Figure 1 , Figure 1 which is a schematic structural diagram of a remote sensing image terrain correction optimization device based on deep learning for the hardware operating environment involved in the solution of the embodiment of the present invention.

[0048] As Figure 1 shown, the remote sensing image terrain correction optimization device based on deep learning may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. Among them, the communication bus 1002 is used to realize the connection and communication between these components. The user interface 1003 may include a display screen (Display) and an input unit such as a keyboard (Keyboard). Optionally, the user interface 1003 may further include a standard wired interface and a wireless interface. The network interface 1004 may optionally include a standard wired interface and a wireless interface (such as a wireless-fidelity (Wi-Fi) interface). The memory 1005 may be a high-speed random access memory (Random Access Memory, RAM), or a stable non-volatile memory (Non-Volatile Memory, NVM), such as a disk memory. Optionally, the memory 1005 may also be a storage system independent of the foregoing processor 1001.

[0049] Those skilled in the art can understand that Figure 1 the structure shown in

[0050] As Figure 1 shown, in the memory 1005 as a storage medium, there may be included an operating system, a network communication module, a user interface module, and a remote sensing image terrain correction optimization program based on deep learning.

[0051] In Figure 1In the terrain correction optimization device for remote sensing images based on deep learning shown, the network interface 1004 is mainly used for data communication with the network server; the user interface 1003 is mainly used for data interaction with the user; the processor 1001 and the memory 1005 in the terrain correction optimization device for remote sensing images based on deep learning of the present invention can be arranged in the terrain correction optimization device for remote sensing images based on deep learning. The terrain correction optimization device for remote sensing images based on deep learning calls the terrain correction optimization program for remote sensing images based on deep learning stored in the memory 1005 through the processor 1001 and executes the terrain correction optimization method for remote sensing images based on deep learning provided by the embodiments of the present invention.

[0052] An embodiment of the present invention provides a terrain correction optimization method for remote sensing images based on deep learning. Refer to Figure 2 , Figure 2 which is a schematic flowchart of the first embodiment of the terrain correction optimization method for remote sensing images based on deep learning of the present invention.

[0053] In this embodiment, the terrain correction optimization method for remote sensing images based on deep learning includes the following steps:

[0054] Step S10: Preprocess the remote sensing image samples.

[0055] It should also be noted that the selection of remote sensing image samples needs to cover complex terrains (such as hills, mountains, and plains) and scenes with obvious illumination changes to ensure that the images contain diverse terrain features and illumination conditions. And the remote sensing image samples can be Sentinel-2 optical remote sensing images to ensure the diversity and representativeness of the data. At the same time, the 30-meter resolution digital elevation model (DEM) provided by USGS SRTM is required. The DEM data provides slope, aspect, and other terrain parameters for subsequent analysis.

[0056] In this embodiment, Xuancheng City, Anhui Province can be selected as the experimental area, covering composite landform features such as plains, hills, and mountains. The longitude and latitude range of the research area is from 118.73°E to 119.14°E, and from 29.97°N to 30.41°N. The terrain elevation difference is significant, and the coverage range is from 108 meters to 1769 meters. This area has typical terrain complexity and obvious solar radiation gradient characteristics, which is suitable for the verification of terrain correction methods.

[0057] In specific implementation, the method of preprocessing the remote sensing image samples to obtain the preprocessed remote sensing image samples is to perform spectral information processing on the remote sensing image samples through radiometric calibration and atmospheric correction to obtain standard spectral image samples; determine the slope and aspect according to the DEM data corresponding to the standard spectral image samples; perform resampling and coordinate projection on the DEM data based on the slope and aspect to obtain the preprocessed remote sensing image samples.

[0058] Spectral information includes digital values, atmospheric scattering, and aerosols in the image.

[0059] Preprocessing includes radiometric calibration, atmospheric correction, terrain feature selection, and DEM data resampling. The purpose of radiometric calibration is to convert the digital values (DN values) in the image into radiance or reflectance values with physical significance; the purpose of atmospheric correction is to remove the influence of atmospheric effects and obtain surface reflectance; the purpose of terrain feature selection is to ensure that the study area can represent different complex terrain conditions; the purpose of DEM data resampling is to ensure that the terrain data is consistent with the spatial resolution of the remote sensing image.

[0060] In the specific implementation, the Google Earth Engine (GEE) platform is used to perform radiometric calibration and atmospheric correction on Sentinel-2 images. Radiometric calibration converts the image data into reflectance values through linear scaling, providing a consistent radiation basis for subsequent processing. Atmospheric correction uses the Sen2Cor algorithm built into Sentinel-2 to remove the influence of atmospheric scattering and aerosols, ensuring the spectral consistency of the image and providing accurate input for terrain correction.

[0061] The GEE platform is used to calculate multiple terrain features, such as slope, aspect, and terrain ruggedness index (TRI). Based on these features, the study area is divided into: hilly and low mountain terrain (slope between 15° and 30°); typical steep mountain terrain (local slope greater than 30°); low-relief flat areas (slope less than 15°). This division ensures that the selected area can cover typical steep mountains, gentle hills, and low mountain terrains. By dividing the area according to this criterion and selecting specific experimental areas for testing, it is ensured that the selected area has rich terrain complexity and lighting changes. This division method enables the study to evaluate the performance and adaptability of deep learning models under complex terrain conditions.

[0062] The USGS SRTM DEM is resampled from 30m resolution to 10m using bilinear interpolation to improve the spatial accuracy and reduce the impact of scale differences on the terrain correction accuracy. At the same time, the resampled DEM is reprojected using the WGS84 coordinate reference system (EPSG:4326) to ensure its consistency with the coordinates of the remote sensing image.

[0063] In this embodiment, the remote sensing image data used in the research comes from the Sentinel-2A satellite, and the image acquisition time is December 8, 2019. The multi-spectral data of Sentinel-2A has a spatial resolution of 10-60 meters, and the band coverage range includes from visible light to short-wave infrared (443nm to 2190nm). The GEE platform is used to perform radiometric calibration on the Sentinel-2 image, multiply the data by 0.0001 for standardization, and convert it into reflectance values; subsequently, the Sen2Cor algorithm is used for atmospheric correction to remove the influence of atmospheric scattering and aerosols, ensure the spectral consistency of the image, and provide accurate input for terrain correction. Then, the USGS SRTM digital elevation model (DEM) is used to calculate the slope and aspect to characterize the terrain undulation and its influence on lighting conditions; in the research, the bilinear interpolation method is used to resample the USGS SRTM DEM from 30m to 10m, and then the resampled DEM is re-projected using the WGS84 coordinate reference system to ensure its consistency with the Sentinel-2 image coordinates.

[0064] Using the pre-processed remote sensing images, terrain correction based on the SE (Statistical-Empirical) model is implemented on the Google Earth Engine (GEE) platform. The correction process includes calculation of lighting factors, calculation of regression correction coefficients, and terrain correction of the SE model, to construct a standardized data set suitable for deep learning network training. The cropping size of the data set in this case is set to 512×512 pixels to ensure that the image blocks contain rich terrain information and can effectively reduce computational redundancy. The cropped image data is stored in 16-bit integer format to ensure the consistency of the data format and optimize the numerical range of the model input.

[0065] Step S20: Input the pre-processed remote sensing image samples into the SE model to obtain SE-corrected image samples.

[0066] Furthermore, the specific implementation method of inputting the pre-processed remote sensing image samples into the SE model (semi-empirical model) to enable the SE model to output SE-corrected image samples is as follows:

[0067] To achieve efficient terrain correction, the pre-processed remote sensing image samples are used as input data, including band information such as blue light, green light, and near-infrared, as input data for the deep learning model to use for training; the SE model using a physical-statistical hybrid method is used for terrain correction, and the corrected image is used as the reference data set.

[0068] Calculation of lighting factors: Extract the solar azimuth angle and solar zenith angle θ z of the pre-processed remote sensing image samples, and calculate the slope θ of the study areas and slope aspect Then, calculate the local illumination factor (IL) using the following formula:

[0069]

[0070] SE terrain correction implementation: To reduce the influence of terrain undulation on the image reflectance, first perform a linear calculation on each spectral band (B2, B3, B4, B8, B11, B12) in the preprocessed remote sensing image samples. Take the illumination factor (IL) as the independent variable and the original reflectance ρ T as the dependent variable, and obtain the two correction coefficients of the regression equation by finding the regression relationship: the slope a and the intercept b. Then, combine the calculated correction coefficients a and b, and combine the illumination factor IL (cosΥ i ) to correct each spectral band (B2, B3, B4, B8, B11, B12), reducing the influence of terrain undulation on the image reflectance. The SE correction formula:

[0071]

[0072] In the above formula, ρ SE represents the surface reflectance after terrain correction; ρ T represents the reflectance of the original inclined surface; represents the average value of the reflectance of the original inclined surface; θ s represents the slope; represents the terrain slope aspect; IL represents the illumination factor, i.e., cosΥ i ; a and b respectively represent the slope and intercept of the linear fitting regression equation.

[0073] The SE-corrected SE-corrected image samples include information such as terrain features and illumination effects. Therefore, the deep learning model (i.e., the terrain correction model) can automatically extract image-related information through an end-to-end learning process without relying on fixed terrain parameter inputs.

[0074] Step S30: Respectively use the preprocessed remote sensing image samples and the SE-corrected image samples as inputs to the initial terrain correction model, and the initial terrain correction model is constructed based on an encoder-decoder and a discriminator of a generative adversarial network.

[0075] Reference Figure 3 , Figure 3This is the working principle diagram of the deep learning network for the first embodiment of the remote sensing image terrain correction optimization method based on deep learning. The initial terrain correction model based on deep learning is constructed based on an encoder, a decoder, and a discriminator of a generative adversarial network. The encoder-decoder structure is used as the generator of the generative adversarial network to perform terrain correction on the uncorrected Sentinel-2 image (i.e., the preprocessed remote sensing image sample). The SE-corrected image is used as a supervision signal to guide the model learning (i.e., providing the SE-corrected image sample to the discriminator of the generative adversarial network as the true reference). The encoder extracts terrain features through multiple layers of convolution and enhances multi-scale information fusion through skip connections; the decoder uses transposed convolution to gradually restore the image resolution, ensuring illumination consistency and accurate reconstruction of terrain features, and obtaining the generated image (i.e., the terrain correction image sample). The generated image maintains the original bands of Sentinel-2 and the 16-bit integer format to ensure data consistency.

[0076] Step S40: Generate a terrain correction image sample through the codec according to the preprocessed remote sensing image sample.

[0077] Furthermore, the processing method of generating a terrain correction image sample through the codec according to the preprocessed remote sensing image sample is to extract local terrain features and global terrain features from the preprocessed remote sensing image sample through the encoder; perform image reconstruction on the local terrain features and the global terrain features through the decoder to generate a terrain correction image sample.

[0078] In specific implementation, the encoder-decoder structure is utilized and follows the basic principle of a multi-layer convolutional neural network. Through the hierarchical feature extraction and detail reconstruction mechanism, this architecture can effectively capture the multi-scale terrain features of the image.

[0079] Multi-scale feature extraction: The model encoder consists of 8 layers of convolutional networks. The number of convolutional kernels increases layer by layer, and the multi-scale features of the image are extracted by reducing the spatial resolution. Specific design: 64 convolutional kernels are used in the first layer of the encoder, and the number of convolutional kernels increases in each subsequent layer, finally reaching 512 convolutional kernels in the eighth layer. This design can capture details and local features at low levels while extracting global features at high levels. Each layer of convolution operation uses a 4×4 convolutional kernel with a stride of 2 to gradually reduce the spatial resolution of the image.

[0080] Image Detail Reconstruction: The decoder gradually restores the spatial resolution of the image through deconvolution operations to ensure the integrity of details at different scales. The decoder structure consists of 8 deconvolution layers, and each layer has a skip connection with the feature map of the corresponding layer in the encoder, directly transmitting the multi-scale features extracted from each layer to the corresponding layer in the decoder. While restoring the spatial resolution, the deconvolution layers gradually reduce the number of channels, from 512 channels at the top layer to 1 channel at the output layer, completing the restoration from global features to detailed features and ensuring the hierarchical reconstruction process of the image.

[0081] Step S50: Calculate the pixel-level loss function value and the terrain consistency loss function value based on the terrain-corrected image sample and the SE-corrected image sample respectively, and calculate the adversarial loss function value through the discriminator.

[0082] The quality of the corrected image is quantitatively evaluated by designing a set of loss functions to ensure high-precision requirements in terms of detail restoration and illumination consistency. Taking the discriminator module of the generative adversarial network (GAN) as the core, combining three loss functions: adversarial loss (BCE Loss), pixel-level loss (L1 Loss), and terrain consistency loss (Terrain Loss), comprehensively evaluates the global and local quality of the image.

[0083] The discriminator of the generative adversarial network constrains the spectral authenticity of the generated image through the adversarial loss (BCE Loss), optimizes the spectral accuracy with the pixel-level loss (L1 Loss), and ensures the stability of the slope and aspect features of the image with the terrain consistency loss (Terrain Loss), thereby reducing the influence of terrain effects.

[0084] Furthermore, the processing method for calculating the adversarial loss function value based on the terrain-corrected image sample and the SE-corrected image sample is to obtain the pixel values corresponding to the SE-corrected image sample; determine the pixel prediction probability and pixel supplementation probability corresponding to the terrain-corrected image sample based on the pixel values corresponding to the SE-corrected image sample, and determine the number of pixels between the terrain-corrected image sample and the SE-corrected image sample; calculate the adversarial loss function value through the adversarial loss function based on the pixel values corresponding to the SE-corrected image sample, pixel prediction probability, pixel supplementation probability, and the number of pixels by the discriminator of the generative adversarial network.

[0085] The adversarial loss function (BCE Loss) is a loss function used to measure the similarity of the distribution between the generated image and the target image, guiding the generator to produce an output closer to the target image. Its significance lies in optimizing the global distribution of the image, ensuring the authenticity of the overall spectral features of the generated image, and helping the generator learn the correct spectral information.

[0086]

[0087] Wherein, N is the number of pixels between the terrain-corrected image sample and the SE-corrected image sample; yi is the i-th pixel value of the SE-corrected image sample, indicating whether it is a real image, with a value range of [0, 1]; p(yi) is the predicted probability corresponding to the i-th pixel value of the generated image; 1 - p(yi) is the complementary probability corresponding to the i-th pixel value of the generated image, representing the probability of not being the target image.

[0088] Furthermore, the method for calculating the pixel-level loss function value based on the terrain-corrected image sample and the SE-corrected image sample is to determine the pixel value corresponding to the terrain-corrected image sample, and calculate the pixel-level loss function value through the pixel-level loss function according to the pixel value corresponding to the terrain-corrected image sample, the pixel value corresponding to the SE-corrected image sample, and the number of pixels.

[0089] The pixel-level loss function (L1 Loss) is a loss function used to measure the pixel-by-pixel difference between the generated image and the target image. By minimizing the absolute error, it ensures that the generated image is consistent with the target image in terms of brightness, texture, and details. In addition, on the basis of the BCE Loss function, the L1 Loss function further optimizes the local pixel-level error to ensure that the spectral values, brightness, and details of the generated image are more accurate.

[0090]

[0091] N is the total number of pixels (i.e., the number of pixels) in the terrain-corrected image sample and the SE-corrected image sample; G(xi) is the i-th pixel value of the generated image; yi is the i-th pixel value of the SE-corrected image sample.

[0092] Furthermore, the method for calculating the terrain consistency loss function value based on the terrain-corrected image sample and the SE-corrected image sample is to separately determine the pixel slope value and the pixel aspect value corresponding to the terrain-corrected image sample and the SE-corrected image sample; and calculate the terrain consistency loss function value through the terrain consistency loss function according to the pixel slope value, the pixel aspect value, and the number of pixels.

[0093] The terrain consistency loss function (Terrain Loss) is a loss function designed for terrain feature correction, which is used to ensure the consistency between the generated image and the target image in terms of terrain details and morphology. As the final constraint, Terrain Loss ensures that the terrain features of the generated image, such as slope and aspect, are consistent with the target image, thereby improving the accuracy of terrain correction.

[0094]

[0095] N is the total number of pixels in the image; S G (x i ) and S y (xi ) respectively represent the slope values corresponding to the i-th pixel of the generated image and the SE-corrected image sample; A G (x i ) and A y (x i ) respectively represent the aspect values corresponding to the i-th pixel of the generated image and the SE-corrected image sample; α and β are weight parameters that control the influence ratio of the slope and aspect loss terms.

[0096] Step S60: Optimize the initial terrain correction model based on the adversarial loss function value, the pixel-level loss function value, and the terrain consistency loss function value to obtain a remote sensing terrain correction model.

[0097] It should also be noted that through collaborative optimization, these three loss functions gradually improve the image quality layer by layer. BCELoss ensures the consistency of the global spectral distribution, L1 Loss constrains the pixel-level error, and Terrain Loss ensures the consistency of terrain features. The three work together to optimize the image.

[0098] During the training process, the discriminator improves its ability to distinguish the authenticity of the image. The generator (constructed by the encoder-decoder) optimizes the correction strategy under the constraints of multiple loss functions, and adjusts the parameters of the encoder-decoder through backpropagation and the Adam optimizer to improve the terrain adaptability and correction accuracy of the remote sensing terrain correction model, and finally generates a high-quality corrected image with consistent spectrum and stable terrain features.

[0099] Step S70: Perform terrain correction on the remote sensing image to be processed based on the remote sensing terrain correction model.

[0100] In this embodiment, preprocess the remote sensing image to be processed, input the preprocessed remote sensing image into the remote sensing terrain correction model, and output the corrected remote sensing image through the encoder-decoder.

[0101] It should also be noted that the preprocessed remote sensing image can be input into the SE model to obtain the SE-corrected remote sensing image. Then, input the SE-corrected remote sensing image into the remote sensing terrain correction model, calculate the loss function values (i.e., the adversarial loss function value, the pixel-level loss function value, and the terrain consistency loss function value) based on the corrected remote sensing image and the SE-corrected remote sensing image, optimize the remote sensing terrain correction model through the loss function values, and use the optimized remote sensing terrain correction model for the next image terrain correction.

[0102] Correction effect evaluation:

[0103] By comparing the experimental results of the pre - calibration image, the image of the traditional calibration method SE model, and the calibration image of this embodiment, the calibration effect is comprehensively verified from four aspects: visual analysis, within - class interquartile range reduction rate (IQRR), root - mean - square error and mean absolute error (RMSD and MAE), and correlation analysis. At the same time, the possible local errors are optimized.

[0104] (1) Visual analysis: Visual analysis mainly evaluates the spectral consistency, brightness balance, and the elimination effect of terrain effects of the images under different terrain conditions by comparing the original image, the image calibrated by the SE model, and the image calibrated by the method proposed in this study. Special attention is paid to whether the spectral information in the high - slope area of the calibrated image tends to be uniform, and whether there are phenomena of over - calibration or information loss in the image.

[0105] In this embodiment, referring to Figure 4 and Figure 5 , Figure 4 is a comparison chart of false - color results of calibration based on a deep - learning model and a traditional terrain - correction method in the first embodiment of the method for optimizing terrain correction of remote - sensing images based on deep learning of the present invention; Figure 5 is a comparison chart of the results of calibration based on a deep - learning model and a traditional terrain - correction method in the first embodiment of the method for optimizing terrain correction of remote - sensing images based on deep learning of the present invention; The comparison results of images before and after calibration using different terrain - correction methods, including various traditional methods and the calibration method D - GTCM based on a deep - learning model, highlight the different performances before and after calibration of each method.

[0106] (2) Within - class interquartile range reduction rate (IQRR): The reduction of within - class radiance variability is an important indicator for evaluating the terrain - correction effect. The change of terrain usually leads to differences in radiance in different regions of the image. Especially under high - slope or complex terrain conditions, the radiance consistency of the image is often significantly affected by terrain effects. In this study, the interquartile range reduction rate (IQRR) is introduced as a new quantitative index to measure the improvement of radiance consistency in the image after terrain correction. The interquartile range (IQR) is an important statistic reflecting the data distribution range, defined as the difference between the upper quartile (Q3) and the lower quartile (Q1), representing the variability of the data. After terrain correction, the radiance consistency of the image should be significantly improved, resulting in a reduction of the within - class interquartile range. Therefore, IQRR is defined as the change amount of the interquartile range of the image before and after calibration, and the calculation formula:

[0107]

[0108] Ideally, terrain correction should enhance the radiometric consistency of the image, manifested as a significant reduction in the interquartile range, thereby significantly increasing the IQRR value. When the IQRR value is greater than 20%, it generally indicates that the correction method effectively reduces the impact of terrain on the image and the correction effect is good. On the contrary, if the IQRR value is negative or close to 0, it indicates that the correction effect is poor, possibly failing to effectively remove the interference of terrain on the image or even exacerbating the terrain effect in some bands.

[0109] It should also be noted that Table 1 shows the IQRR values of different terrain correction methods in each band. The Variable Empirical Coefficient Algorithm (VECA), C-Correction, Empirical Rotation Correction (ER), Sun-Canopy-Sensor + C Correction (SCS + C), Statistical-Empirical Correction (SE), and Deep-GAN Topographic Correction Method (D-GTCM) are compared. From the comparison results, the terrain correction model (D-GTCM) shows obvious advantages in multiple bands, especially in the blue, green, and near-infrared bands. In the blue band, the IQRR value of D-GTCM is 22.93, higher than other traditional methods, indicating its superiority in reducing the variability of image radiance and effectively removing the spectral inconsistency caused by terrain effects. The IQRR value in the green band is 36.53, showing an advantage compared to other methods, further proving the good correction effect of this model in this band. In the near-infrared band, the IQRR value of D-GTCM is 46.48, superior to the traditional methods SE and ER, demonstrating higher correction efficiency and reflecting its adaptability to complex terrain.

[0110] In the SWIR1 and SWIR2 bands, D-GTCM shows stable and effective correction effects, especially the advantage in the SWIR1 band is particularly prominent, demonstrating its superior performance in the infrared band. Overall, the IQRR values of D-GTCM in multiple bands are relatively high, indicating that it can effectively reduce the impact of terrain effects during the correction process, maintain the spectral consistency of the image, and reflect the advantages of a deep learning model.

[0111] Table 1 shows the IQRR values of different terrain correction methods in each band

[0112] Calibration method Red light Green light Blue light Near infrared Short-wave infrared 1 Short-wave infrared 2 VECA 22.39 30.59 17.73 43.58 35.68 26.18 C 17.79 27.85 16.31 38.44 28.55 17.79 ER 25.77 31.51 18.44 47.89 42.82 35.20 SCS+C 22.70 29.68 17.02 42.00 33.25 23.76 SE 26.07 31.96 18.44 47.97 42.82 35.07 D-GTCM 30.37 36.53 22.93 46.48 50.21 38.06

[0113] (3) Root Mean Square Error (RMSD) and Mean Absolute Error (MAE): RMSD (Root Mean Square Deviation) and MAE (Mean Absolute Error) are important statistical indicators for measuring the image correction effect. RMSD measures the overall error between the corrected image and the target image. The smaller the value, the better the correction effect. MAE measures the average error of each pixel, and the smaller the value, the better the correction effect. These two indicators are often used to quantitatively evaluate the accuracy of image correction algorithms, especially in terms of image details and spectral consistency. The calculation formulas for RMSD and MAE are as follows:

[0114]

[0115] Table 2 shows the Root Mean Square Error (RMSD) values for each band

[0116] Calibration method Red light Green light Blue light Near infrared Short-wave infrared 1 Short-wave infrared 2 VECA 163.45 103.42 163.45 1315.97 1836.28 1409.71 C 168.06 106.30 168.06 1377.03 1927.45 1483.82 ER 146.53 104.89 146.53 564.94 591.88 363.01 SCS+C 154.42 100.18 154.42 1202.13 1706.05 1306.94 SE 142.34 101.93 142.34 548.43 574.56 352.43 D-GTCM 140.98 120.36 140.98 474.08 507.87 310.23

[0117] Table 3 shows the Mean Absolute Error (MAE) values for each band

[0118] Calibration method Red light Green light Blue light Near infrared Short-wave infrared 1 Short-wave infrared 2 VECA 119.66 84.78 50.36 539.06 624.43 391.04 C 119.21 84.57 50.27 543.24 632.26 396.58 ER 119.11 85.20 50.63 459.17 481.13 295.10 SCS+C 116.06 82.66 49.22 517.14 595.92 371.45 SE 119.06 85.17 50.65 458.70 480.65 294.86 D-GTCM 112.17 97.96 55.80 373.93 400.20 242.25

[0119] Judging from the calculation results of RMSD and MAE in Table 2 and Table 3, D-GTCM shows advantages in multiple bands, especially in the Near Infrared (NIR), SWIR1, and SWIR2 bands. Since these bands are particularly sensitive to terrain changes, especially in complex terrains with a mixture of hills, mountains, and plains, terrain undulations and illumination changes usually cause reflectance fluctuations. D-GTCM can effectively reduce the influence of terrain effects on images and improve the spectral consistency of these bands.

[0120] In the red, green, and blue bands, D-GTCM shows balanced correction effects. Especially in the red and green bands, D-GTCM significantly reduces RMSD and MAE, outperforming traditional methods and being able to effectively reduce reflectance fluctuations caused by terrain and improve spectral consistency. However, in the blue band, the performance of D-GTCM is relatively inferior, with an RMSD of 71.94 and an MAE of 55.80, slightly higher than other methods, probably because the blue band is more sensitive to atmospheric and illumination changes.

[0121] In the Near Infrared (NIR) band, the RMSD value of D-GTCM is 474.08 and the MAE value is 373.93, significantly lower than traditional methods, indicating that the correction effect in the NIR band is better than traditional methods. Especially in areas with large terrain undulations such as hills and mountains, the reflectance in the NIR band is greatly affected by vegetation, terrain, and illumination. D-GTCM can effectively reduce these effects and improve image quality.

[0122] In the SWIR1 and SWIR2 bands, D-GTCM shows advantages. The RMSD value in the SWIR1 band is 507.87, and the MAE is 400.20. The SWIR band is sensitive to the reflection characteristics of water, vegetation, and soil, and is greatly affected by terrain and illumination changes. D-GTCM can effectively remove these effects and demonstrate superior calibration capabilities.

[0123] Overall, the superior performance of D-GTCM in multiple bands verifies its effectiveness in complex terrain environments, demonstrating the advantages of this method in improving image quality and spectral consistency.

[0124] (4) Correlation analysis: In terrain correction, the key index for the image correction effect is the correlation between the reflectance and the solar incidence angle. Ideally, for an effectively corrected image, its reflectance should no longer significantly depend on the solar incidence angle, and the correlation should be close to zero. Terrain effects usually cause the spectral characteristics in different regions of the image to fluctuate with the change of the solar angle. Effective terrain correction should reduce or eliminate this effect.

[0125] To quantify this relationship, regression analysis is often used to calculate the slope and correlation coefficient (R) between the image reflectance and the solar incidence angle. The slope reflects the strength of the linear relationship between the two. Ideally, the slope of the corrected image should be close to 0, that is, the reflectance no longer changes significantly with the solar incidence angle; the R value represents the goodness of fit, and the closer the value is to 0, the better the image correction effect.

[0126] Table 4 shows the slope values between the reflectance of each band and the solar incidence angle

[0127] Calibration method Red light Green light Blue light Near infrared Short-wave infrared 1 Short-wave infrared 2 Original 488.02 369.27 208.40 2149.28 2240.97 1345.94 VECA 34.34 16.03 -0.35 312.62 332.90 218.44 C 36.38 16.74 -0.30 341.87 369.42 243.03 ER -27.13 -2.60 -7.37 169.87 132.05 47.48 SCS+C 39.75 20.12 2.03 303.15 323.91 217.14 SE -27.95 -3.06 -7.40 169.95 132.05 48.12 D-GTCM 111.71 58.23 3.75 366.81 295.36 281.61

[0128] Table 5 shows the correlation coefficient (R) values between the reflectance of each band and the solar incidence angle

[0129] Calibration method Red light Green light Blue light Near infrared Short-wave infrared 1 Short-wave infrared 2 Original 0.54 0.62 0.52 0.74 0.73 0.65 VECA 0.05 0.04 0.00 0.16 0.16 0.14 C 0.05 0.04 0.00 0.16 0.16 0.14 ER -0.04 -0.01 -0.02 0.09 0.07 0.03 SCS+C 0.06 0.05 0.01 0.15 0.15 0.14 SE -0.04 -0.01 -0.02 0.09 0.07 0.03 D-GTCM 0.17 0.14 0.01 0.20 0.17 0.20

[0130] Tables 4 and 5 show the results of the correlation analysis of the slope and correlation coefficient R values. The results show that D-GTCM significantly reduces the correlation between the image reflectance and the solar incidence angle, and the correction effect is better. Especially in the blue band, the slope and R value of D-GTCM are lower, and the correction effect is relatively ideal. However, in the SWIR1 and SWIR2 bands, the correction effect is relatively weak, and the slope and R value are still relatively high, indicating that there is still room for improvement in the correction effect of this method in these bands.

[0131] Generally speaking, the correction effect of D-GTCM is relatively balanced in the six bands. The slope and R values in different bands are relatively close, and there is no serious overcorrection phenomenon, reflecting the global adaptability and stability of this method. This result demonstrates the comprehensive advantages of D-GTCM for full-band correction under complex terrain.

[0132] In actual operation, the solar incidence angle has a certain impact on the image reflectance. Especially in areas with large topographic undulations, the reflectance difference between the sunny slope and the shady slope is relatively large, which may lead to a high correlation. By analyzing the change in the correlation of the images before and after correction, it is possible to evaluate whether the correction method can effectively reduce the topographic effect and improve the spectral consistency of the images.

[0133] Model calculation efficiency:

[0134] Quantitatively analyze the calculation efficiency of the deep learning model in the terrain correction task from three aspects: training calculation cost, model convergence speed, and computational resource occupancy. During the training process, monitor the training time, GPU utilization, and video memory occupancy of each epoch to evaluate resource consumption; use a fixed batch size and set the learning rates of the generator and discriminator to ensure the stability of gradient updates and improve the optimization efficiency.

[0135] To ensure training stability, monitor the gradient change of the loss function, and combine the weighted optimization strategy of L1 pixel-level loss, BCE adversarial loss, and Terrain Loss topographic consistency loss to improve the convergence of the model. At the end of each epoch, use the test dataset for inference, measure the inference time and computational latency, and evaluate the real-time performance of the model in different computing environments to provide a reference for computational efficiency for subsequent optimization.

[0136] Model generalization test:

[0137] Adopt a cross-region verification method to evaluate the adaptability of the deep learning model in different terrain environments to ensure that the terrain correction method has good generalization ability. During the training process, the model learns the lighting characteristics of typical terrain regions and makes adjustments according to different terrain differences. To verify the applicability of the model under various terrain conditions such as plains, hills, and mountains, select representative remote sensing images for testing and evaluate their correction effects.

[0138] The evaluation indicators of the generalization test include spectral consistency, terrain adaptability, and image radiation stability to ensure that the model can maintain a stable correction effect under different terrain conditions. By comparing the performance in different terrain regions, analyze the applicability of the model in complex terrains and verify the cross-region promotion potential of the method.

[0139] In this embodiment, first, preprocess the remote sensing image samples, input the preprocessed remote sensing image samples into the SE model to obtain SE-corrected image samples, and then use the preprocessed remote sensing image samples and the SE-corrected image samples as inputs to the initial terrain correction model respectively. Based on the SE-corrected image samples, generate terrain-corrected image samples through an encoder-decoder. The initial terrain correction model is constructed based on an encoder, a decoder, and a discriminator of a generative adversarial network. Then, calculate the adversarial loss function value, pixel-level loss function value, and terrain consistency loss function value according to the terrain-corrected image samples and the SE-corrected image samples respectively. Finally, optimize the initial terrain correction model based on the adversarial loss function value, pixel-level loss function value, and terrain consistency loss function value, so that the optimized terrain correction model performs terrain correction on the to-be-processed remote sensing image. Compared with the prior art where the terrain effect in complex terrain areas is highly non-linear and traditional correction methods are difficult to effectively eliminate this effect through linear or quasi-linear models, in this embodiment, it is not necessary to rely on specific parameters of the image (such as solar zenith angle, slope, aspect, etc.), and the terrain correction model based on deep learning automatically extracts image features, thereby achieving efficient terrain correction based on the image features.

[0140] Refer to Figure 6 , Figure 6 which is the structural block diagram of the first embodiment of the remote sensing image terrain correction optimization system based on deep learning of the present invention.

[0141] As Figure 6 shown, the remote sensing image terrain correction optimization system based on deep learning proposed in the embodiment of the present invention includes:

[0142] A processing module 6001 for preprocessing remote sensing image samples;

[0143] The processing module 6001 is further configured to input the preprocessed remote sensing image samples into the SE model to obtain SE-corrected image samples;

[0144] A correction module 6002 for using the preprocessed remote sensing image samples and the SE-corrected image samples as inputs to the initial terrain correction model respectively, where the initial terrain correction model is constructed based on an encoder-decoder and a discriminator of a generative adversarial network;

[0145] The correction module 6002 is further configured to generate terrain-corrected image samples through the encoder-decoder according to the preprocessed remote sensing image samples;

[0146] An optimization module 6003 for calculating the pixel-level loss function value and the terrain consistency loss function value according to the terrain-corrected image samples and the SE-corrected image samples respectively, and calculating the adversarial loss function value through the discriminator;

[0147] The optimization module 6003 is further configured to optimize the initial terrain correction model based on the adversarial loss function value, the pixel-level loss function value, and the terrain consistency loss function value to obtain a remote sensing terrain correction model;

[0148] The correction module 6002 is further configured to perform terrain correction on the to-be-processed remote sensing image based on the remote sensing terrain correction model.

[0149] For other embodiments or specific implementation manners of the remote sensing image terrain correction optimization system based on deep learning of the present invention, reference may be made to the above method embodiments, which will not be elaborated herein.

[0150] It should be noted that in this article, the term "including", "comprising" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or further includes elements inherent to such process, method, article or system. Without more limitations, an element defined by the statement "including a..." does not exclude the existence of additional identical elements in the process, method, article or system including the element.

[0151] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages and disadvantages of the embodiments.

[0152] Through the description of the above embodiments, those skilled in the art can clearly understand that the above embodiment methods can be implemented by means of software plus a necessary general hardware platform. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation manner. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. The computer software product is stored in a storage medium (such as a read-only memory / random access memory, magnetic disk, optical disk), and includes several instructions for causing a terminal device (which may be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods described in various embodiments of the present invention.

[0153] The above are only the preferred embodiments of the present invention, and do not limit the patent scope of the present invention accordingly. Any equivalent structure or equivalent process transformation made by using the description of the present invention and the accompanying drawings, or directly or indirectly applied to other related technical fields, shall be included in the patent protection scope of the present invention in the same way.

Claims

1. An optimized method for terrain correction of remote sensing images based on deep learning, characterized in that, Including the following steps: Preprocess the remote sensing image samples; Input the preprocessed remote sensing image samples into the SE model to obtain SE corrected image samples; Respectively use the preprocessed remote sensing image samples and the SE corrected image samples as the inputs of the initial terrain correction model, where the initial terrain correction model is constructed based on an encoder-decoder and the discriminator of a generative adversarial network; Generate terrain correction image samples through the encoder-decoder according to the preprocessed remote sensing image samples; Calculate the pixel-level loss function value and the terrain consistency loss function value according to the terrain correction image samples and the SE corrected image samples respectively, and calculate the adversarial loss function value through the discriminator; Optimize the initial terrain correction model based on the adversarial loss function value, the pixel-level loss function value and the terrain consistency loss function value to obtain a remote sensing terrain correction model; Perform terrain correction on the to-be-processed remote sensing image based on the remote sensing terrain correction model.

2. The method according to claim 1, characterized in that, The step of preprocessing the remote sensing image samples includes: Perform spectral information processing on the remote sensing image samples through radiometric calibration and atmospheric correction to obtain standard spectral image samples; Perform resampling and coordinate projection according to the DEM data corresponding to the standard spectral image samples.

3. The method according to claim 1 or 2, characterized in that, The step of generating terrain correction image samples through the encoder-decoder according to the preprocessed remote sensing image samples includes: Extract terrain local features and terrain global features from the preprocessed remote sensing image samples through the encoder; Perform image reconstruction on the terrain local features and the terrain global features through the decoder to generate terrain correction image samples.

4. The method according to claim 3, wherein The step of calculating the pixel-level loss function value according to the terrain correction image samples and the SE corrected image samples includes: Determine the pixel values corresponding to the terrain correction image samples; Calculate the pixel-level loss function value through the pixel-level loss function according to the pixel values corresponding to the terrain correction image samples, the pixel values corresponding to the SE corrected image samples, and the number of pixels between the terrain correction image samples and the SE corrected image samples.

5. The method according to claim 3, wherein The step of calculating the terrain consistency loss function value according to the terrain correction image samples and the SE corrected image samples includes: Respectively determine the pixel slope values and pixel aspect values corresponding to the terrain correction image samples and the SE corrected image samples; Calculate the terrain consistency loss function value through the terrain consistency loss function according to the pixel slope values, the pixel aspect values, and the number of pixels between the terrain correction image samples and the SE corrected image samples.

6. The method according to claim 3, characterized in that The step of calculating the adversarial loss function value through the discriminator according to the terrain correction image samples and the SE corrected image samples includes: Obtain the pixel values corresponding to the SE corrected image samples; Determine the pixel prediction probability and pixel supplementation probability corresponding to the terrain correction image samples based on the pixel values corresponding to the SE corrected image samples, and determine the number of pixels between the terrain correction image samples and the SE corrected image samples; Based on the discriminator, the adversarial loss function value is calculated through the adversarial loss function according to the pixel values corresponding to the SE corrected image samples, the pixel prediction probability, the pixel supplementation probability, and the number of pixels.

Images