Accumulated snow depth downscaling method combining ResNet and accumulated snow attenuation curve

Abstract

The invention relates to the technical field of accumulated snow depth downscaling, in particular to an accumulated snow depth downscaling method combining ResNet and an accumulated snow attenuation curve. Acquiring accumulated snow data of the target area and preprocessing the accumulated snow data to obtain a first accumulated snow depth image, a first auxiliary depth image and a second auxiliary depth image; inputting the first snow depth image and the first auxiliary depth image into a model for training, and obtaining a first snow prediction image and a second snow prediction image by using the trained model; obtaining a fine-scale residual value of the first snow depth image and the first snow prediction image, and correcting the second snow prediction image to obtain downscaling snow depth data; performing secondary correction on the downscaling accumulated snow depth data by using the accumulated snow attenuation curve to obtain optimal accumulated snow depth data; and performing accumulated snow depth mapping according to the optimal accumulated snow depth data. According to the invention, accurate snow depth downscaling can be carried out on snow depth data of a large-scale area, and efficient snow depth prediction is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of snow depth downscaling, and specifically relates to a snow depth downscaling method combining ResNet and a snow attenuation curve. Background Art

[0002] Snow cover is the most widely distributed and seasonally variable natural cover on the earth, and is also one of the most sensitive indicators of climate change; due to its high surface reflectivity and low thermal conductivity, snow cover plays a crucial role in the energy budget and radiation balance of the earth's surface. Snow depth is one of the important variables describing the snow accumulation in a region, and is a key parameter indispensable for research such as global climate change and hydrological cycle. In particular, the seasonal snow cover in high mountain areas is mostly the source of rivers, and plays a key role in providing water resources supply and ecological security for the population. However, global warming is accelerating the melting of glaciers and snow cover. Excessive meltwater not only exerts great pressure on the ecosystem, but also increases the frequency of natural disasters including floods, thus posing a serious threat to human safety and the sustainability of the livestock industry. Therefore, accurately measuring and analyzing snow depth is crucial for understanding regional water resources, climate change dynamics and disaster prediction.

[0003] Traditional snow cover observation methods mainly rely on automatic weather stations and manual field observations to obtain accurate snow depth data. However, these conventional observation methods have significant limitations due to the limitations of human factors and geographical environment: the distribution of observation stations is uneven, the spatial coverage is limited, and there are gaps in the time series. These limitations restrict the comprehensive understanding of snow cover dynamics at large regional scales and also impede the in-depth analysis of regional climate change and water resources management. To overcome these limitations, the development of remote sensing technology provides a new perspective, which can obtain continuous, large-scale snow cover information at a relatively low cost, especially showing unique advantages in difficult-to-reach areas with extreme climates and complex topographies. Hyperspectral and high spatio-temporal resolution optical remote sensing sensors can effectively identify the snow cover distribution range and time variation trend, and have been widely used for snow cover area monitoring. However, since the visible and infrared bands are difficult to effectively penetrate clouds, and snow cover and clouds have similar reflection characteristics, this limits the application of optical remote sensing in snow depth inversion. In contrast, passive microwave remote sensing technology makes it possible to monitor snow cover under all-weather conditions due to its strong penetration and cloud-free constraints. Especially in high-latitude and high-altitude regions, due to continuous cloud cover and frequent extreme weather conditions, passive microwave remote sensing has become an ideal choice for snow cover monitoring in these regions and is one of the most effective means of snow depth inversion currently.

[0004] Currently, estimating large-scale snow depth mainly relies on the microwave radiation signals naturally emitted by the snow layer received by passive microwave sensors. Based on this theoretical basis, a variety of snow depth algorithms have been developed, and multiple global snow depth / snow water equivalent products and Chinese regional snow depth products have been released. Although these datasets have achieved significant improvements in accuracy compared to earlier products, the spatial resolutions of 10 km and 25 km still restrict the application of snow depth at the regional scale, especially in areas with complex terrain and high heterogeneity of snow distribution. This resolution limitation not only affects the in-depth study of climate change but also restricts the accuracy of hydrological simulations. Therefore, developing snow depth products with higher spatial resolutions to meet the needs of regional-scale applications has become the focus and challenge of current research.

[0005] Currently, the method to improve the spatial resolution of snow depth products mainly realizes spatial refinement by exploiting the complementary characteristics of optical and passive microwave remote sensing data. Among them, the snow cover data extracted from optical remote sensing products are combined to improve the AMSR2 snow depth data. However, these linear regression-based methods have two limitations: one is that the model accuracy is vulnerable to the autocorrelation between variables. The other is that it is difficult to accurately characterize the complex non-linear relationship between snow depth and terrain environment, especially in mountainous areas with rugged terrain. To overcome the limitations of linear methods, machine learning methods have gradually been introduced into snow depth inversion research. The application of models such as artificial neural networks, random forests, and support vector machines has verified the advantages of machine learning in dealing with complex non-linear problems. With the development of technology, deep learning has shown great potential in remote sensing inversion, data fusion, downscaling and other research due to its powerful learning ability. However, for regions where snow distribution shows significant spatial heterogeneity and is mainly shallow snow, existing machine learning and deep learning algorithms often lead to systematic biases in the estimation of shallow snow when dealing with such characteristics. Therefore, there is an urgent need to develop a snow depth downscaling method with higher spatial resolution and the ability to accurately identify shallow snow areas to improve the snow depth monitoring ability at a large scale. Summary of the Invention

[0006] To solve the problems existing in the prior art, the present invention provides a snow depth downscaling method combining ResNet and snow attenuation curve. The method preprocesses the snow data of the target area to obtain the first snow depth image, the first auxiliary depth image, and the second auxiliary depth image; inputs the first snow depth image and the first auxiliary depth image into the model for training, and uses the trained model to obtain the first snow prediction image and the second snow prediction image; obtains the fine-scale residual value between the first snow depth image and the first snow prediction image to correct the second snow prediction image to obtain the downscaled snow depth data; uses the snow attenuation curve to perform secondary correction on the downscaled snow depth data to obtain the optimal snow depth data; and performs snow depth mapping according to the optimal snow depth data. The present invention can perform relatively accurate snow depth downscaling on the snow depth data of large-scale areas and achieve efficient snow depth prediction.

[0007] The present invention adopts the following technical solutions. A snow depth downscaling method combining ResNet and snow attenuation curve includes: Obtaining the snow data of the target area; the snow data includes the original snow depth data and the auxiliary data; Preprocessing the original snow depth data and the auxiliary data respectively to obtain the first snow depth image corresponding to the original snow depth data, and the first auxiliary depth image and the second auxiliary depth image corresponding to the auxiliary data; Inputting the first snow depth image and the first auxiliary depth image into the ResNet model for training to obtain the trained ResNet model; Inputting the first auxiliary depth image and the second auxiliary depth image into the trained ResNet model respectively to obtain the first snow prediction image and the second snow prediction image of the target area; Obtaining the fine-scale residual value between the first snow depth image and the first snow prediction image, and using the fine-scale residual value to correct the second snow prediction image to obtain the downscaled snow depth data; Using the snow attenuation curve to perform secondary correction on the downscaled snow depth data to obtain the optimal snow depth data; Performing snow depth mapping according to the optimal snow depth data at the set scale.

[0008] Further, the auxiliary data includes: longitude, latitude, digital elevation data, slope data, aspect data, snow cover degree data, snow cover date data, and vegetation cover data.

[0009] Further, the method for preprocessing the original snow depth data and the auxiliary data respectively is: Mask and clip the original snow depth data to obtain the first snow depth image with a spatial resolution of 25 km for the target area; Mask and clip the auxiliary data, and use the nearest neighbor interpolation method to sample the auxiliary data to spatial resolutions of 500 m and 25 km respectively; Extract the auxiliary data sampled to spatial resolutions of 500 m and 25 km to point data, and obtain the vector file corresponding to each grid point; Extract the longitude and latitude from the vector file to obtain the longitude data and latitude data with spatial resolutions of 500 m and 25 km; Perform standardization processing on the auxiliary data and the original snow depth data sampled to spatial resolutions of 500 m and 25 km; Stack the auxiliary data sampled to a spatial resolution of 500 m and the auxiliary data sampled to a spatial resolution of 25 km respectively to obtain the second auxiliary depth image corresponding to a spatial resolution of 500 m and the first auxiliary depth image corresponding to a spatial resolution of 25 km.

[0010] Further, the method for obtaining the fine-scale residual value of the first snow depth image and the first snow prediction image is: Obtain the difference between the first snow depth image and the first snow prediction image as the coarse-scale residual value; Calculate the fine-scale residual value according to the coarse-scale residual value using the bicubic interpolation method.

[0011] Further, the method for correcting the second snow prediction image using the fine-scale residual value is: Among them, is the fine-scale residual at a spatial resolution of 500 m; is the snow depth prediction data set at a spatial resolution of 500 m, that is, the second snow prediction image; is the downscaled snow depth data.

[0012] Further, using the fine-scale residual value to correct the second snow prediction image, specifically: Among them, is the downscaled snow depth data after secondary correction, SCF is the snow cover fraction data corresponding to each pixel, is the downscaled snow depth data.

[0013] The beneficial effects of the present invention are as follows: The present invention proposes a snow depth downscaling method combining the ResNet model and the snow attenuation curve. By using ResNet to fuse high-resolution optical remote sensing auxiliary data, it is possible to better learn the complex non-linear relationship between snow depth and complex surface, terrain and other features, realize efficient prediction of snow depth, and accurately identify shallow snow areas and snow-free areas by combining the snow attenuation curve and snow cover data, enabling relatively accurate snow depth downscaling of snow depth data in large-scale regions. Compared with traditional downscaling methods, the new method has higher accuracy, can provide a snow depth downscaling effect with a spatial resolution of 500 meters, and improves the ability of snow depth monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0015] Figure 1 Schematic flow chart of a snow depth downscaling method combining ResNet and snow attenuation curve according to an embodiment of the present invention; Figure 2 Schematic diagram of spatial detail comparison before and after downscaling in a small area one according to an embodiment of the present invention; Figure 3 Schematic diagram of spatial detail comparison before and after downscaling in a small area two according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0016] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0017] A schematic flow chart of a snow depth downscaling method combining ResNet and snow attenuation curve according to an embodiment of the present invention is as Figure 1 shown and includes: Obtain snow data of the target area; In the embodiments of the present invention, the snow cover data includes original snow depth data and auxiliary data. Among them, the snow depth data is the original coarse-resolution snow depth data at a spatial resolution of 25 kilometers, and the auxiliary data is at a spatial resolution of 500 meters and includes, but is not limited to: longitude, latitude, digital elevation model, slope data, aspect data, snow cover fraction data, snow cover days data, and vegetation cover data.

[0018] Preprocess the original snow depth data and the auxiliary data respectively to obtain the first snow depth image corresponding to the original snow depth data, and the first auxiliary depth image and the second auxiliary depth image corresponding to the auxiliary data; In the embodiments of the present invention, the preprocessing operation specifically includes: after projecting and cropping the coarse-resolution snow depth data, crop the snow depth area to be downscaled with a mask, crop the auxiliary data with the snow depth area mask to be downscaled, and resample all the auxiliary data to 500 meters and 25 kilometers spatial resolution using the nearest neighbor interpolation method; then extract the point data using the resampled snow depth data at 500 meters spatial resolution and the original coarse-resolution snow depth data at 25 kilometers spatial resolution, obtain the vector file for each grid point, and extract the longitude and latitude of the vector file to obtain the longitude data and latitude data at 500 meters and 25 kilometers; obtain all the basic data with spatial resolutions of 25 kilometers and 500 meters, and perform standardization processing on them to process all the data into data within the range of 0-1; for all the standardized data unified to 500 meters and 25 kilometers resolution, use Matlab software to process the basic data into a mat file, stack the original snow depth data at 25 kilometers spatial resolution and all the auxiliary data at 25 kilometers spatial resolution (longitude, latitude, digital elevation model, slope data, aspect data, snow cover fraction data, snow cover days data, and vegetation cover data) into a three-dimensional image matrix, that is, obtain the first snow depth image and the first auxiliary depth image, stack the auxiliary data at 500 meters spatial resolution (longitude, latitude, digital elevation model, slope data, aspect data, snow cover fraction data, snow cover days data, and vegetation cover data) into a three-dimensional image matrix with a resolution of 500 meters, that is, obtain the second auxiliary depth image, and output it as an HDF file for model training.

[0019] Input the first snow depth image and the first auxiliary depth image into the ResNet model for training to obtain the trained ResNet model; In the embodiments of the present invention, the Residual Neural Network (ResNet) is selected as the core algorithm. This is based on its excellent performance in processing complex spatial patterns. Compared with traditional random forests, backpropagation neural networks, and deep belief networks, ResNet has stronger feature extraction capabilities and a more stable training process. Moreover, this network can effectively capture and integrate multi-scale local and global spatial information, which is crucial for snow depth estimation in areas with complex terrain. Not only that, when the snow cover distribution within a region shows significant spatial heterogeneity and is mainly shallow snow cover, existing machine learning and deep learning algorithms often lead to systematic biases in shallow snow cover estimation. The present invention can effectively improve the accuracy of snow depth inversion under different snow cover conditions by using the snow attenuation curve and snow cover fraction data. By combining deep learning technology with snow cover fraction and snow attenuation curve, high-precision snow depth data reconstruction at a spatial resolution of 500 m in areas with strong snow heterogeneity is achieved. The specific training steps are as follows: The preprocessed first snow depth image and first auxiliary depth image at a spatial resolution of 25 km are input into the established ResNet model for model training. A pseudo snow depth image is obtained by calculating the feature map; and the loss function is calculated using the snow depth value estimated by the model and the original snow depth value, and the model is optimized through backpropagation.

[0020] The first auxiliary depth image and the second auxiliary depth image are respectively input into the trained ResNet model to obtain the first snow prediction image and the second snow prediction image of the target area.

[0021] In the embodiments of the present invention, the first snow prediction image at a spatial resolution of 25 km is simulated through the trained ResNet model( ); subsequently, the three-dimensional image (second auxiliary depth image) at a spatial resolution of 500 m after preprocessing is input into the ResNet model, and the second snow prediction image at a spatial resolution of 500 m is predicted( ) The fine-scale residual value between the first snow depth image and the first snow prediction image is obtained, and the second snow prediction image is corrected using the fine-scale residual value to obtain the downscaled snow depth data; In the embodiments of the present invention, first, the simulated first snow prediction image( ) is subtracted from the original snow depth image at a spatial resolution of 25 km( ) to calculate the simulated residual , and the calculation formula is as follows: where is the coarse-scale residual at a spatial resolution of 25 km; is the snow depth image at the original 25-kilometer spatial resolution; is the first snow cover prediction image at 25-kilometer spatial resolution.

[0022] Secondly, the coarse-scale residuals are calculated using the Rc in the original snow depth product and the grid values corresponding to the simulated 25-kilometer snow depth image ( ), and then the fine-scale residuals are obtained through bicubic interpolation , and the calculation formula is as follows: Among them, is the fine-scale residual at a resolution of 500 meters; Fbicubic is the bicubic interpolation function; is the original snow depth data; is the difference between the original snow depth image at 25-kilometer spatial resolution and the first snow cover prediction image at 25-kilometer spatial resolution.

[0023] Finally, the predicted 500-meter snow depth image is corrected using the fine-scale residuals to obtain the final downscaled snow depth product (SD d ), and the calculation formula is as follows: Among them, is the fine-scale residual value; is the second snow cover prediction image at 500-meter spatial resolution; is the downscaled snow depth data.

[0024] The downscaled snow depth data is corrected twice using the snow attenuation curve to obtain the optimal snow depth data; In the embodiment of the present invention, the downscaled snow depth data obtained based on the ResNet algorithm is further corrected twice using the snow attenuation curve, that is, the empirical relationship between the snow depth and the snow cover fraction during the snowmelt process is used to obtain the corrected optimal snow depth data ( ). Different from the downscaling method, this method dynamically adjusts the snow depth value according to the snow cover fraction, so as to more accurately reflect the distribution of snow cover during the melting period. The calculation formula is as follows: Among them, is the downscaled snow depth value corrected by SDC; is the snow cover fraction corresponding to each pixel, is the downscaled snow depth data.

[0025] Snow depth mapping is performed according to the optimal snow depth data at the set scale.

[0026] In another specific embodiment of the present invention: Obtain the long-term snow depth dataset of China at the National Tibetan Plateau Scientific Data Center, convert the ASCII file into raster data to obtain the snow depth data of China; the daily snow depth data of meteorological stations come from the official website of the National Meteorological Science Service Center; the digital elevation data, snow cover fraction, and snow cover days data are obtained from the official website of the National Tibetan Plateau Scientific Data Center, and the original spatial resolution is 500 meters; the continuous vegetation field data comes from the MODIS MOD44B product in the Google Earth Engine (GEE) platform.

[0027] Preprocess the obtained data: After projecting and cropping the long-term snow depth data of China using ArcMap software, mask and crop the area that needs to reflect the snow depth, and mask and crop the auxiliary data with the area reconstructed by the required snow depth. For the digital elevation model, snow cover fraction data, and snow cover days data, resample them to 25 km and 500 m spatial resolutions using the bilinear interpolation method. Use the Slope and Aspect tools in ArcMap to extract the corresponding slope and aspect data respectively, and resample them to 25 km and 500 m spatial resolutions using the bilinear interpolation method; for the snow depth data with 25 km and 500 m spatial resolutions, use the Raster to Point tool in ArcMap to extract the vector files of each grid point one by one, and perform feature extraction on the obtained point data to obtain longitude and latitude data; use the Point to Raster tool in ArcMap to assign the longitude data and latitude data to obtain the longitude image map and latitude image map with 25 km and 500 m spatial resolutions; use Google Earth Engine (GEE) to extract the continuous vegetation field data of the required area and process it into 500 m and 25 km resolutions using the GEE platform.

[0028] For all data unified to 500 m and 25 km resolutions, use Matlab software to process the basic data into mat files, stack the original snow depth data of 25 km and all 25 km auxiliary data (longitude, latitude, digital elevation model, slope data, aspect data, snow cover fraction data, snow cover days data, and vegetation cover data) into a three-dimensional image matrix and output it as an HDF file for model training; stack the longitude, latitude, digital elevation model, slope data, aspect data, snow cover fraction data, snow cover days data, and vegetation cover data with 500 m resolution into a three-dimensional image matrix with 500 m resolution for model prediction.

[0029] An embodiment of the present invention realizes the construction of a snow depth downscaling model based on the ResNet algorithm on the pycharm software platform. The three-dimensional image matrix used for model training is input into the model for training. During the training process, the convolutional layer performs pixel-by-pixel scanning on the original data and the auxiliary data set, and uses the ResNet network to deeply mine the image information to obtain the feature map. Finally, through the ResNet model, the loss function is calculated using the estimated snow depth value and the original snow depth value, and then this result is backpropagated and the model is optimized. Finally, simulated 25-kilometer snow depth data is generated. Specifically, the network structure includes several key stages: First, the network extracts features through two consecutive convolutional layers: the initial layer extracts shallow features from the auxiliary data , and then the second layer is used to further extract features from and use it as the input of the residual dense block ( ). In , the output of each module is calculated as follows: where is a composite function representing convolution and activation operations. The activation function selected here is the rectified linear unit (ReLU) function, The calculation of the latter layer involves the use of the previous layer. The ReLU function is used to enhance the non-linear characteristics of the image.

[0030] Secondly, the dense feature fusion process integrates the outputs of all to extract information features in the global range. Subsequently, these global features are connected with the shallow features to generate the final dense features ( ). Then, the final output result is obtained through another convolution. Finally, the consistency of the model structure is constrained by calculating the loss function using the original snow depth value. The calculation of the loss function is as follows: where, is the sample logarithm, is the network parameter, is the training model, is the input data of the i-th sample, is the original snow depth data of the i-th sample. In the embodiment of the present invention, all convolutional layers are configured with a kernel size of 3×3, and 64 filters are used in the shallow feature extraction layer and the local and global feature fusion layers respectively.

[0031] Set some numerical ranges for the parameters of the model, including the number of epochs (epoch), learning rate (learning_rate), number of layers in the residual dense block (nDense), growth coefficient (growthRate), batch size (batch_size), number of features (nFeat), etc. Through repeated adjustment and comparative experiments, the model with the optimal accuracy is obtained. The specific values of the parameters of the optimal model in this example are: epoch = 300, learning_rate = 0.001, nDense = 5, growthRate = 32, batch_size = 2, nFeat = 64.

[0032] Subtract the simulated snow depth image at a spatial resolution of 25 km from the original 25 km snow depth image to calculate the simulation residual. Secondly, calculate the coarse-scale residual using the grid values corresponding to the simulated 25 km snow depth image in the original snow depth product, and then obtain the fine-scale residual through bicubic interpolation. Finally, use the fine-scale residual to correct the prediction to obtain the 500 m snow depth image, and obtain the optimal downscaled snow depth data. Denormalize the optimal downscaled snow depth data to restore it to the original scale, and obtain the final downscaled snow depth data based on the ResNet model.

[0033] The downscaled snow depth data obtained based on the ResNet algorithm is corrected twice using the snow attenuation curve, that is, the empirical relationship between snow depth and snow cover during the snowmelt process, to obtain the corrected optimal snow depth data. Different from the downscaling method, this method dynamically adjusts the snow depth value according to the snow cover, so as to more accurately reflect the distribution of snow during the melting cycle; map the corrected optimal snow depth data at a spatial resolution of 500 m.

[0034] Verify the original data, the 500 m snow depth downscaled data based on the ResNet model, and the 500 m downscaled snow depth data corrected based on the snow attenuation curve. The indicators include root mean square error (RMSE), bias (BIAS), and mean absolute error (MAE). The verification results of the accuracy are shown in Table 1.

[0035] Table 1 Schematic table of accuracy verification of different snow depth data In the embodiment of the present invention, two small regions are further selected in the Qinghai-Tibet Plateau as examples to compare the long-term snow depth product of China with a original spatial resolution of 25 km and the digital elevation model, such as Figure 2 and Figure 3As shown, the downscaled snow depth model of the present invention is relatively consistent with the spatial distribution of the original snow depth data. The resolution of the original snow depth product is low, showing the same snow depth value over a large area. However, the terrain of the Qinghai-Tibet Plateau is complex, and the actual snow cover distribution should exhibit obvious heterogeneity, mostly in a patchy distribution. From this, it can be seen that the accuracy of the model proposed by the present invention after downscaling is more delicate and has good consistency with the terrain features, and can better identify snow-free areas and shallow snow areas.

[0036] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Claims

1. A snow depth downscaling method combining ResNet and snow attenuation curves, characterized in that, Including: Obtaining snow cover data of a target area; the snow cover data includes original snow depth data and auxiliary data; Preprocessing the original snow depth data and the auxiliary data respectively to obtain a first snow depth image corresponding to the original snow depth data, and a first auxiliary depth image and a second auxiliary depth image corresponding to the auxiliary data; Inputting the first snow depth image and the first auxiliary depth image into a ResNet model for training to obtain a trained ResNet model; Inputting the first auxiliary depth image and the second auxiliary depth image into the trained ResNet model respectively to obtain a first snow cover prediction image and a second snow cover prediction image of the target area; Obtaining the fine-scale residual value between the first snow depth image and the first snow cover prediction image, and using the fine-scale residual value to correct the second snow cover prediction image to obtain downscaled snow depth data; Performing secondary correction on the downscaled snow depth data using a snow attenuation curve to obtain optimal snow depth data; Performing snow depth mapping according to the optimal snow depth data at a set scale.

2. A snow depth downscaling method combining ResNet and snow attenuation curves according to claim 1, characterized in that: The auxiliary data includes: longitude, latitude, digital elevation data, slope data, aspect data, snow cover fraction data, snow cover date data, and vegetation cover data.

3. A snow depth downscaling method combining ResNet and snow attenuation curves according to claim 1, characterized in that: The method for preprocessing the original snow depth data and the auxiliary data respectively is: Performing mask cropping on the original snow depth data to obtain a first snow depth image with a 25-kilometer spatial resolution of the target area; Performing mask cropping on the auxiliary data, and respectively sampling the auxiliary data to 500-meter and 25-kilometer spatial resolutions using the nearest neighbor interpolation method; Extracting the auxiliary data sampled to 500-meter and 25-kilometer spatial resolutions into point data, and obtaining a vector file corresponding to each grid point; Extracting longitude and latitude from the vector file to obtain longitude data and latitude data with 500-meter and 25-kilometer spatial resolutions; Performing standardization processing on the auxiliary data and the original snow depth data sampled to 500-meter and 25-kilometer spatial resolutions; Stacking the auxiliary data sampled to 500-meter spatial resolution and the auxiliary data sampled to 25-kilometer spatial resolution respectively to obtain a second auxiliary depth image corresponding to 500-meter spatial resolution and a first auxiliary depth image corresponding to 25-kilometer spatial resolution.

4. A snow depth downscaling method combining ResNet and snow attenuation curve according to claim 1, characterized in that: The method for obtaining the fine-scale residual value between the first snow depth image and the first snow cover prediction image is: Obtaining the difference between the first snow depth image and the first snow cover prediction image as a coarse-scale residual value; Calculating the fine-scale residual value according to the coarse-scale residual value using the bicubic interpolation method.

5. A snow depth downscaling method combining ResNet and snow attenuation curve according to claim 1, characterized in that: The method for correcting the second snow cover prediction image using the fine-scale residual value is: Among them, is the fine-scale residual with a spatial resolution of 500 meters; is the snow depth prediction dataset with a spatial resolution of 500 meters, that is, the second snow prediction image; is the downscaled snow depth data.

6. A snow depth downscaling method combining ResNet and snow attenuation curve according to claim 1, characterized in that: Correcting the second snow cover prediction image using the fine-scale residual value, specifically: Among them, is the downscaled snow depth data after secondary correction, and SCF is the snow cover fraction data corresponding to each pixel. is the downscaled snow depth data.

Images