Abstract:Accurate snow depth estimation is of significant importance, particularly for preventing avalanche disasters and predicting flood seasons. The predominant approaches for such snow depth estimation, based on deep learning methods, typically rely on passive microwave remote sensing data. However, due to the low resolution of passive microwave remote sensing data, it often results in low-accuracy outcomes, posing considerable limitations in application. To further improve the accuracy of snow depth estimation, in this paper, we used active microwave remote sensing data. We fused multi-spectral optical satellite images, synthetic aperture radar (SAR) images and land cover distribution images to generate a snow remote sensing dataset (SRSD). It is a first-of-its-kind dataset that includes active microwave remote sensing images in high-latitude regions of Asia. Using these novel data, we proposed a multi-scale feature perception and aggregation neural network (MFPANet) that focuses on improving feature extraction from multi-source images. Our systematic analysis reveals that the proposed approach is not only robust but also achieves high accuracy in snow depth estimation compared to existing state-of-the-art methods, with RMSE of 0.360 and with MAE of 0.128. Finally, we selected several representative areas in our study region and applied our method to map snow depth distribution, demonstrating its broad application prospects.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is to improve the accuracy of high - resolution snow depth estimation. Specifically, the existing snow depth estimation methods based on deep learning mainly rely on passive microwave remote sensing data. However, due to the low spatial resolution of passive microwave remote sensing data, the accuracy of the estimation results is poor, and there are great limitations in application. To solve this problem, this paper uses active microwave remote sensing data and fuses multispectral optical satellite images, synthetic aperture radar (SAR) images and land cover distribution images to construct a snow depth remote sensing data set (SRSD). On this basis, a multi - scale feature perception and aggregation neural network (MFPANet) is proposed, aiming to improve the ability to extract features from multi - source images, so as to achieve higher - precision snow depth estimation. ### Main contributions of the paper 1. **Constructing a multi - source data set**: This paper constructs a data set for snow cover remote sensing in high - latitude areas of Asia. This data set fuses multispectral optical satellite images, SAR images and land cover distribution images, and uses the surface snow depth measurements at meteorological stations as the ground truth. 2. **Proposing a multi - scale neural network**: Different from the "point - to - point" prediction that ignores spatial characteristics, the model proposed in this paper is a "region - to - point" deep model for snow depth estimation. The network includes a multi - branch feature extraction unit (MBFE), a multi - scale feature atrous aggregation module (MSFAA) and a high - and low - level feature fusion module (HLF). These components endow the new model with multi - scale feature perception ability, which is especially helpful to reduce the spatial interference in non - snow areas, so as to achieve high - precision snow depth estimation. 3. **Drawing snow depth distribution maps**: By optimizing the model parameters, a snow depth distribution map with a resolution of 320 meters is generated in the study area. Based on this method, it can be predicted that high - resolution snow depth maps will be generated in any area of interest. ### Method overview 1. **Multi - branch feature extraction unit (MBFE)**: A dual - branch residual network structure is used to extract and fuse features at different levels. The down - sampling order of each branch is different, and the feature complementarity is enhanced through the concatenation operation. 2. **Multi - scale feature atrous aggregation module (MSFAA)**: It contains multiple depth - wise separable convolutions with different dilation rates, as well as 1x1 convolution and 2x2 average pooling operations. Features are aggregated through the pyramid pooling concept to reduce the interference of non - snow factors. 3. **High - and low - level feature fusion module (HLF)**: Two branches are designed to input high - level semantic information and low - level feature information respectively. High - and low - level features are fused through weighting operations and up - sampling to avoid information distortion and loss of diversity caused by simple combination. ### Experimental setup The experiment uses the PyTorch framework and is configured with an Nvidia GeForce RTX4070Ti GPU. The optimizer adopts the adaptive moment estimation (Adam), and the learning rate strategy is the "ploy" strategy. The specific formula is as follows: \[ \text{lr} = \text{base\_lr} \times \left(1 - \frac{\text{epoch}}{\text{num\_epoch}}\right)^{\text{power}} \] where \(\text{lr}\) is the updated learning rate, \(\text{base\_lr}\) is the baseline learning rate, \(\text{epoch}\) is the number of iterations, \(\text{num\_epoch}\) is the maximum number of iterations, and \(\text{power}\) controls the shape of the curve (usually greater than 1). In the model of this paper, \(\text{power}\) is set to 0.9, and the final fixed number of iterations is 250. No pre - trained parameters are used in the training process, and the batch size is set to 64. To prevent the model from over - fitting, multiple methods such as data augmentation, dropout and normalization are adopted. In addition, a ten - fold cross - validation experiment is also designed, and a part of the data is selected as the test set in turn.

MFPANet: Multi-Scale Feature Perception and Aggregation Network for High-Resolution Snow Depth Estimation

Downscaling Snow Depth Mapping by Fusion of Microwave and Optical Remote-Sensing Data Based on Deep Learning

Snow depth mapping in agricultural areas of Northeast China based on deep learning and multi-temporal Sentinel-1 data

Snow depth retrieval from microwave remote sensing by combining wavelet transform and machine learning models in Northern Xinjiang, China

Comparison of Machine Learning-Based Snow Depth Estimates and Development of a New Operational Retrieval Algorithm over China

Deep Dense Multi-Scale Network for Snow Removal Using Semantic and Depth Priors

An Approach to Improve the Spatial Resolution and Accuracy of AMSR2 Passive Microwave Snow Depth Product Using Machine Learning in Northeast China

Retrieval of Snow Depth over Arctic Sea Ice Using a Deep Neural Network

Cross-dimensional feature attention aggregation network for cloud and snow recognition of high satellite images

Deep Dense Multi-scale Network for Snow Removal Using Semantic and Geometric Priors

Snow Cover Mapping for Mountainous Areas by Fusion of MODIS L1B and Geographic Data Based on Stacked Denoising Auto-Encoders

A combined multi-source data and deep learning approach for retrieving snow depth on Antarctic Sea ice during the melting season

A machine learning approach for estimating snow depth across the European Alps from Sentinel-1 imagery

Multi-Scale and Multi-Level Feature Fusion Network for Quantitative Precipitation Estimation with Passive Microwave

Estimation of Snow Depth from AMSR2 and MODIS Data based on Deep Residual Learning Network

PCSSR‐DNNWA: A Physical Constraints Based Surface Snowfall Rate Retrieval Algorithm Using Deep Neural Networks With Attention Module

Controllably Deep Supervision and Multi-Scale Feature Fusion Network for Cloud and Snow Detection Based on Medium- and High-Resolution Imagery Dataset

GNSS-IR Snow Depth Retrieval Based on the Fusion of Multi-Satellite SNR Data by the BP Neural Network

Multiscale and Multilevel Feature Fusion Network for Quantitative Precipitation Estimation With Passive Microwave

Extraction of snow cover from high-resolution remote sensing imagery using deep learning on a small dataset

Snowpack Estimation in Key Mountainous Water Basins from Openly-Available, Multimodal Data Sources