A Self-learning Weight Calibration Based Residual Dense Network for Soil Moisture Downscaling

Soil moisture is a key state variable in the ecosystem. However, existing soil moisture products often cannot have both high spatial resolution and long time series. Therefore, it is important to downscale the soil moisture products for fine application. So far，the soil moisture downscaling methods can be summarized as satellite-based methods, model-based methods, and learning-based methods. Satellite-based methods can integrate the advantageous features of different satellite, but face challenging in practical applications due to spatiotemporal differences and cloud coverage. Model-based methods offer superior interpretability of physical processes, but the model parameters are difficult to obtain and cannot fully characterize the non-linear mapping relationship in real physical processes. Learning-based methods have the prominent ability to fit nonlinear relationship, while existing learning-based methods do not take the correlation and redundancy between various covariates into consideration, and the extraction of key features is insufficient. Hence, we proposed Self-learning Weight calibration based soil moisture Downscaling Network (SWDN) to couple the learning-based model with the weight calibration strategy for improving the products accuracy, as shown in Fig 1. The proposed SWDN constructs a complex mapping relationship model from multi-factor geoscience parameters to soil moisture. Under this framework, the residual dense connection network is adopted as the backbone for feature extraction and guides the reconstruction of soil moisture spatial information. The spatial weight and multi-factor weight self-learning modules are designed to adaptively calibrate feature weights of spatial direction and multi-factors, respectively. By updating parameters of above two modules, the weights of key features are enhanced and the weights of redundant features are weakened to achieve efficient extraction of discriminative features. Subsequently, under the assumption of scale invariance, model from geoscience parameters to soil moisture is fully trained on low spatial resolution data and applied on high spatial resolution geoscience parameters to generate high-precision soil moisture products. Experiments on the Western Continental United States dataset show that the proposed SWDN method exhibits superior performance with higher consistency with in-situ measurements and richer spatial texture information over the comparison methods. Compared to the state of the art downscaling methods, results demonstrate that the R and RMSE reach 0.44 and 0.077 , which improve 16% and 5% respectively. The maps of soil moisture distribution before and after downscaling are shown in Fig 2. Fig.1. The structure of the SWDN method for soil moisture downscaling. Fig.2. Mapping of SM distribution before and after downscaling on 2017.7.17. (a) Original SMAP; (b) BPNN downscaled SM; (c) DBN downscaled SM; (d) RDN downscaled SM;(e) SWDN downscaled SM.