Two-stage downscaling and correction cascade learning framework for generating long-time series seamless soil moisture

Abstract

Soil moisture (SM) is a key state variable in agricultural, hydrological, and ecological studies. Microwave remote sensing can retrieve soil moisture at regional or global scales, but is limited by coarse spatial resolution. In order to generate large-scale, spatiotemporally seamless soil moisture of high precision, we propose a two-stage downscaling and correction cascade learning framework by fusing multi-sourced remote sensing and in-situ data. Under the framework, the Hybrid Attention based residual dense Network for soil moisture Downscaling (HAND) is developed to downscale the Soil Moisture Active Passive (SMAP) SM products from 36 km to 1 km effectively. The Random Forest method is subsequently employed to correct the downscaled SM products by in-situ measurements and the 1 km seamless daily SM products of high precision are then generated. The soil moisture downscaling network adopts the Residual Dense connection Network (RDN) as the backbone and embeds a multi-factor interactive attention module, a cross-attention module, and the hybrid attention block with increased/ decreased receptive field, to comprehensively extract the complex relationships between geoscience parameters and soil moisture. The western continental of the United States is served as the study area of this paper, covering 2016–2020. The Pearson correlation (R, unitless) and the Unbiased Root-Mean-Square Error (UbRMSE, $m^3/m^3$) values of the HAND downscaled products with SMAP are 0.65 and 0.066 $m^3/m^3$, showing the ability of HAND model to achieve satisfactory accuracy while maintaining consistency with original SMAP products, as well as restoring fine spatial details. After the in-situ correction, the R and UbRMSE values of ten-folder cross validation against the in-situ SM reach 0.92 and 0.033$m^3/m^3$, while the metrics of SMAP SM against the in-situ SM are 0.46 and 0.083 $m^3/m^3$, which demonstrates great potential of the proposed method in water resources management at regional scale.