A reduced latency regional gap-filling method for SMAP using random forest regression

Authors: WANG, X. - Hohai University; LU, H. - Hohai University; Crow, Wade; ZHU, Y. - Hohai University; SU, J. - Chinese Academy Of Sciences, Nanjing Branch; ZHENG, J. - Hohai University; GOU, Q. - Hohai University

Submitted to: iScience
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 12/19/2022
Publication Date: 12/22/2022
Citation: Wang, X., Lu, H., Crow, W.T., Zhu, Y., Su, J., Zheng, J., Gou, Q. 2022. A reduced latency regional gap-filling method for SMAP using random forest regression. iScience. 20. Article 105853. https://doi.org/10.1016/j.isci.2022.105853.
DOI: https://doi.org/10.1016/j.isci.2022.105853

Interpretive Summary: Soil moisture estimates obtained from satellite-based sensors are potentially of value for a broad range of agricultural applications. However, these data sets typically contain temporal gaps and do not provide unique estimates every single day. Such gaps can cause problems for agricultural applications where day-to-day variations in soil moisture are often critical. Based on a novel machine-learning approach, this paper proposes a gap-filling procedure capable of generating a continuous daily soil moisture product. Validation results using extensive ground-based soil moisture measurements indicate that the gap-filling approach is robust and accurate. The results of this study will be used to continuously track the availability of soil moisture at fine temporal scales and better mitigate the impact of soil moisture extremes on local agricultural productivity.

Technical Abstract: Microwave remote sensing is a direct way to acquire quasi-global soil moisture state estimates in near-real-time. The L-band NASA Soil Moisture Active/Passive (SMAP) mission represents a significant advance in these efforts. However, significant temporal gaps in SMAP soil moisture retrieval products limit their value for certain operational applications. This study aims to estimate missing SMAP L3 composite (SMAP_L3) soil moisture retrievals using a low-latency Random Forest regression approach. To build the Random Forest gap-filling model, we make use of on-orbit satellite data and SMAP official ancillary data as inputs and utilize SMAP_L3 soil moisture retrievals as a training target. Resulting gap-filled estimates (RF-SMAP) are evaluated using both in-situ and triple collocation (TC) approaches to determine their accuracy relative to existing on-orbit passive microwaved products (i.e., AMSR2, SMOS_L3, and SMAP_L3) and the (time-continuous) SMAP Level 4 Surface and Root-zone Soil Moisture data assimilation product. Although RF-SMAP cannot fully match the skill of the SMAP_L3 retrievals on which it is based, our proposed model predicts soil moisture status during periods where SMAP fails to retrieve data (due to, e.g., dense vegetation or intense rainfall). Overall, our predicted data shows good consistency with the in-situ observations collected within the Huai River Basin during the wet season (median R = 0.39, median ubRMSD = 0.05 m3/m3, median Rbias = 7 %) and closely matches the cumulative distribution function of the in-situ data during autumn. Furthermore, after integrating the original SMAP_L3 data and our (gap-filling) RF-SMAP estimates, the resulting seamless RF-SMAP product exhibits good temporal correlation with in-situ data (i.e., median R values reach 0.40, 0.47, and 0.34 for the entire period, wet season, and dry season, respectively). Given the moderate data latency of the SMAP_L4 soil moisture product (up to 7 days), RF-SMAP can be used as a low-latency solution to enhance the near-real time data availability of continuous SMAP_L3 soil moisture retrievals.