IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (Jan 2022)

An Effective Land Type Labeling Approach for Independently Exploiting High-Resolution Soil Moisture Products Based on CYGNSS Data

  • Yan Jia,
  • Shuanggen Jin,
  • Qingyun Yan,
  • Patrizia Savi,
  • Rongchun Zhang,
  • Wenmei Li

Journal volume & issue
Vol. 15
pp. 4234 – 4247


Read online

Recently, soil moisture (SM) has been estimated using Cyclone Global Navigation Satellite System (CYGNSS) data. Machine learning (ML) algorithms for CYGNSS SM estimation can minimize unpredictable influences and help improve the accuracy of SM retrieval. However, ML-based CYGNSS SM estimation requires ancillary data from other sources, and thus, the uncertainty, internal errors, and even dependence on external parameters of this process may complicate and limit SM estimation. In this article, a simple land type (LT) digitization strategy that incorporates the idea of classification is proposed with feature optimization to achieve an effective and independent SM retrieval without any other auxiliary data. The input features are chosen from the CYGNSS data themselves, and the corresponding labels (digitized stable LTs) are used in the training stage of the SM estimation model. During the fine-tuning stage, several input features (such as the dielectric constant and incident angle) are compared and selected after optimization to achieve better results. Moreover, the CYGNSS data are gridded at 9 × 9 km to validate the enhanced soil moisture active passive mission SM products at a resolution of 9 km. Only three input variables are adopted for the SM learning model, which are directly derived from the CYGNSS data for independently estimating SM at a high spatial resolution. Powerful performance is achieved by extreme gradient boosting based on a LT digitalization strategy, with root-mean-square error (RMSE) and unbiased RMSE (ubRMSE) values of 0.063 cm3/cm3 and a correlation coefficient (R) of 0.71 for the entire dataset. The performances of different ML learning models for various LTs are presented. The mean ubRMSE and RMSE are 0.041 cm3/cm3 and 0.057 cm3/cm3, respectively. The results demonstrate the effectiveness of the proposed LT digitization strategy for retrieving SM from CYGNSS data with various ML methods and the capability of SM estimation using the CYGNSS product as a new independent source.