CYGNSS based Mapping of Inundation Dynamics in Congo Peatlands Using Machine Learning

Abstract

Peatlands store approximately one-third of the world’s soil carbon, making accurate mapping of surface water dynamics essential for understanding their role in the global carbon cycle. Global Navigation Satellite System - Reflectometry (GNSS-R) provides an effective means for long-term, large-scale monitoring of surface water, particularly in densely vegetated tropical peatlands due to the strong penetration capability of L-band signals.In this study, we map inundation over the Congo Basin using data from the CYGNSS mission. The training, validation and testing data consist of 559 temporally sparse inundation fraction samples (March 2017-August 2021) derived from water table depth observations at four in-situ stations combined with information on spatial variability of ground elevation. Multiple features are extracted from CYGNSS delay–doppler maps and their retrieved reflectivity, including signal-to-noise ratio (SNR), statistical moments (mean, variance), and waveform-based indicators such as leading-edge slope (LES) and trailing-edge slope (TES). These GNSS-R features are combined with auxiliary variables including NDVI and precipitation and are used to train machine learning models (Random Forest) for estimating inundation fraction.Model performance is evaluated using a leave-one-spatial-cluster-out cross-validation strategy to ensure spatial independence between training and testing data. The results demonstrate that models based on multiple CYGNSS features significantly outperform those using single features alone. At high inundation level, models based solely on CYGNSS tend to underestimate surface water coverage, whereas the inclusion of precipitation significantly reduces this bias and improves R² during highly saturated conditions. These findings highlight the strong potential of GNSS-R combined with machine learning for large-scale tropical peatland hydrological monitoring.