Soil Salinity Estimation Over Coastal Wetlands Based on Random Forest Algorithm and Hydrological Connectivity Metric

Abstract

Owing to climate warming and human activities (irrigation and reservoirs), sea level rise and runoff reduction have been threatening the coastal ecosystem by increasing the soil salinity. However, short-term sparse in situ observations limit the study on the response of coastal soil salinity to external stressors and thus its effect on coastal ecosystem. In this study, based on hydrological connectivity metric and random forest algorithm (RF), we develop a coastal soil salinity inversion model with in situ observations and satellite-based datasets. Using Landsat images and ancillary as input variables, we produce a 30-m monthly grid dataset of surface soil salinity over the Yellow River Delta. Based on the cross-validation result with in situ observations, the proposed RF model performs higher accuracy and stability with determination coefficient of 0.89, root mean square error of 1.48 g·kg-1, and mean absolute error of 1.05 g·kg-1. The proposed RF model can gain the accuracy improvements of about 11–43% over previous models at different conditions. The spatial distribution and seasonal variabilities of soil salinity is sensitive to the changing signals of runoff, tide, and local precipitation. Combining spatiotemporal collaborative information with the hydrological connectivity metric, we found that the proposed RF model can accurately estimate surface soil salinity, especially in natural reserved regions. The modeling results of surface soil salinity can be significant for exploring the effect of seawater intrusion and runoff reduction to the evolution of coastal salt marsh ecosystems.