Enhanced Monitoring of Suspended Solids via Sentinel‐2A and Machine Learning: Impact of Temporal Resolution

ABSTRACT

The quality of surface water observation plays a pivotal role in the proper management of the environment, particularly where pollution and sedimentation are likely to occur. In this paper, the performance of various machine learning models to predict suspended solids (SS) concentration with the help of multi‐temporal composites based on Sentinel‐2A Level‐2A satellite images is assessed. The process includes the initial processing of satellite observations by cloud masking, and in‐situ measurements and temporal aggregation of satellite observations to 2‐h, 6‐h, and daily composites. The input features were spectral bands (B2‐B12) and the Scene Classification Layer (SCL). Various machine learning models were developed and trained in 70/30 and 80/20 data divisions, and their performance was measured using RMSE, MAE, MSE, R2, and RMSLE. The results indicate that ensemble models, especially XGBoost and random forest models, are the lowest error achievers, especially when using 2‐h and 6‐h composites. The scale of 6‐h aggregation gives the most optimal balance between the reduction of noise and the retention of information. On the other hand, deep learning models (RNNs) and linear regression do not generalize well, which is probably because of the small size of the dataset, noise in the high‐frequency observations, and insufficient ability to find complicated nonlinear interactions. Moreover, well‐structured temporal sequences are also required by RNNs, which is limited by the intermittent and aggregated characteristics of satellite data. In general, the paper reveals that short‐time temporal composites and strong ensemble techniques can enhance the estimation of satellite‐based SS.