A comprehensive evaluation of inland water quality monitoring using the long-term Landsat archive and mixture density networks

Lorena Pinheiro Silva , Xiaoxu Guo, Rebecca Lee North, Greg Silsbe

Nutrient enrichment of inland and coastal waters is visible from space. In these regions, chlorophyll a (chl-a), total suspended solids (TSS), and Secchi depth (ZSD) constitute the most commonly targeted water quality parameters. A growing number of studies have shown that, despite not being strictly optically active, total nitrogen (TN) and total phosphorus (TP) can nonetheless be estimated from remotely sensed data with varying accuracy, underscoring that in some environments water color is a robust indicator of trophic state. This study provides a comprehensive analysis of Mixture Density Networks (MDN) model performance across a suite of water quality parameters (TN, TP, chl-a, ZSD, and TSS) using Landsat 7–9 co-located with nearly 100 sampled reservoirs in Missouri, USA. MDN models were deployed across several pipelines that represent different combinations of atmospheric correction processors (LEDAPS/LaSRC, ACOLITE, and RAdCor), matchup windows (within 1 and 3 -day), feature engineering strategies (single versus multi-feature input), and model output configurations (single versus multiple-output). The median symmetric accuracy (MdSA) of model predictions was variable across parameters and pipelines. TN had the best-performing pipeline (MdSA=16.1%) and pipeline ensemble (MdSA=25.9%) followed by ZSD, TSS, and TP. Chl-a had worst ensemble accuracy (MdSA=71.1%) and was consistently the least accurate water quality metric across pipelines. That in this dataset definitive metrics of eutrophication that include non-optically active constituents (TN and TP) can be more accurately estimated than the most widely targeted water quality parameter (chl-a) is discussed from theoretical and applied perspectives.