A generalized machine learning approach for dissolved oxygen estimation at multiple spatiotemporal scales using remote sensing.

Dissolved oxygen (DO) is an effective indicator for water pollution. However, since DO is a non-optically active parameter and has little impact on the spectrum captured by satellite sensors, research on estimating DO by remote sensing at multiple spatiotemporal scales is limited. In this study, the support vector regression (SVR) models were developed and validated using the remote sensing reflectance derived from both Landsat and Moderate Resolution Imaging Spectroradiometer (MODIS) data and synchronous DO measurements (N = 188) and water temperature of Lake Huron and three other inland waterbodies (N = 282) covering latitude between 22-45 °N. Using the developed models, spatial distributions of the annual and monthly DO variability since 1984 and the annual monthly DO variability since 2000 in Lake Huron were reconstructed for the first time. The impacts of five climate factors on long-term DO trends were analyzed. Results showed that the developed SVR-based models had good robustness and generalization (average R = 0.91, root mean square percentage error = 2.65%, mean absolute percentage error = 4.21%), and performed better than random forest and multiple linear regression. The monthly DO estimates by Landsat and MODIS data were highly consistent (average R = 0.88). From 1984 to 2019, the oxygen loss in Lake Huron was 6.56%. Air temperature, incident shortwave radiation flux density, and precipitation were the main climate factors affecting annual DO of Lake Huron. This study demonstrated that using SVR-based models, Landsat and MODIS data could be used for long-term DO retrieval at multiple spatial and temporal scales. As data-driven models, combining spectrum and water temperature as well as extending the training set to cover more DO conditions could effectively improve model robustness and generalization.