Inland Water Trophic State Identification using Remote Sensing data and Machine Learning Approach

Abstract

Providing information on inland waters’ trophic state levels (oligotrophic, mesotrophic, eutrophic, and hypereutrophic) is considered more effective in communicating the lake’s conditions to the public and policymakers than water quality. It is highly demanding for a straightforward method to estimate inland waters’ trophic state. This study proposes a model for estimating the lake trophic state levels from remote sensing data. We used simulation data to overcome data limitations in building the classification model. The simulation data consists of one nanometer (nm) interval spectra, paired with calculated Secchi Disk Depth (SD in m) representing various water types. We convert the spectra into Landsat band configuration and classify SD into four trophic state levels. We compare four machine learning classification models, i.e., Classification and Regression Trees (CART), k-Nearest Neighbour (k-NN), Support Vector Machine (SVM), Random Forest (RF), and select the best model using the Kappa Index. We apply the selected model to satellite images and estimate trophic state levels of various lakes across time and space. Our result demonstrated that the developed model could robustly identify the lake trophic state levels. This rapid identification procedure could provide valuable spatial and temporal information of the lake’s conditions for the public and policymakers to support inland water sustainable management.