An AI-Driven Multi-Layer Perceptron Model for Early Detection of Lake Eutrophication

Abstract

Early detection of eutrophication is crucial for water body protection because it requires substantial time and resources to restore the water quality once polluted. Conventional methodology of determining water quality mandated repetitious and protracted manual labour. Many on-field sensors and monitoring systems have been proposed to ease the burden. However, though in-situ systems could reduce manual chemical analyses, AI models classifying the quality of lakes have drawn relatively little attention as a solution to simplifying the process. Therefore, this study suggests a multi-layer perceptron (MLP) model for early detecting eutrophic lakes. MLP models are easily adaptable to different environments and features, assisting the researchers by greatly shortening the detection time. Water assessment data received from WaterAtlas were sparse and therefore were preprocessed using linear regression. The data was randomly split to tune the number of epochs and the depth of the model. Afterwards, the model was tested by lake-based datasets to observe the model performance for unknown lakes only by existing data. The model had classified the dataset with 0.8911 accuracy and precision, recall, and f1-score of eutrophic lakes all reaching above 0.90, proving that the model could accurately evaluate the water quality. The macro-ROC and micro-ROC curves had the AUC scores of 0.96 and 0.95, respectively. This result with a narrow gap indicated that the model can precisely estimate the quality of lakes though the input data type is not evenly distributed. Therefore, this study shows that MLP is a suitable approach for early detection of lake eutrophication.