Implementation of Remote Sensing and Deep Learning Techniques For Lake Water Quality Classification

Recent advances in Deep Learning (DL) have enabled the development of autonomous systems capable of extracting complex patterns from large datasets. When combined with Remote Sensing (RS), these methods provide powerful tools for environmental monitoring by offering extensive observations of ecosystems. This study focuses on classifying the Water Quality of an Azorean Lake. The temporal measure is based on over 4 years of in-situ measurements, combined with Sentinel-2 satellite imagery, which was preprocessed with cloud masking, temporal averaging, and spatial alignment with sampling dates. The applied DL models were DenseNet, ResNet, and UNet and tuned using grid-search, with both original and pre-trained variants evaluated at 2 spatial scales, in-situ site location and full-lake imagery. The results show that pre-trained DenseNet and UNet models achieve the highest F1-score, 0.988, and that full-lake images generally improve classification performance. The study demonstrates the strong potential of satellite-based approaches for water quality assessment and highlights the benefits of model pre-training and careful spatial configuration.