Estimating chlorophyll-a in tropical inland waters is difficult because of complex optical conditions, limited field data, and frequent failures in atmospheric correction (over 70%). Traditional algorithms (R2 below 0.45) do not perform well in Case-2 waters such as Laguna Lake in the Philippines. In this study, we introduce a two-stage transfer learning approach using 3D Convolutional Neural Networks (3D-CNNs). We use simulated pre-training with 126,000 samples, optimize spatial context with patch sizes from 5×5 to 11×11, and apply geometric augmentation to increase the dataset size by six times. Our process includes quality filtering with six Water Quality and Science Flags, per-band Z-score normalization, and stratified sampling to evaluate Sentinel-3 OLCI 16-band images at 300 m resolution. The best results came from the 9×9 patch model, which reached R2 = 0.5315, RMSE = 0.6870, and MAE = 0.3221 log10 μg/L on 21,135 test samples. This improved baseline performance by 17.9% and outperformed traditional methods by 18 to 28%. Transfer learning was key, giving a 40% R2 increase over direct training, and the two-stage method (simulated pre-training, head adaptation, full fine-tuning) led to further improvements. These findings show that deep learning with transfer learning and spatial context optimization (using a 9×9 patch) can greatly improve chlorophyll-a estimation in complex tropical lakes.
Maghfiroh et al. (Thu,) studied this question.