ESTIMATION OF CHLOROPHYLL-A CONCENTRATIONS IN WETLANDS USING REMOTE SENSING.

Abstract

Remote sensing techniques with an integration of machine learning algorithms is increasingly being tested as a method for water quality monitoring, particularly for the estimation of Chlorophyll-a concentrations. In this study, similar methods are used to estimate Chlorophylla from small-scale wetlands with open water in Sweden, where in-situ Chlorophyll-a field data has been collected. Images captured by the drone and Sentinel 2A images were used to obtain wavelength bands (Blue, Green, Red, Red-edge and Near-infrared) data which was used as input in the Random Forest models. This study compared the performance of random forest models developed using Drone imagery and Sentinel-2A multispectral imagery for Chlorophyll-a estimation in two wetlands in Sweden. The random forest model clearly predicted Chlorophyll-a from both Drone imagery (R²=0.845 training) and Sentinel 2A imagery (R²= 0.903 training). Sentinel-2A model achieved higher training R² value than Drone imagery model, indicating stronger predictive performance. The cross-validation results for both indicated substantially lower predictive ability (R² ≈ 0.02–0.05) indicating overfitting. Variable importance analysis revealed substantial differences between the two datasets. In the Drone Imagery model, the Red-edge band variable was identified as the most important predictor for Chlorophyll-a while Near-infrared variable was the most important for Sentinel2A model. Overall, the findings show that effectiveness of retrieving Chlorophyll-a by remote sensing using Random Forest models is possible, however it depends on the interaction between spatial resolution, spectral characteristics, and model structure.