| dc.description.abstract | Monitoring and assessing the characteristics of surface water is critical for managing and improving its quality. While the existing in situ measurement methods are precise, they have limitations in terms of spatial and temporal coverage and high acquisition costs. Remote sensing data, specifically satellite imagery, has the potential to provide an invaluable complementary source of information at multiple scales of analysis. Incorporating in situ measurements with satellite imagery provides a way to estimate water quality parameters at broader scales and with greater temporal frequency. The knowledge gap lies in effectively integrating in situ measurements with satellite imagery using deep learning for accurate water quality assessments. Closing this knowledge gap, the application of deep learning, demonstrated efficacy in diverse domains, shows potential for enhancing the water quality parameter estimation process. Although multiple regression techniques and machine learning methods were studied previously to predict water quality parameters, a high-performance deep learning technique that learns higher order statistical relationships between surface reflectance values and water quality parameters is still lacking, and thus, needed. This research proposes a new and cost-effective technique coupling Sentinel-2 imagery and Deep Neural Network (DNN), a deep learning technique to estimate two water quality parameters: Chlorophyll-a and turbidity. The study was focused on Lake Okeechobee, Florida's largest freshwater lake. Surface reflectance values from Sentinel-2 image bands (5304 band images for Chl-a and 5772 band images for turbidity) combined with auxiliary band variables were used to predict the water quality parameters based on in situ datasets. A total of 25 feature variables were used for Chl-a and 19 feature variables were used for turbidity. We adopted a common partition ratio of 70/30 for training and testing purposes. We evaluated two machine learning models: Support Vector Regression (SVR), Random Forest (RF), and one DNN model. The proposed exponentially decreasing DNN approach produced results with significantly higher R2 values, and lower MSE, RMSE, MRE, and MAE values. When applied to the test dataset, the DNN approach showed the highest correlation (R2=0.62, MSE=0.14, RMSE=0.38, MRE=12.15%, MAE=0.30) for Chl-a and for turbidity (R2=0.72, MSE=0.14, RMSE=0.37, MRE=9.60%, MAE=0.28). For the Chl-a parameter, the DNN approach outperformed the RF model by 12.7% and the SVR model by 16.9%, when assessed using R2. Similarly, for the turbidity parameter the DNN approach outperformed the RF model by 18.03% and the SVR model by 28.5% when assessed using R2. The weights of the trained DNN model can be adapted and fine-tuned for other inland lakes and can be tested on other research areas in the future. | en_US |
