THE RETRIEVAL of CHLOROPHYLL-A CONCENTRATIONS in INLAND WATER USING CONVOLUTION NEURAL NETWORK on SATELLITE IMAGERY

The retrieval of chlorophyll-a (Chl-a) concentration, which is a crucial indicator in monitoring water quality across inland waters, remains a challenging task by using satellite data. Former studies based on semi-empirical and analytical approaches have achieved great progress. However, the Chl-a retrieval models from these approaches suffer from a wide range of uncertainties that originate from inter-seasonal variations of optical water properties and insufficient quantity of in situ samples. Most inland lakes in tropical regions experience different trophic states across wet and dry seasons. The optically water properties are complex and varied due to different trophic states over different seasons that pose difficulty to remote sensing-based models in accurately estimating Chl-a concentrations. To overcome this problem, a season-insensitive model based on a multi-task convolution neural network with a multi-output structure is proposed. In addition, a layer-sharing network structure with data augmentation is adopted to alleviate the problem of insufficient quantity of in situ Chl-a samples in model calibration and validation. To evaluate the proposed method, a largest lake in the Philippines, Laguna Lake, is selected as the study area. The lake is characterized by oligotrophic and mesotrophic conditions in wet season, whereas the states change to mesotrophic and eutrophic conditions in dry season. Several Sentinel-3 OLCI level-2 images matched with 409 in situ Chl-a measurements in range from 1.24 to 22.30 mg m-3 are collected. Over 5-fold cross validation, the average coefficient of determination (R2) and root mean square error (RMSE) of the proposed model are 0.74 and 2.06 mg m-3, respectively. In comparison, the estimation accuracy of our model is improved than that of related semi-empirical models. The slopes (m) of regressed lines generated from estimated and in situ Chl-a samples also demonstrate the ability of our proposed model to properly capture seasonal patterns of Chl-a in Laguna Lake.