Detection of Complex Formations in an Inland Lake from Sentinel-2 Images Using Atmospheric Corrections and a Fully Connected Deep Neural Network

The detection of complex formations, initially suspected to be oil spills, is investigated using atmospherically corrected multi-spectral satellite images and deep learning techniques. Several formations have been detected in an inland lake in Northern Greece. Four atmospheric corrections (ACOLITE, iCOR, Polymer, and C2RCC) that are specifically designed for water applications are examined and implemented on Sentinel-2 multispectral satellite images to eliminate the influence of the atmosphere. Out of the four algorithms, iCOR and ACOLITE are able to depict the formations sufficiently; however, the latter is chosen for further processing due to fewer uncertainties in the depiction of these formations as anomalies across the multi-spectral range. Furthermore, a number of formations are annotated at a pixel level for the 10 m bands (Red, Green, Blue, and NIR), and a deep neural network (DNN) is trained and validated. Our results show that the 4-band configuration provides the best model for the detection of these complex formations. Despite not being necessarily related to oil spills, studying these formations is crucial for environmental monitoring, pollution detection, and the advancement of remote sensing techniques.