Harmonizing SAR and optical data to map surface water extent: a deep learning approach

In this work, we demonstrate how harmonized optical and SAR satellite imagery can be utilized for robust identifica- tion of open water surfaces at a global scale. We train an image segmentation architecture based convolutional neural network (CNN) to extract the most salient features from the input data and generate a per-pixel water/not-water classifica- tion. We find that combining optical and radar imagery helps reduce false positive and false negative inferences, illustrating the effectiveness of this harmonization. The resulting model is able to classify water surfaces at the resolution of the SAR sensor (12.5 meters) with a validation set precision and recall of 0.74 and 0.81 respectively. We also demonstrate that the trained model is capable of generating inferences beyond the geographic bounds of the training data.