A deep learning algorithm to detect and classify sun glint from high-resolution aerial imagery over shallow marine environments

Abstract

Sun glint contamination is a significant problem for high-resolution remote sensing over aquatic environments. Sun glint is a particular issue for researchers using aerial imagery to assess shallow water benthic communities, as it may be wrongly classified as benthic substrates of interest. Although various methods are available to correct for sun glint using multispectral and hyperspectral imagery, no method has been developed to detect sun glint within high-resolution RGB imagery, such as that collected with drones. Here we developed an artificial neural network capable of automated detection of sun glint in high-resolution imagery. Training data were classified using an object-based image analysis workflow and a semantic image segmentation algorithm was developed based on a modified U-net architecture. The model correctly identified 99.58% of background in our test images, and 66.07% of sun glint in our test images. These accuracies were achieved despite a highly imbalanced dataset, with sun glint only accounting for 1.19% of pixels in the testing dataset. Overall, 99.18% of predictions in this model were correct. Given this, we contend that this algorithm is a simple solution for the instant detection of sun glint from high-resolution imagery. We offer this semantic image segmentation as an open source solution for the detection and classification of sun glint in high-resolution imagery.