Deep Learning Algorithms to Detect and Remove Sun Glint from High-Resolution Aerial Imagery

Abstract:

Aerial remote sensing of aquatic environments presents many challenges that can lead to significant data losses. The advent of drones have allowed for the acquisition of increasingly high-resolution imagery whenever required, with polarising filters available to minimise image contamination. Despite these advances, some degree of sun-glint contamination usually remains in any drone imagery captured over aquatic environments. These instances can cause significant inaccuracies during an image classification process, particularly in fine-scale ecological studies in which the area of interest is not surface water but underwater organisms, such as corals or seagrass beds. No method currently exists to effectively remove sun glint contamination from high-resolution imagery.

With the use of AI, it is possible to identify instances of sun glint and remove them from further analysis. We present a deep learning algorithm capable of detecting and classifying sun glint in any drone image. Our training images were classified using an object-based image analysis workflow, and our deep learning algorithm is based on a U-net architecture. Our model performed with high levels of training and validation accuracies, with a 99.74% chance of accurately predicting sun glint as not present, and a 63.07% chance of predicting sun glint as present. Further, this model is able to achieve these accuracies despite a highly imbalanced dataset, with sun glint only accounting for 0.8% of pixels in the dataset. Overall, 99.42% of predictions in our model are correct. We believe that this packaged, fully trained AI algorithm can be of use to any researcher needing to instantly detect and remove sun glint contamination from high-resolution imagery collected in an aquatic environment.