A Self-Trained Model for Cloud, Shadow and Snow Detection in Sentinel-2 Images of Snow- and Ice-Covered Regions

Abstract

Screening clouds, shadows, and snow is a critical pre-processing step in many remote-sensing data processing pipelines that operate on satellite image data from polar and high mountain regions. We observe that the results of the state-of-the-art Fmask algorithm are not very accurate in polar and high mountain regions. Given the unavailability of large, labeled Sentinel-2 training datasets, we present a multi-stage self-training approach that trains a model to perform semantic segmentation on Sentinel-2 L1C images using the noisy Fmask labels for training and a small human-labeled dataset for validation. At each stage of the proposed iterative framework, we use a larger network architecture in comparison to the previous stage and train a new model. The trained model at each stage is then used to generate new training labels for a bigger dataset, which are used for training the model in the next stage. We select the best model during training in each stage by evaluating the multi-class segmentation metric, mean Intersection over Union (mIoU), on the small human-labeled validation dataset. This effectively helps to correct the noisy labels. Our model achieved an overall accuracy of 93% compared to the Fmask 4 and Sen2Cor 2.8, which achieved 75% and 76%, respectively. We believe our approach can also be adapted for other remote-sensing applications for training deep-learning models with imprecise labels.