Training sensor-agnostic deep learning models for remote sensing: Achieving state-of-the-art cloud and cloud shadow identification with OmniCloudMask

Abstract

Deep learning models are widely used to extract features and insights from remotely sensed imagery. However, these models typically perform optimally when applied to the same sensor, resolution and imagery processing level as used during their training, and are rarely used or evaluated on out-of-domain data. This limitation results in duplication of efforts in collecting similar training datasets from different satellites to train sensor-specific models. Here, we introduce a range of techniques to train deep learning models that generalise across various sensors, resolutions, and processing levels. We applied this approach to train OmniCloudMask (OCM), a sensor-agnostic deep learning model that segments clouds and cloud shadow. OCM demonstrates robust state-of-the-art performance across various satellite platforms when classifying clear, cloud, and shadow classes, with balanced overall accuracy values across: Landsat (91.5 % clear, 91.5 % cloud, and 75.2 % shadow); Sentinel-2 (92.2 % clear, 91.2 % cloud, and 80.5 % shadow); and PlanetScope (96.9 % clear, 98.8 % cloud, and 97.4 % shadow). OCM achieves this accuracy while only being trained on a single Sentinel-2 dataset, employing spectral normalisation and mixed resolution training to address the spectral and spatial differences between satellite platforms. This approach allows the model to effectively handle imagery from different sensors within the 10 m to 50 m resolution range, as well as higher resolution imagery that has been resampled to 10 m. The OCM library is available as an open source Python package on PyPI.