Snow and Cloud Classification in Historical SPOT Images: An Image Emulation Approach for Training a Deep Learning Model Without Reference Data

The lack of revisit in long-term satellite time series, such as Landsat is an issue to assess ecosystems response to snow cover variations in mountains. A recent release of the Satellites Pour l'Observation de la Terre (SPOT) 1-5 satellite images collection by the SPOT World Heritage (SWH) program offers the opportunity to increase the temporal revisit of Landsat from 1986 to 2015 at 20 m resolution. However, spectral and radiometric limitations of these images hinder the application of well-established pixel-wise methods to extract the snow cover area. As a work-around, deep learning techniques, such as convolutional neural networks can incorporate both spectral and spatial information to classify every pixel as snow, cloud, or snow-free. However, the lack of reference data poses a challenge to the implementation of such data-driven approaches. Here, we develop an emulator of SPOT images, which takes as input Sentinel-2 images. As a result, an emulated SPOT image can be paired with a reference snow map generated from its source Sentinel-2 image to train a deep learning model able to process actual SPOT images. We follow this approach to train a U-Net and evaluate different training strategies. We apply the different models to classify actual SPOT images for which we have reference data for validation. The method yields high precision in detecting snow, with minimal false snow pixel identification. This is at the cost of overestimating cloud pixels around clouds and highly saturated areas. The results confirm the potential of this method to generate time series of snow cover maps using the SWH collection.