Automated Cloud Detection in Landsat-8 Imagery Using Deep Neural Networks

Landsat-8 satellite imagery is a vital earth observation tool for monitoring land surface and coastal processes. However, effective retrieval of important biophysical and surface information is hampered by presence of clouds and cloud shadows. Traditional algorithms such as FMask work well however they (i) function on a purely pixel-level basis, (ii) rely heavily on tuning of input parameters, (iii) need the thermal band to be truly effective, and (iv) do not identify thin clouds. Unfortunately, Landsat 8 thermal bands have been facing a number of issues which call into question the effectiveness of FMask during this period. To this effect, we outline a novel automated cloud detection method based on Deep Convolutional Neural Networks (DCNNs) which only utilizes non-thermal bands while classifying cloud, thin cloud, cloud shadow, and clear-sky pixels.
Our cloud detection network, FertyleNet, has been designed to simultaneously incorporate both local and global contextual features. The net has been trained and validated using patches extracted over 12 Grass/Cropland scenes using the manually labelled Landsat 8 Cloud Cover Assessment Validation dataset. We aimed to address issues of training bias by including equal number of patches per class distributed evenly across all scenes. We also ensure that the model is trained with a large number of patches containing multiple class memberships so it can adapt to complicated cases. FertyleNet yields an overall accuracy of ~95% when validated across the entire 12 scenes whereas FMask achieved a baselines accuracy of ~90% thus showing that DCNNs are a viable tool for classifying clouds and cloud shadows for Landsat 8 data.