Training Strategies of CNN for Land Cover Mapping with High Resolution Multi-spectral Imagery in Senegal

Regional land cover mapping has been relying on medium resolution multi-spectral data (10-30m GSD) such as Landsat and Sentinel-2. However, very high-resolution (VHR) multi-spectral imagery (1-2m GSD) (WorldView-2,-3,WV), is becoming more accessible, allowing detailed land cover mapping, including small agriculture fields and individual trees. There has been a massive proliferation of deep learning methods like convolutional neural networks, notable U-Net, for satellite image segmentation. However, there are major challenges in scaling-up the training process for U-Net models using a patchwork of hundreds of VHR images to create a seamless land cover map. Firstly, the U-Net is a fully supervised learning method that requires significant resources to create training labels for large regions. Second, the availability of archived VHR imagery (2011-21) is highly irregular in time and the multi-spectral signatures of various land cover classes change drastically due to phenology and variable imaging conditions. Third, the study area of Casamance, Senegal experiences rapid changes in vegetation greenness and widespread burned area of various ages, making it hard to distinguish natural vegetation and croplands (active, fallow or burnt). The overall goal was to test various strategies to scale-up U-Net models to classify large amounts of WV imagery to produce land cover maps. The training and validation sets consisted of randomly sampled sets of small tiles from each 5000x5000 pixel WV image. The training set size varied from 5 to 20 images, with a separate 10 images for testing. The U-Net model continuously coped with the increasing image variations in the growing training set with accuracy >90% for all training images. However, the accuracy dropped to 40-80% when applied to test images, due to unique image conditions of some imagery. These are the preliminary results of the capability of U-Net on available training images. Ongoing research is investigating strategies to further scaling-up training set with increase image conditions and to attempt to standardize spectral properties of problematic images to improve regional application of U-Net models.