A Deep Learning Pipeline for Drought Assessment using Spatial Satellite Images and Vision Transformers

Abstract

700 million people are in danger of being displaced due to inept drought prediction and prevention systems. Current research on drought assessment focuses solely on factors such as soil moisture and rainfall, which require painstaking measurements and lab samples, and can often be misleading. This research eliminates this requirement by proposing an end-to-end pipeline to detect and prevent droughts in at-risk areas using satellite images and vision transformers. The dataset is comprised of over 86,000 satellite images labeled by pastoralists and divided with an 80-20 ratio for training and validation. First, using feature filtering, normalization, and a Gaussian filter, the images in the dataset are modified to yield a better performance. Next, a deep vision transformer model with multi-headed attention is constructed, consisting of four heads, three transformer layers, and a patch size of five. The final MLP head produces logits for drought severity prediction level. Overall, the best transformer model achieves 78.3% accuracy in predicting drought conditions on a validation set of 10,000, unseen satellite images. In addition, this method outperforms state-of-the-art convolutional neural networks on this classification task, as compared to VGG-16, ResNet-50 and DenseNet-121 models. The model harnesses AWS cloud computing, deep vision transformers, and specific image augmentation to achieve state-of-the-art results in drought prediction and prevention. With this research, scientists have the potential to assess droughts quickly and accurately, revolutionizing our ability to provide resources and care to those affected by the increasingly common droughts caused by the climate crisis worldwide.