Tri-band assessment of multi-spectral satellite data for flood detection

Multi-spectral satellite data provides vast resources for important tasks such as flood detection, but training and fine tuning models to perform optimally across multi-spectral data remains a significant research challenge. In light of this problem, we present a systematic examination of the role of tri-band deep convolutional neural networks in flood prediction. Using Sentinel-2 data we explore the suitability of different deep convolutional architectures in a flood detection task; in particular we examine the utility of VGG16, ResNet18, ResNet50 and EfficientNet. Importantly our analysis considers the questions of different band combinations and the issue of pre-trained versus non-pre-trained model application. Our experiment shows that a 0.96 F1 score is achievable for our task through appropriate combinations of spectral bands and convolutional neural networks. For flood detection, three-band combinations of RB8aB11 and RB11B outperformed 33 other combinations when trained with pre-trained ResNet18 and other models. Our analysis further demonstrates a strong performance by pre-trained models despite the fact that these pre-trained models were originally trained on different spectral bands.