Water detection in Sentinel-1 data using a Bayesian Convolutional Neural Network: Application of uncertainty estimations to identify error prone areas and improve the results

Kurzfassung

Floods are a natural hazard that can seriously impact the affected communities. Therefore, improvements in flood management are necessary to better prevent and manage flood disasters. These can be achieved by mapping flooded areas using remote sensing data such as Synthetic Aperture Radar (SAR) data. SAR has the advantage of covering large spatial extents and operating weather and daylight independently. While conventional methods exist to detect water in SAR data, Convolutional Neural Networks (CNNs) have produced excellent results. The results, however, do not come without inaccuracies and uncertainties. Therefore, Bayesian Convolutional Neural Networks (BCNNs) have been developed to estimate the uncertainties of the model. This study analyzes the conditions that prevail in misclassified areas. Certain landcover classes like bare soil show higher percentages of wrongly labeled pixels. The behavior of the estimated uncertainties is also tested over pixels that are wrongly and correctly labeled as well as over different landcover classes. It is found that uncertainties are higher over misclassified pixels and certain landcover types like bare soil and herbaceous vegetation. Based on the findings that uncertainties are elevated over falsely labeled pixels, the pixels are turned to their opposite class when exceeding an uncertainty threshold. After the re- labeling, the performance metrics are compared to the initial metrics. In this study, mul- tiple setups for relabeling are tested and compared. The approach is found to be working in certain areas. The study is conducted to confirm the applicability of BCNNs to generate precise flood mapping products and to estimate model uncertainties. The relabeling also aims to shorten the process of training data creation. Training data creation is a resource-intensive step. By improving the results after the classification, less accurate training data might be usa- ble to train the model. As a result, more training data can be efficiently generated to cover more expansive areas globally. The findings provide a basis to create more complete models in the future and further assist flood management.