Water classification using a Random Forest model to limit river erosion in Bangladesh
Summary
Riverbank erosion is a significant issue in Bangladesh that is mainly caused by heavy rainfall in the Meghna estuary. Erosion predictions can help when planning and taking measures to protect land from erosion. One method of obtaining erosion predictions is via the Bangladesh Erosion Monitor, this is a dashboard that predicts what land is at risk of erosion using the Joint Research Center Global Surface Water Explorer (JRC-GSWE) monthly water extents. The use of JRC-GSWE water extents in the Bangladesh Erosion Monitor is not ideal due to the fact the data were last updated in December of 2021, making it impossible to provide recent erosion predictions.
This study aims to provide an alternative water classification method that creates water extents like the JRC-GSWE water extents. The proposed model is a Random Forest model using Sentinel-1 radar data as opposed to the Landsat data used for the JRC-GSWE water extents. An advantage Sentinel-1 data has over Landsat data is that radar data is not affected by cloud cover. The following research question needs to be answered to achieve the study aim:
“How can a Random Forest classification model utilizing Sentinel-1 data be developed to produce near real-time water extents for Bangladesh that are comparable to or surpass the JRC-GSWE water extents?”
To achieve this, a Random Forest model was created that uses available bands from Sentinel-1 data to create near real-time water extents. Different preprocessing methods were tested on the Sentinel-1 data to achieve the best similarity to the JRC-GSWE water extents and performance. Similarity and performance were assessed with the overall accuracy and Kappa coefficient. Furthermore, a validation polygon dataset was created to assess the performance of the proposed Random Forest model water extents and JRC-GSWE water extents when comparing them to the polygons. The proposed Random Forest model showed to be capable of creating reliable predictions with an average overall accuracy of 96,7% and water extents similar to the JRC-GSWE water extents with an average Kappa coefficient of 0,879. When looking at the performance of both models when compared to the validation polygon dataset, they both show to be accurate with an accuracy of roughly 97%. These results show that the proposed Random Forest model can create water extents that are very similar to the JRC-GSWE water extents with the added benefits that they can be created in near real-time and are not affected by cloud cover.