Burnt Area Segmentation with Densely Layered Capsules

In the presented article, deep learning (DL) is employed on Sentinel-2 and Landsat imagery to assess and delineate the wildfire-affected burnt area. The most important aspect is the use of densely layered capsule architecture as an alternative to traditional convolutional neural networks (CNN) and max pooling. The presented investigation is a practical capsule-based wildfire detection and burnt area segmentation. Moreover, the research work presents a comprehensive assessment of capsule-based segmentation architecture. Due to the limitation of freely available open access annotated data, a dataset is manually constructed from scratch. The proposed algorithm provides prognostic delineation of the burnt area with an overall accuracy of 96%. The quantitative measures demonstrate that the proposed capsule-based architecture outperforms other state-of-the-art segmentation models (U-Net, DeepLabV3+). The capsule-based architecture shows good promise and can be scaled to other similar topographic regions.