Automatic Detection of Natural Disaster Effect on Paddy Field from Satellite Images using Deep Learning Techniques

TL;DR

This study develops a deep learning-based method to detect rice field damage from natural disasters in Bangladesh using satellite imagery, NDVI analysis, and ground truth data, achieving promising segmentation results.

Contribution

The paper introduces a novel approach combining NDVI analysis with deep learning to detect crop loss, and provides ground truth data for training segmentation models using high-resolution RGB images.

Findings

FCI-based model achieved IoU 0.51

RGB images achieved IoU 0.41

Ground truth data enables crop loss detection with high-resolution imagery

Abstract

This paper aims to detect rice field damage from natural disasters in Bangladesh using high-resolution satellite imagery. The authors developed ground truth data for rice field damage from the field level. At first, NDVI differences before and after the disaster are calculated to identify possible crop loss. The areas equal to and above the 0.33 threshold are marked as crop loss areas as significant changes are observed. The authors also verified crop loss areas by collecting data from local farmers. Later, different bands of satellite data (Red, Green, Blue) and (False Color Infrared) are useful to detect crop loss area. We used the NDVI different images as ground truth to train the DeepLabV3plus model. With RGB, we got IoU 0.41 and with FCI, we got IoU 0.51. As FCI uses NIR, Red, Blue bands and NDVI is normalized difference between NIR and Red bands, so greater FCI's IoU score than RGB is expected. But RGB does not perform very badly here. So, where other bands are not available, RGB can use to understand crop loss areas to some extent. The ground truth developed in this paper can be used for segmentation models with very high resolution RGB only images such as Bing, Google etc.