Classification of Water Bodies Using Sentinel-2 Image and Artificial Neural Network

Abstract

This research aims to classify water bodies in Iraq using Sentinel-2A satellite images and artificial neural networks (ANNs) techniques. This satellite provides high-resolution multispectral imaging data, which helps neural networks to distinguish water from other features such as lakes, rivers, and reservoirs. An ANN model was trained using pre-labelled datasets to recognize the distinctive patterns of water. In this research, a new index called Water Body Index (WBI) is introduced to extract water bodies, which is based on the difference between six spectral bands of Sentinel-2A: blue (B), red (R), green (G), near infrared (NIR), short infrared 1 (SWIR 1), and short infrared 2 (SWIR 2). Using the neural net library in R I obtained a new linear equation that serves as the index. The new index concept had two distinct proposed versions. WBI-C (complete version). WBI-R (reduced version). The distributions of these new indices were compared with four other previously used indices: NDWI, MNDWI, MUWI-C, and MUWI-R. The results showed that WBI-C and WBI-R were more efficient in classifying water compared to vegetation, buildings, and other elements. The total average accuracy rates were: WBI-R: 99.6%, WBI-C: 99.8%, NDWI: 99.6%, MNDWI: 98.7%, MUWI-C: 98.9%, MUWI-R: 99%. As for the average Kappa Coefficient: WBI-R: 95.4%, WBI-C: 97.7%, NDWI: 85%, MNDWI: 87%, MUWI-C: 87% , MUWI-R: 89%. As for the average F-score:  WBI-R: 97.7%, WBI-C: 98.9%, NDWI: 97%, MNDWI: 93.7%, MUWI-C: 93%, MUWI-R: 94%. These results show that WBI-C and WBI-R outperformed other indices in classifying water bodies with high accuracy.