Convolutional neural networks for water segmentation using sentinel-2 red, green, blue (RGB) composites and derived spectral indices. Issue 14 (18th July 2021)

• Record Type: Journal Article
• Authors: James, Thomas; Schillaci, Calogero; Lipani, Aldo
• Is Part Of: International journal of remote sensing. Volume 42:Issue 14(2021)
• Page Start: 5338
• Page End: 5365
• Publication Date: 2021-07-18
• Subjects: Remote sensing -- Periodicals; Télédétection -- Périodiques; 621.3678
• DOI: 10.1080/01431161.2021.1913298
• Languages: English
• ISSNs: 0143-1161

Abstract

ABSTRACT: Near-real time water segmentation with medium resolution satellite imagery plays a critical role in water management. Automated water segmentation of satellite imagery has traditionally been achieved using spectral indices. Spectral water segmentation is limited by environmental factors and requires human expertise to be applied effectively. In recent years, the use of convolutional neural networks (CNN's) for water segmentation has been successful when used on high-resolution satellite imagery, but to a lesser extent for medium resolution imagery. Existing studies have been limited to geographically localized datasets and reported metrics have been benchmarked against a limited range of spectral indices. This study seeks to determine if a single CNN based on Red, Green, Blue (RGB) image classification can effectively segment water on a global scale and outperform traditional spectral methods. Additionally, this study evaluates the extent to which smaller datasets (of very complex pattern, e.g harbour megacities) can be used to improve globally applicable CNNs within a specific region. Multispectral imagery from the European Space Agency, Sentinel-2 satellite (10 m spatial resolution) was sourced. Test sites were selected in Florida, New York, and Shanghai to represent a globally diverse range of waterbody typologies. Region-specific spectral water segmentation algorithms were developed on each test site, to represent benchmarks of spectral index performance. DeepLabV3-ResNet101 was trained on 33, 311 semantically labelled true-colour samples. The resulting model was retrained on three smaller subsets of the data, specific to New York, Shanghai and Florida. CNN predictions reached a maximum mean intersection over union result of 0.986 and F1-Score of 0.983. At the Shanghai test site, the CNN's predictions outperformed the spectral benchmark, primarily due to the CNN's ability to process contextual features at multiple scales. In all test cases, retraining the networks to localized subsets of the dataset improved the localized region's segmentation predictions. The CNN's presented are suitable for cloud-based deployment and could contribute to the wider use of satellite imagery for water management.