Segmentation of waterbodies in remote sensing images using deep stacked ensemble model

Identifying surface water resources is considered as one of the principal applications of remote sensing image analysis that plays a crucial role in controlling optimal use of these resources, and preventing floods and crises such as drought. Traditional machine learning methods for extracting waterbodies require complex spectral analysis and selection of features based on previous knowledge. Although applying deep learning-based approaches, which has been considered in recent years, has eliminated the necessity of extracting manual features, they require too many training data and computational resources to achieve high performance. Consequently, each presented deep architecture can detect some of the existing patterns in the predefined conditions. This paper trains and optimizes three robust deep architectures, presented in various fields, using surface water data, and combines their results to achieve a robust model for detecting surface water. To this end, a deep hybrid architecture called deep stacked ensemble model is employed on the outputs of three independent deep sub-models and extracts the final segmentation mask of the water areas more accurately. We evaluated our proposed model on a water body detection dataset provided by artificial intelligence crowd landsat (AIcrowd(1) LNDST) challenge. The proposed technique improves the semantic segmentation performance and surpasses state-of-the-art results. (C) 2022 Elsevier B.V. All rights reserved.

Automated summary

This paper presents a deep learning-based approach for detecting surface water by training and optimizing three robust deep architectures and combining their results. By employing a deep stacked ensemble model, the proposed technique achieves more accurate segmentation masks of water areas and outperforms state-of-the-art results in a water body detection dataset.