[Paper Review] An automatic water detection approach based on Dempster-Shafer theory for multi spectral images

Na Li, Arnaud Martin+1 more|arXiv (Cornell University)|Aug 9, 2017

Remote-Sensing Image Classification22 references3 citations

TL;DR

This paper proposes an unsupervised water detection method for multi-spectral images using Dempster-Shafer theory to fuse a spectral model based on near-infrared reflectance with a supervised learning classifier. The approach automatically identifies principal water bodies while labeling ambiguous features—such as half-dry streams, shadows, clouds, and built-up areas—as 'ignorance,' providing critical uncertainty-aware outputs for downstream classification with 7.41% water, 87.92% non-water, and 4.67% ignorance detected.

ABSTRACT

Detection of surface water in natural environment via multi-spectral imagery has been widely utilized in many fields, such land cover identification. However, due to the similarity of the spectra of water bodies, built-up areas, approaches based on high-resolution satellites sometimes confuse these features. A popular direction to detect water is spectral index, often requiring the ground truth to find appropriate thresholds manually. As for traditional machine learning methods, they identify water merely via differences of spectra of various land covers, without taking specific properties of spectral reflection into account. In this paper, we propose an automatic approach to detect water bodies based on Dempster-Shafer theory, combining supervised learning with specific property of water in spectral band in a fully unsupervised context. The benefits of our approach are twofold. On the one hand, it performs well in mapping principle water bodies, including little streams and branches. On the other hand, it labels all objects usually confused with water as `ignorance', including half-dry watery areas, built-up areas and semi-transparent clouds and shadows. `Ignorance' indicates not only limitations of the spectral properties of water and supervised learning itself but insufficiency of information from multi-spectral bands as well, providing valuable information for further land cover classification.

Motivation & Objective

• To address the challenge of misclassifying water bodies due to spectral similarity with built-up areas, shadows, and clouds in multi-spectral imagery.
• To develop an automatic, fully unsupervised approach that avoids manual threshold selection common in spectral index methods.
• To integrate spectral response properties of water with supervised learning results using Dempster-Shafer theory for uncertainty-aware classification.
• To label ambiguous or indistinguishable features as 'ignorance' to reflect model limitations and guide future classification refinement.

Proposed method

• A spectral model is applied to identify water based on strong near-infrared (NIR) absorption, using a threshold derived from spectral response characteristics.
• Supervised learning (e.g., SVM or similar) is trained on outputs from the spectral model to refine water detection.
• Dempster-Shafer theory fuses evidence from the spectral model and the supervised classifier, treating them as dependent sources with discounting coefficients.
• The discounting coefficient for the spectral model is calculated using the supervised model’s output as pseudo-ground truth, enabling uncertainty quantification.
• A three-class output is generated: 'water', 'non-water', and 'ignorance'—where 'ignorance' captures features too ambiguous to classify with confidence.
• The fusion process uses mass functions to represent belief in each class, with the final decision based on plausibility and belief measures under uncertainty.

Experimental results

Research questions

• RQ1 Can Dempster-Shafer theory effectively fuse a spectral model and a supervised classifier in a fully unsupervised context for water detection?
• RQ2 How can ambiguity in spectral signatures—especially between water and similar features like shadows or built-up areas—be formally represented and managed?
• RQ3 To what extent does the proposed method improve detection of small water bodies (e.g., streams) compared to traditional spectral indices?
• RQ4 Can the 'ignorance' class serve as a meaningful indicator of model and data limitations for subsequent classification refinement?

Key findings

• The method successfully detected 7.41% of the image as 'water', including small and narrow streams often missed by conventional methods.
• 87.92% of the image was classified as 'non-water' with high confidence, indicating strong separation from water bodies.
• 4.67% of the image was labeled as 'ignorance', encompassing half-dry streams, thin clouds, shadows, and built-up areas that are spectrally similar to water.
• Manual verification using ENVI software confirmed that the method accurately identified principal rivers and their branches while clearly isolating ambiguous features.
• The 'ignorance' class effectively captures the limitations of both the spectral model and the supervised classifier, as well as insufficiencies in multi-spectral data, particularly the absence of SWIR bands.
• The approach outperforms traditional spectral indices by avoiding manual threshold tuning and by providing uncertainty-aware outputs that support further refinement with additional data (e.g., MIR bands).