A one-class classification model for burned-area detection based on mutual ordering of normalized differences

• Published in: IEEE transactions on geoscience and remote sensing Vol. 61; p. 1
• Main Author: Zanetti, Massimo
• Format: Journal Article
• Language: English
• Published: New York IEEE 01.01.2023; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Barium; Classification; Climate studies; Detection; hyper-dimensional binary spaces; Indexing; Mathematical analysis; Mathematical properties; Meteorology; Neural networks; normalized difference indexes; One-class classification; remote sensing; Robustness (mathematics); Space heating; Support vector machines; Training; Vegetation mapping
• ISSN: 0196-2892; 1558-0644
• DOI: 10.1109/TGRS.2023.3301056

Global scale assessment of burned area (BA) is essential for climate studies and open access satellite-borne multispectral (MS) imagery is vital for the mapping purpose. Global characterization of BAs via MS analysis is difficult as it usually requires ancillary data to model local factors such as vegetation types and local eco-climate systems. This paper proposes a novel classification model that exploits certain mathematical properties of normalized difference indexes (NDIs) to build an abstract space of features where the BA class can be learned globally and solely using MS images. The core idea is that, although NDIs are subject to strong intra-class variations, their mutual order (i.e., the sign of their difference) can be robust enough for characterization. By encoding every possible such ordering relation in a binary domain, the feature space turns out to be hyper-dimensional, with abstraction capabilities similar to that of Neural Network (NN) layers. The proposed classification model is one-class, therefore very convenient as it only requires training samples for the positive class to be collected. The model is experimentally validated in an extensive BA detection exercise with Sentinel-2 images that involves recently published global BA reference data. Results are promising as we report higher F1-score than those reported for state-of-the-art BA products currently available, which are obtained through hybrid techniques and multiple data sources. The model also outperforms the well-known and largely used one-class SVM (OC-SVM), which is tested in this work for the first time for BA detection at global scale.