Combined use of R-VSPI and VSPI for enhanced quantification of fire severity in south-eastern Australian forests

Abstract

Wildfires, intensified by climate change, necessitate advanced methods for accurate and near-real-time fire severity mapping to improve emergency response and post-fire recovery strategies. Satellite remote sensing, combined with supervised learning approaches, enhances the accuracy and efficiency of fire severity mapping. This study introduces Decision-Based Hierarchical Learning (DBHL), a novel multi-sensor fire severity classification model that integrates Synthetic Aperture Radar (SAR; Sentinel-1 backscatter) and optical (Sentinel-2 reflectance) data. The model was applied to assess wildfire impacts on temperate forests during the 2019/20 "Black Summer" wildfire season in south-eastern Australia. DBHL incorporated SAR-based RADAR-Vegetation Structure Perpendicular Index (R-VSPI) and optical-based Vegetation Structure Perpendicular Index (VSPI) as candidate indices. By integrating these complementary datasets, DBHL leverages both structural and physiological changes as fire severity indicators, addressing limitations in single-sensor approaches. A pixel-wise approach was employed to spatially upscale the applicability of the R-VSPI and VSPI indices for fire severity assessment across the entire region. Using field data, the sensitivities of the R-VSPI and VSPI indices were validated during the immediate post-fire to one-year post-fire period. DBHL was trained and evaluated with a focus on comparing its performance against independent R-VSPI and VSPI classifications. The findings reveal the unique strengths of each index across various fire severity classes, demonstrating their complementary value. R-VSPI is more sensitive to structural changes in forests, while VSPI excels in identifying changes related to canopy-level disturbances. One-year post-fire recovery analysis shows distinct spatial patterns, with VSPI indicating faster recovery in surface vegetation and R-VSPI highlighting prolonged structural recovery. The DBHL model demonstrates the complementary strengths of the indices, allowing fire severity assessments to be contextualized across vertical vegetation strata, distinguishing between canopy-based damage indicators and underlying structural changes. DBHL outperformed single-sensor approaches, achieving the highest classification accuracy (overall accuracy=88.89%, kappa=0.86), particularly improving differentiation of Moderate (partial canopy scorch) and High (full crown scorch) severity with a producer's accuracy of 100%, and 80%, respectively. Future research is aimed at integrating multi-wavelength SAR, including L-band (1.25 GHz) and P-band (0.43 GHz), along with LiDAR measurements to enhance structural fire severity assessments. © 2025 The Author(s). Published by Elsevier Inc. This is an open access article under the CC BY license. http://creativecommons.org/licenses/by/4.0/