Work

Enhanced burn severity mapping using machine learning and sentinel-2 spectral indices in tropical montane forests of Tanzania

Year: 2026
Type: article
Abstract: Forest fires pose a significant threat to ecosystems, biodiversity, and human settlements, necessitating accurate and timely detection of burned areas for post-fire management. Remote sensing-based techniques coupled with google earth cloud computing platform, has emerged a best technique for spatial and temporal monitoring of the impact of forest fires. However, application of these techniques in the mountain’s forests are still limited particularly in Tanzania. Therefore, this study aimed to evaluate the effectiveness of Sentinel-2 derived burn severity indices (dNBR and dNDVI) integrated with a Random Forest machine learning framework for mapping and assessing forest fire severity in the West Usambara mountain forests of Magamba, Shagayu, Mkusu, and Baga. Pre-fire conditions were assessed using imagery acquired between 2021-01-01 and 2022-12-30, while post-fire conditions were evaluated using imagery from 2023-01-01 to 2023-12-30. Sentinel-2 data obtained through the Google Earth Engine platform were used to compute the differenced Normalized Burn Ratio (dNBR) and differenced Normalized Difference Vegetation Index (dNDVI), which were subsequently applied to generate burn severity maps across the four forest reserves. The results revealed that dNBR and dNDVI captured complementary burn severity patterns. dNBR highlighted low to moderate-high severity areas, with the largest affected zones in Baga (25.69% low, 0.29% moderate-high) and Magamba (14.38% low, 0.28% moderate-high), while dNDVI was more sensitive to severe canopy damage, detecting very high severity in Shagayu (0.15%) and Magamba (0.031%). Across all forests, unburned areas remained the dominant class, although their proportions varied among forest reserves and burn severity indices, suggesting highly patchy fire behaviour potentially influenced by the complex topography, heterogeneous vegetation structure, and varying microclimatic conditions characteristic of tropical montane forests. Random Forest classification achieved high overall accuracies ranging from 98.2% in Baga to 99.7% in Shagayu, demonstrating the effectiveness of integrating spectral, topographic, and texture predictors for robust burned-area classification. By integrating complementary spectral indices with machine learning, this study advances burn severity assessment in tropical montane forests by improving the detection of spatially heterogeneous fire effects, thereby supporting more effective ecological restoration and post-fire management. Although evaluated in only four forest reserves, the framework provides a cost-effective and scalable approach that can be readily transferred to other tropical montane forests using freely available Sentinel-2 imagery and Google Earth Engine.
Language: English
Open Access status: gold
Estimated APC paid: $590
OpenAlex ID: W7223084963
FWCI:
Cites: 64
Cited by:
Related to: 9