Biomass burning is a global environmental concern that leads to loss of vegetation, pollution and health hazards. Therefore, it is crucial to quantify biomass burning and its impact on various ecological processes. The study focuses on Central India, which has a high frequency of forest fires, agricultural burning and experiences a long dry pre-monsoon period, making it highly susceptible to fires. The region exhibits strong seasonality in the types of fires recorded, with forest fires peaking in March and agricultural fires peaking in April.
Current fire monitoring tools that use coarse-resolution data underestimate the extent and the frequency of fires, especially for smaller burn patches (<25 Ha), highlighting a need for high-resolution mapping to understand fire dynamics better.
The study aims to evaluate the accuracy of different Spectral Indices (SI) and compare Supervised Classification with ML-based mapping, with and without the inclusion of spectral indices. It distinguishes between the performance of global fire product Fire Information for Resource Management System (FIRMS) and Landsat images in detecting fires. And also highlights the automation potential in burned area detection.
To represent the landscape, a total of nine Landsat scenes were randomly selected, spanning the period from 1993 to 2011. The study compared the accuracy of different Spectral Indices in segregating the burned areas from other classes. It assessed supervised classification using maximum likelihood estimation, with and without including spectral indices and hillshade. Two machine learning algorithms, random forest (RF) and support vector machine (SVM), were also tested for their ability to classify burned areas. It investigated the potential for automating burned area detection across Central India by considering factors such as sun position, location, and season.
The results of the study show that the separability of spectral indices varies depending on the month of image acquisition and whether the burned area is of forest or agriculture. Total eight indices were analysed; NDVI, GEMI, BAI, BAIMS, NBR, MIRBI, BAIML, and NDWI (full forms at the bottom of this page). In March, certain indices like GEMI, BAIML, NDVI, and NBR can effectively differentiate between forest burned area and other land cover classes, while in April, only MiRBI is able to differentiate forest burned area from forest and agriculture. Similar patterns are observed for agricultural burned areas. However, the impact of the month is less drastic for agricultural burned areas.
The supervised classification using only spectral bands achieves varying accuracies depending on the month. In March, the classification performs relatively well for forest burned areas, while in April, the accuracy decreases significantly. The classification struggles to differentiate between forest and agricultural burned areas. Including spectral indices in the classification does not consistently improve accuracy. Considering ML Algorithms, the RF algorithm is sensitive to parameter selection, while the SVM algorithm tends to overfit the data and obtain higher but potentially misleading accuracies. Random Forest (RF) had more consistent results than Support Vector Machine (SVM), which tended to overfit the data. However, in specific parameters, RF misclassified less frequently present land cover types, such as water and shadow. Both ML algorithms outperformed supervised classification in terms of accuracy. Regarding automating burned area detection across Central India, the classification accuracy is influenced by factors such as azimuth angle and month of image acquisition. Including scene-specific variables, such as solar azimuth, solar elevation, and month, can improve classification accuracy across multiple images.
For spectral indices (SIs), the commonly used indices such as NBR and BAI did not perform well in this study. NBR's performance was influenced by forest biomass, leading to lower accuracy in April when vegetation was reduced. On the other hand, indices like GEMI, NDVI, and NDWI consistently showed better performance in separating burned areas from unburned areas due to their ability to capture changes in vegetation. However, relying on these indices alone is not sufficient. ML algorithms showed more consistent results compared to supervised classification. Unlike supervised classification, ML algorithms required less training data and were not limited by the minimum number of samples per class. Comparing the results to FIRMS, the commonly used satellite-based fire detection system, Landsat-based methods demonstrated significantly higher accuracy. FIRMS data underestimated burned areas, especially for agricultural fires, leading to a poor understanding of fire patterns and drivers. When combined with ML algorithms, Landsat imagery showed potential for more accurate and comprehensive mapping of burned areas. For automating burned area detection, the findings suggest prioritizing lower commission errors and choosing images acquired later in the fire season (April-May).
In conclusion, the study emphasizes the importance of considering multiple spectral indices, employing ML algorithms, and using Landsat imagery for improved burned area detection. The machine learning algorithms outperformed supervised classification. While RF and SVM have strengths and limitations, careful parameter selection (and not necessarily selecting parameters with highest accuracy) is crucial to minimize commission errors and improve the accuracy of burned area mapping.
Original Paper: Chandel, A., Sarwat, W., Najah, A., Dhanagare, S., & Agarwala, M. (2022). Evaluating methods to map burned area at 30-meter resolution in forests and agricultural areas of Central India. Frontiers in Forests and Global Change, 5, 933807.
NDVI: Normalized Difference Vegetation Index GEMI: Global Environmental Monitoring Index BAI: Burnt Area Index; BAIMS Burn Area Index Modified SWIR NBR: Normalized Burn Ratio MiRBI: Mid-Infrared Burn Index BAIML: Burn Area Index Modified-lSWIR NDWI: Normalized Difference Water Index