Developing a Semi-automated Random Forest Classification Scheme to Analyze Burned Area Using Landsat Imagery in Southern India

Abstract

Fire regimes, or the pattern, frequency, and intensity of fires over a landscape, has begun to shift over much of the dry tropical regions of the world. The complexity of the pressure fire exerts on the landscape is intensified by the intimate relationship humans have cultivated with this process, especially in regions where traditional burning has taken place. However, the departure from the pre-historic and more recent historical use of fire as a management tool, generally due to conservation regulations, presents unique research challenges to understand how major shifts in human relationships to fire, management regulations of fire prone ecosystems, and vegetation composition, may be driving changing fire behavior and regime. The Biligiriranganatha Hills Temple Reserve (BRT) and Sathyamangalam Forest Reserve (SFR) are adjacent protected areas southern India have similar ecologies, rainfall seasonality, yet have different histories of regulation, species invasions, and human resource use, within the last two decades that may result in key differences in fire regimes. This study therefore sought to construct a fire history across this region to quantify potential differences in burning patterns, compare these patterns to those available with MODIS burned area product (MCD64A1.006), and to explore the potential of this detection method to evaluate and characterize frequency and the proportion of area burned across landcover types (forest, savanna, and other). I used Landsat 7 remotely sensed data and the Random Forest algorithm in Google Earth Engine to develop a semi-automatic method to detect burned area across a time series. The RF classifier was trained using a subset images in which burned pixels were manually classified and then used identify burned area across all available images from 2003-2019 (N= 115 images). Separate classifiers were developed over ~ 5-year time steps due to computational limitations but demonstrated high overall accuracy at 99% for unclassified pixels within the training images and 98% for images not included in the training data. Classified images were composited to develop annual fire maps that were accurate to the Landsat resolution (30M). the Landsat product performed far better at detecting small fires not flagged by MODIS and, in cases of large fires, created fire perimeters similar to MODI. The majority of burned pixels in both reserves experienced burns approximately every 3-4 years. The proportion of burned area by land cover type was highest for savanna in both reserves and the proportion of burned area in forest was higher in BRT than in SFR, potentially indicating landscape-level shifts in ecosystem flammability. Overall, this study developed an efficient semi-automatic classification system for Landsat 7 data that created 30m resolution annual fire summaries which can be used to better inform on the ground research by Indian conservation organizations.