Role of phenology as discriminant in vegetation type mapping

Abstract

Vegetation-type maps are crucial for landscape conservation and management, and provide essential data for decision-making. Due to cost and time constraints, traditional field-based methods have limitations in terms of scale and frequency in updates. The increasing availability of satellite sensors that offer high-temporal-frequency multispectral imagery has spurred the development of more efficient and accurate vegetation mapping techniques. Here, the potential of Sentinel-2 time series consisting of multiseason images was assessed to enhance the accuracy of the vegetation-type map, compared to a single image of the leaf-off season, using the Random Forest algorithm in the Similipal Biosphere Reserve, Odisha. Time-series imagery is pivotal for assessing vegetation phenology, offering comprehensive means to track temporal vegetation dynamics and discern seasonal patterns. The Biosphere Reserve is predominantly covered by tropical sal-mixed moist deciduous forest, followed by tropical sal-mixed dry deciduous and semievergreen forests. The classified output using Sentinel-2 time-series attained an overall accuracy of 97.1% with a kappa value of 0.96, whereas the classification based on a single image achieved 86.6% accuracy and a kappa value of 0.84. The overall accuracy of the output map suggests that leveraging phenological patterns extracted from multitemporal images, compared to a single image from the dry season, could enhance model performance by 9.7%.