High-resolution Remote Sensing to Identify Tree Plantations from Natural Forests and Agriculture in Southern India

Abstract

Tree plantations play an important role in tropical and subtropical countries, including providing economic benefits as well as ecosystem services. Over the past decade, the amount of land under agroforestry and plantations has increased rapidly in smallholder systems. One way to identify the extent of agroforestry and plantations is to use remote sensing, which can map land cover at large spatiotemporal scales. However, remote sensing classifiers often confuse agroforestry and plantations with forest cover; this is because these land cover classes often have similar spectral signatures, particularly high Near-infrared (NIR) reflectance and Normalized Difference Vegetation Index (NDVI) values. In addition, smallholder plantations in tropical areas are especially difficult to identify due to the small size of the plantation plots and the high cloud cover during the monsoon season. However, with the launch of Sentinel-2, which has additional spectral bands in the red edge, it may be possible to better classify these land cover types. Our study objective was to develop a general classification model using high spatial- and temporal-resolution Sentinel-1 and Sentinel-2 imagery to identify smallholder plantations using random forest algorithms. We developed this algorithm in southern India, in the four states with the largest plantation areas - Kerala, Karnataka, Tamil Nadu, and Andhra Pradesh. We find that using only Sentinel-1 imagery has lower classification accuracy (~70%) than using only Sentinel-2 imagery (~90%). Additionally, the combination of Sentinel-1 and Sentinel-2 can be used to map smallholder plantations with high accuracy: 93.91% (Kerala), 93.40% (Karnataka), 88.38% (Tamil Nadu) and 92.91% (Andhra Pradesh). Our results demonstrate the feasibility of systematically identifying tree plantations using high-resolution remote sensing data and machine learning algorithms