Characterization of tropical forests at community level: combining spectral, phenological, structural datasets using random forest algorithm

Abstract

Since the inception of satellite remote sensing as a technology, characterization of forests has been one of the major application of it. Characterization of forests at Species level is essential for sustainable management of biodiversity. Recent advancements in remote sensing as a technology has enable us to observe not only the re ectance spectra of forests from space, but also their phenology and structure. In this study Earth Observation (EO) datasets were divided into 3 parts namely spectral data, structural and phenological data. Random forest algorithm was applied on the 3 sets of EO data and eld inventorybased tree community classes to generate tree community maps of Purna wildlife sanctuary. Classi cation accuracy achieved from spectral datasets (79.08% to 87.23%) was more than phenological dataset (80.94%) which was more than structural datasets (74.11% to 81.49%). A model with combination of predictors from the 3 datasets increased the classi cation accuracy to 90.29%. Some salient ndings of this study are 1) in general with the current sensors the accuracies achieved for tree community mapping is Spectral datasets> Phenological datasets> Structural datasets 2) Signi cant increase in accuracy can be achieved by combining the three datasets 3) In case of spectral datasets, the last image before the start of monsoon season gave the best accuracy 4) In the case of spectral datasets, relatively modern spectral bands contributed signi cantly more to the model as compared to trivial bands.