Tree species classification based on machine learning techniques: mapping Chir pine in Indian Western Himalayas

Abstract

Detailed information on the extent and composition of tree species in a forest is crucial for scientific studies and forest management plans. In the recent years, remote sensing has emerged as a powerful tool in gathering different vegetation biophysical parameters. Further, the advancements in the state-of-art machine learning techniques have enhanced the process of combining multi-temporal, multi-sensor datasets, handling and analyzing large variable datasets to produce results with higher precision and accuracies. Ground-based census and mapping of tree species is a cumbersome, time taking and expensive exercise. Thus, Machine Learning Algorithm (MLA) based classification of remotely-sensed datasets has become a topical aspect of research interest. There are only a handful of studies mapping the distribution of forest tree species in the Western Himalayas using satellite imagery. The present study is amongst the pioneer studies undertaken in India to map one of the dominant tree species i.e. Pinus roxburghii, commonly known as Chir pine, by integrating machine learning techniques with the remote sensing technology. A supervised Random Forest MLA has been employed to discriminate between tree species based upon various spectral and topographical variables from Sentinel imagery. The methodology employed produced reliable maps of distribution of Chir pine forests in the study area, Uttarakhand, India with adequate accuracy. As the study has been undertaken using open and freely-available datasets such as Sentinel, Google Earth Engine (GEE) platform and ML libraries, it has the potential for adoption in countries like India for forestry research and inventory in a cost-effective manner. The methodology can be reproduced for delineation of different forest tree species to produce distribution maps with enhanced accuracies.

Deep transfer learning for dominant tree species classification based on google earth engine Sentinel-2: a case study in Kunming City

“…The research process consists of three main steps. Firstly, we conducted preprocessing of Sentinel-2 images for the Kunming region and downloaded these images from the Google Earth Engine (GEE) platform [10][11][12]. Next, we selected pure forest stands from the Forest Resource Second-Class Survey data, covering over 65% dominance of target tree species, to serve as our sample data.…”