A Hybrid Learning Framework for Bamboo Mapping Using Supervised and Unsupervised Classification

Abstract

Accurate mapping of bamboo resources is important for ecological assessment, forest management and regional resource planning, particularly in landscapes where bamboo occurs with other vegetation types. This study presents a hybrid learning framework for bamboo mapping using Sentinel-2 surface reflectance imagery acquired from January to December 2024 for Karbi Anglong and Dima Hasao districts of Assam, India. The proposed approach integrates unsupervised K-means clustering with supervised Random Forest classification to improve the discrimination of bamboo-related land-cover classes. Spectral bands B2, B3, B4 and B8, together with the Normalised Difference Vegetation Index, were used as input features. K-means clustering was first applied to identify spectrally homogeneous regions and to support the refinement of bamboo training samples. The refined training dataset was then used to classify six land-cover classes: water, land, forest, mixed vegetation, pure bamboo and bamboo-dominated areas. A total of 1,138 training samples was used, and the dataset was divided using a 70:30 train-test split. The proposed Hybrid Random Forest model achieved an overall accuracy of 95.3%, a kappa coefficient of 0.94 and a bamboo F1-score of 0.95. The results indicated improved classification performance compared with the standalone Random Forest and Support Vector Machine models evaluated in the study. Area-wise assessment showed substantial bamboo presence in both study districts, with pure bamboo and bamboo-dominated classes forming major components of the classified landscape. The findings suggest that cluster-assisted training sample refinement can improve bamboo classification in heterogeneous forest landscapes using medium-resolution multispectral satellite data, provided that the training samples and outputs are carefully validated.