Comparative Evaluation of Fast-Learning Classification Algorithms for Urban Forest Tree Species Identification Using EO-1 Hyperion Hyperspectral Imagery

Balabathina Veera Narayana^1*Surender Mishra¹Suhas Sharma¹Piyush Kumar¹Akash Narayan²

• ¹Panacea Geospatial, New Delhi, New Delhi, India
• ²TERI School of Advanced Studies, New Delhi, India

Abstract— Accurate identification of forest tree species is vital for sustainable forest management, biodiversity assessment, and environmental monitoring. This study presents a comparative evaluation of thirteen supervised classification algorithms for species-level tree classification in the heterogeneous Hauz Khas Urban Forest, New Delhi, India, using EO-1 Hyperion hyperspectral imagery. The evaluated classifiers include traditional spectral/statistical methods—Maximum Likelihood, Mahalanobis Distance, Minimum Distance, Parallelepiped, Spectral Angle Mapper (SAM), Spectral Information Divergence (SID), and Binary Encoding—as well as machine learning algorithms such as Decision Tree (DT), K-Nearest Neighbor (KNN), Support Vector Machine (SVM), Random Forest (RF), and Artificial Neural Networks (ANN). To address spectral redundancy and high dimensionality, dimensionality-reduction techniques (Principal Component Analysis—PCA and Minimum Noise Fraction—MNF) along with band-selection strategies based on the Average Pairwise Absolute Difference (APAD) metric and Species-specific band-ratio indices were employed to enhance classification performance. Comprehensive ground-truth samples collected from field surveys and cross-validated with very high-resolution Pléiades imagery ensured reliable training and validation of the classification models. A total of 21 tree species were identified, with Random Forest and Decision Tree emerging as the most effective among all tested classifiers. Random Forest achieved the highest species-level accuracy (95% for Peepal, Medlar) and demonstrated superior overall performance when applied to PCA-transformed data (Overall Accuracy = 82.56%, Kappa = 0.81). The findings demonstrate that integrating dimensionality reduction and optimal band selection with ensemble learning significantly enhances classifier efficiency and accuracy. The core contribution of this study lies in identifying the most effective fast-learning classifiers for accurate species-level mapping in spectrally heterogeneous urban forests, underscoring the potential of hyperspectral imaging and ensemble learning for scalable and operational urban forest monitoring.