Tree Species Classification in a Madagascar Tropical Dry Forest Using Multi-Temporal UAV Multispectral Imagery and Machine Learning

Abstract

Accurate tree species mapping in tropical dry forests remains challenging because strong seasonal phenological variation and spectral overlap can limit discrimination from single-date imagery. This study evaluated the integration of wet- and dry-season unoccupied aerial vehicle (UAV) multispectral imagery, vegetation indices, Gray-Level Co-occurrence Matrix (GLCM) texture metrics, and digital surface model (DSM)-derived information for individual-tree classification in Kirindy Mitea National Park, Madagascar. UAV imagery at 11.3 cm ground sampling distance was linked to manually delineated crowns of 231 field-identified trees representing six canopy species. After missing-data filtering, 230 trees were retained and stratified into training (n = 186) and independent (n = 44) test datasets. Predictor selection was conducted exclusively on the training set: 68 candidate crown-level predictors were screened using Kruskal–Wallis tests with Benjamini–Hochberg false-discovery-rate correction, followed by Spearman correlation filtering (ρ ≥0.80), yielding 34 predictors. Random Forest (RF) and radial Support Vector Machine (SVM) classifiers were tuned using the same repeated five-fold cross-validation framework (five repetitions), without synthetic oversampling. On the independent test set, RF achieved an overall accuracy of 65.9% (95% CI: 50.1–79.5%), Cohen’s kappa of 0.45, balanced accuracy of 0.65, Macro-F1 of 0.40, and MCC of 0.46. SVM achieved an overall accuracy of 59.1% (95% CI: 43.3–73.7%), kappa of 0.33, balanced accuracy of 0.64, Macro-F1 of 0.38, and MCC of 0.36. The paired difference in overall correctness was not significant (McNemar p = 0.606). RF permutation importance identified wet-season CTVI median, dry-season NDWI median, wet-season NIR contrast median, and wet-season KNDVI intra-crown variability among the highest-ranking predictors. These results show that multi-season spectral, textural, and DSM-derived information provides complementary information for tree-species discrimination, while class-specific performance and uncertainty indicate that broader validation and larger samples are required before operational transfer to other sites.