Machine learning classification of mixed savannah environments.

The inherent spatial heterogeneity of savannah landscapes limits accurate discrimination of different vegetation components at the spatial scale of medium resolution sensors. We address this issue using very high resolution (VHR) WorldView-3 (WV-3) imagery. Previous studies in similar contexts provide limited information on the utility of specific bands and derived image features from VHR data for optimized savannah vegetation mapping. Against this backdrop, we (i) compared RF, SVM, CART and MLP models, and (ii) investigated the relative contributions of original WV-3 bands, and vegetation and texture indices for discriminating four plant functional types (PFTs) - woody evergreen, woody deciduous, bunch grasses, grazing lawns - in Kruger National Park, South Africa. All models achieved high F-scores (F1) and overall accuracy (OA) scores, with RF (F1 = 92.50 ± 0.004%, OA = 91.16 ± 0.004%), SVM (F1 = 94.75 ± 0.002%, OA = 93.02 ± 0.002%) and MLP (F1 = 95.25 ± 0.003%, OA = 94.27 ± 0.003%) outperforming CART (F1 = 86.75 ± 0.006%, OA = 84.93 ± 0.003%). A combination of original image bands, spectral and texture indices enhanced discrimination capacities of the machine learning models. Across all models, the most important features were: (i) Coastal blue and Yellow for original bands; (ii) Global Environmental Monitoring Index and Modified Soil Adjusted Vegetation Index-2 for vegetation indices, and (iii) Mean and Sum average for texture
features. Our research contributes a robust workflow for optimized use of WV-3 imagery and algorithms for accurate high-resolution monitoring of savannah vegetation structure and phenology, while providing timely information
critical for the understanding of habitat heterogeneity in southern Africa savannahs.