19 June 2026 in Remote Sensing Applications: Society and Environment
11 April 2026 in SSRN Electronic Journal
Accurate species inventorying is essential for mangrove monitoring and conservation but is often hindered by challenging terrain and limited accessibility. With recent related studies exploring Unmanned Aerial Vehicles (UAVs), hyperspectral sensors were adopted to support this task. While promising, these approaches often involve complex operational requirements and high costs for commercial UAVs and their associated equipment. This study explores an alternative approach, training machine learning models on leaf-based hyperspectral data to classify mangrove genus and species in UAV-acquired data. Leaf-based data were collected using a low-cost, handheld spectrometer for two mangrove genera (Rhizophora and Avicennia) and two Rhizophora species (R. apiculata and R. mucronata) in Malaysia. Baseline Multilayer Perceptron (MLP) classifiers trained on untransformed bands were compared against models trained on features derived from Principal Component Analysis (PCA), and Linear Discriminant Analysis (LDA). Models trained and validated with leaf-based datasets demonstrated high accuracies for genus-level classification (98.17% to 99.57%) but showed higher variability in species-level classification (82.86% to 90.78%). When tested on UAV-based data, models trained on leaf-based PCA and untransformed features showed significantly lower accuracy. Conversely, LDA effectively aligned leaf and UAV datasets into a compatible feature space for accurate cross-domain classification, achieving accuracies of 98.63% and 91.96% at genus and species levels, respectively. These results demonstrate the effectiveness of combining leaf-trained models with LDA feature transformation for accurate cross-domain classification of UAV-based hyperspectral data, potentially reducing operational risks and high costs of extensive UAV data collection.