Tree species classification of pine and miombo using UAV-LiDAR data and machine learning algorithms

Tree species information is essential for forest inventory, biodiversity monitoring, carbon stock assessment and sustainable resource management. However, traditional forest inventories remain the sole tree species classification method applied in Malawi to date. To overcome the limitations of these localised, labour-intensive approaches, remote sensing techniques are increasingly adopted globally, offering a synoptic view of the landscape and covering wider geographical areas. This study aimed to identify significant LiDAR tree metrics and assess classification accuracy of random forest (RF) and support vector machine (SVM) models in pine and miombo species. A Euclidean distance-based method was applied for the co-registration of randomised tree identifications. The models were evaluated using overall accuracy (OA), Kappa accuracy (KA) and K-fold cross-validation strategies. The RF and SVM models performed well for the pine species; RF achieved a higher OA (0.78) and KA (0.66) than SVM (OA = 0.75, KA = 0.62), although the difference was not statistically significant (p = 0.2356). Conversely, model performance dropped sharply among the miombo species; RF achieved an OA of 0.34 and a KA of 0.05, while SVM achieved a marginally better OA of 0.41 but a comparable KA of 0.04; however, no statistically significant differences were observed (p = 0.4049). The study demonstrates that tree crown area, height and height percentiles are the primary structural metrics required for accurate classification, with Pinus patula and Uapaca kirkiana emerging as the most reliably identified species in their respective forest ecosystems. In essence, tree species classification accuracy varies significantly across forest types.