Comparison of Machine Learning Algorithms for Broad Leaf Species Classification Using UAV-RGB Images

Knowing the tree species combination of forests provides valuable information for studying the forest’s economic value, fire risk assessment, biodiversity monitoring, and wildlife habitat improvement. Fieldwork is often time-consuming and labor-required, free satellite data are available in coarse resolution and the use of manned aircraft is relatively costly. Recently, unmanned aerial vehicles (UAV) have been attended to be an easy-to-use, cost-effective tool for the classification of trees. In fact, given the cost-efficient nature of UAV derived SfM, coupled with its ease of application, it became a popular choice. The type of imagery is an important factor in classification analysis because the spatial and spectral resolution can influence the accuracy of classification. On the other hand, classification algorithms also play an important role in the accuracy of tree species identification. So, this study investigated the performance of four classifiers for tree species classification using UAV-based high-resolution imagery in broadleaf forests and takes a comparative approach to examine the three non-parametric classifiers including support vector machines (SVM), random forest (RF), artificial neural network (ANN), and one parametric classifier including linear discriminant analysis (LDA) classifiers in heterogeneous forests of Noor city located in Mazandaran province. In June 2019, the study area was photographed. The field survey was carried out to record the species and position of the mature overstory trees which were clearly identifiable on the orthomosaics. Individual tree crowns were clipped by one-meter buffer and the digit numbers were summarized at for each tree by computing descriptive statistics from the orthomosaics. Using zonal statistics, mean, standard deviation, variance, unique, range, mode, and median were calculated for raw bands (Red, Green, Blue), vegetation indices (NRB, NGB), and band ratios (G/R, R/B) from RGB orthomosaics. We classified the tree into 4 classes: Parrotia persica (Ironwood tree), Populus capsica (Caspian poplar), Ulmus minor (Common Elm), and Quercus castaneifolia (Chestnut-leaved oak). Finally, the classification algorithms were applied using R software. The classification accuracy for identified trees was performed using 10-fold cross-validation by computing the producer’s accuracy, user’s accuracy, and Overall accuracy. All algorithms resulted in overall accuracies above 80%. Of course, the results showed that, as a parametric algorithm, LDA with an overall accuracy of 0.87 provided the best results for tree classification, because it does not require the tuning of free parameters. As for parameter value, the mean was the most important that this can be related to the similarity of this feature in any sample. Caspian poplar with user accuracy of 0.97 and Ironwood tree with user accuracy of 0.72 had the highest and lowest classification accuracy, respectively. Caspian poplar high accuracy is probably due to its crown color which is quite different from the other species. The main error (misclassification) is a classification between “Ironwood tree” and “Common Elm” classes. This may be caused by the fact that the spectral signatures between Ironwood tree and Common Elm trees are very similar. In general, our study showed that UAV derived orthomosaic can be used for tree classification with very high accuracy in mix broadleaf forests by different algorithms.