A deep-learning-based tree species classification for natural secondary forests using unmanned aerial vehicle hyperspectral images and LiDAR

Accurate tree species classification is essential for forest resource management and biodiversity assessment. However, classifying tree species becomes challenging in natural secondary forests due to the difficulties in outlining the tree crown boundary. In this study, an object-based framework for tree species classification in the Experimental Forestry Farm of Northeast Forestry University, located in Heilongjiang Province, China, was developed based on unmanned aerial vehicle (UAV) hyperspectral images (HSIs) and UAV light detection and ranging (LiDAR) data using convolutional neural networks (CNNs). The study area was characterized by representative natural secondary forests that encompass diverse tree species, such as Korean pine (Pinus koraiensis Sieb. et Zucc.), White birch (Betula platyphylla Suk.), Siberian elm (Ulmus pumila L.), and Manchurian ash (Fraxinus mandshurica Rupr.). This study included two key processes: (1) the u-shaped network (U-net) algorithm was employed with the simple linear iterative clustering (SLIC) algorithm, that is, the U-SLIC algorithm, for individual tree crown delineation (ITCD), and (2) the performances of one-dimensional CNN (1D-CNN), two-dimensional CNN (2D-CNN), and three-dimensional CNN (3D-CNN) models for tree species classification were compared while investigating the role of an attention mechanism (convolutional block attention module, CBAM) added to CNN models (1D-/2D-/3D-CNN + CBAM). The results showed that the U-SLIC algorithm obtained a satisfactory accuracy for the ITCD procedure, with a recall of 0.92, precision of 0.79, and F-score of 0.85. The feature selection effectively enhanced the CNN models' performances for tree species classification. Furthermore, adding the CBAM resulted in overall accuracy (OA) improvements of 0.08, 0.11, and 0.09 for the 1D-CNN, 2D-CNN, and 3D-CNN, respectively. The 1D-CNN + CBAM model performed best with an OA of 0.83 when utilizing the selected HSI and LiDAR features. This framework highlighted the utilization and integration of multiple deep-learning algorithms in complex natural forests, serving as prerequisites for forest management decisions, biodiversity conservation, and carbon stock estimation.