More appropriate DenseNetBL classifier for small sample tree species classification using UAV-based RGB imagery

Ni Wang; Tao Pu; Yali Zhang; Yuchan Liu; Zeyu Zhang

doi:10.1016/j.heliyon.2023.e20467

More appropriate DenseNetBL classifier for small sample tree species classification using UAV-based RGB imagery

Heliyon. 2023 Sep 26;9(10):e20467. doi: 10.1016/j.heliyon.2023.e20467. eCollection 2023 Oct.

Authors

Ni Wang¹, Tao Pu^{1

2}, Yali Zhang¹, Yuchan Liu^{1

3}, Zeyu Zhang¹

Affiliations

¹ School of Geographic Information and Tourism, Chuzhou University, Chuzhou, 23900, China.
² School of Geomatics, Anhui University of Science and Technology, Huainan, 232001, China.
³ School of Civil Engineering, Nanjing Forestry University, Nanjing, 210037, China.

Abstract

To effectively classify tree species within datasets characterized by limited samples, we introduced a novel approach named DenseNetBL, founded upon the fusion of the DenseNet architecture and a pivotal bottleneck layer. This bottleneck layer, encompassing a compact convolutional component, played a central role in our methodology. The evaluation of DenseNetBL was conducted under varying conditions, encompassing small-sample tree species data, extensive remote sensing datasets, and state-of-the-art classifiers. Furthermore, a quantitative assessment was executed to extract tree species areas. This was achieved by quantifying pixel areas within manually delineated tree species maps and classifier-generated counterparts. The findings of our study indicated that, in scenarios devoid of pre-trained weights, DenseNetBL consistently outperformed its DenseNet counterpart with equivalent layer numbers. In the realm of small-sample situations, both the Swin Transformer and Vision Transformer exhibited inferior performance when juxtaposed with DenseNet and DenseNetBL. Remarkably, among the shallow architectures, DenseNet33BL showcased superior aptitude for small-sample tree species classification, culminating in the most commendable results (Overall Accuracy (OA) = 0.901, Kappa = 0.892). Conversely, the Vision Transformer yielded the least favorable classification outcomes (OA = 0.767, Kappa = 0.708). The amalgamation of DenseNet33BL and simple linear iterative clustering emerged as the optimal strategy for attaining robust tree species area extraction results across two prototypical forests. In contrast, DenseNet121 exhibited suboptimal performance in the same forests, attaining the least satisfactory tree species area extraction results. These comprehensive findings underscore the efficacy of our DenseNetBL approach in addressing the challenges associated with small-sample tree species classification and accurate tree species area extraction.

Keywords: DenseNetBL; Forests tree species classification; SLIC; Small sample classification.