Table 1 The confusion matrix of SVM classification
results.
The three classification methods demonstrated variability in their
results across different tree species (Figure 5). Two common issues were
identified (Figure 6): commission and omission, and patch fragmentation.
Commission and omission were mainly observed in Artemisia, Salix, and
Corn communities. UAV images with high spectral resolution facilitated
the identification of tree crown shadows, affecting spectral
heterogeneity and leading to misclassification of shadows as Artemisia.
Meanwhile, the spectral similarity between Artemisia and Salix resulted
in increased misclassification. Within the same vegetation environment,
variations in age, growth conditions, sizes, and shapes of plant species
contributed to significant within-species variability. Additionally,
image errors such as irregular changes in intensity density caused by
stitching might cause commission and omission, which are unavoidable.
The fragmentation of patches was caused by two factors: high spatial
resolution, enabling clear identification of gaps between tree and shrub
branches, and sparsely distributed grass, resulting in many small
fragmented patches that hindered plant spatial structure. Large spectral
variation within the same type introduces uncertainty during computer
processing, leading to ‘speckle’ noise. To address these,
post-classification analyses (clump analysis and majority/minority
analysis) were conducted to refine the results and serve as the basis
for calculating plant diversity.