Early Detection of Pine Wilt Disease at the Pre-Visual Stage Using UAV-Based Hyperspectral Imagery

Pine wilt disease (PWD) causes forest decline, but pre-visual spectral responses are weak and spatially heterogeneous within crowns. We developed a UAV hyperspectral framework combining targeted spectral-index construction with patch-based probability aggregation to enhance spectral discrimination and preserve localised spectral responses. We retrospectively labelled 477 Masson pine crowns from two survey plots using multi-temporal observations. Pre-visual candidates remained green at image acquisition and the first follow-up but subsequently developed visible discolouration. Sensitive bands were screened with emphasis on the healthy–pre-visual and pre-visual–early boundaries, and narrow-band indices were constructed using the training partition of Dataset A. Two three-band indices, VI2 (609, 699, and 737 nm) and VI3 (699, 737, and 765 nm), were retained. A random forest trained on crown-mean index features generated class-probability responses for non-overlapping 3 × 3 patches, which were aggregated into crown-level descriptors and classified using shrinkage linear discriminant analysis. On the held-out Dataset A validation set, the framework achieved 90.80% overall accuracy, 86.81% balanced accuracy, 88.45% macro-F1, and 77.78% pre-visual F1. Pre-visual F1 exceeded the NDVI baseline by 16.67 percentage points and improved by 12.78 and 7.19 points over crown-mean classification and pixel-level aggregation, respectively. For transfer from Dataset A to Dataset B, adaptation with three labelled crowns per class improved mean OA from 78.69% to 81.98% over the zero-shot baseline across 100 trials. These results indicate that targeted spectral indices and patch-based spatial representation can capture weak and localised spectral responses for crown-level screening of pre-visual PWD candidates under the evaluated conditions.