Plant Methods Биотехнология

HSIFreezeNet: a lightweight random forest assisted framework for hyperspectral segmentation of camelina freezing damage

GC, Sunil, Ali, Aizaz, Jalal Ud Din, Mian, Rahman, Mukhlesur, Anderson, James V., Kandel, Jinita Sthapit, Horvath, David, Sun, Xin
Springer Nature · 2026-10-19 · DOI 10.1186/s13007-026-01602-w

Аннотация

Highlights Random forest-generated pseudo-labels reduced pixel-wise annotation effort for HSI freeze damage segmentation. HSIFreezeNet-V1 achieved 75% mIoU at 71%% lower inference time and 759-fold fewer parameters than UPerEfficientNet-B7 (79% mIoU). Independent expert annotation revealed a performance gap beyond pseudo-label evaluation. Lightweight architecture enables full hyperspectral cube inference in 2.78 s on embedded hardware. Freezing damage quantification is essential for understanding phenotypic-genotypic relationships, yet automated, scalable framework exists for pixel-wise freeze damage segmentation from hyperspectral imagery remain limited. Current assessments rely on visual inspection, which is subjective and time-consuming. Hyperspectral data, while rich in information, are high-dimensional and large in size, making pixel-wise manual annotation labor-intensive and slow. Deep learning models further increase computational demands when trained on large datasets. To overcome these challenges, we proposed a pseudo-labeling segmentation pipeline in which a random forest model trained on sparse region-of-interest annotation generates full-image masks for training deep learning architectures, substantially reducing labeling effort relative to dense pixel-wise annotation. Manual annotation covered 57 of 442 plants, from which dense masks were generated for the full dataset. We benchmarked thirteen state-of-the-art models and introduced a lightweight architecture, HSIFreezeNet-V1. On a plant-grouped held-out test partition, UPerEfficientNet-B7 achieved the highest mean IoU (0.785 ± 0.005), while HSIFreezeNet-V1 achieved 0.752 ± 0.018 using 759-fold fewer parameters and requiring 71% less inference time on an NVIDIA Jetson AGX Orin (2.78 s versus 9.57 s per hyperspectral cube). To assess accuracy beyond agreement with machine-generated labels, we additionally evaluated all architectures against an independent expert-annotated test set, providing a direct measure of pseudo-label quality. Preliminary qualitative transfer to strawberry and barley was also observed. These findings show that lightweight models trained on pseudo-labeled hyperspectral data can enable rapid quantitative damage assessment on embedded hardware, while independent expert evaluation provides a realistic measure of practical segmentation performance.

Читать у издателя →