Hyperspectral Imaging with cost-sensitive learning for high-throughput screening of loblolly pine (Pinus Taeda L.) seedling for freeze tolerance


Loblolly pine (Pinus taeda L.) is a commercially important timber species planted across a wide temperature gradient in the southeastern U.S. It is critical to ensure that the planting stock is suitably adapted to the growing environment to achieve high productivity and survival. Long-term field studies, although considered the most reliable method for assessing cold hardiness of loblolly pine, are extremely resource-intensive and time-consuming. The development of a highthroughput screening tool to characterize and classify freeze tolerance among different genetic entries of seedlings will facilitate accurate deployment of highly productive and well-adapted families across the landscape. This study presents a novel approach using hyperspectral imaging to screen loblolly pine seedlings for freeze tolerance. A diverse population of 1549 seedlings raised in a nursery were subjected to an artificial mid-winter freeze using a freeze chamber. A customassembled hyperspectral imaging system was used for in-situ scanning of the seedlings before and periodically after the freeze event, followed by visual scoring of the frozen seedlings. A hyperspectral data processing pipeline was developed to segment individual seedlings and extract the spectral data. Examination of the spectral features of the seedlings revealed reductions in chlorophylls and water concentrations in the freeze-susceptible plants. Because the majority of seedlings were freeze-stressed, leading to severe class imbalance in the hyperspectral data, a cost-sensitive learning technique that aims to optimize a class-specific cost matrix in classification schemes was proposed for modeling the imbalanced hyperspectral data, classifying the seedlings into healthy and freeze-stressed phenotypes. Cost optimization was effective for boosting the classification accuracy compared to regular modeling that assigns equal costs to individual classes. Full-spectrum, costoptimized support vector machine (SVM) models achieved geometric classification accuracies of 75% to 78% before and within 10 days after the freeze event, and up to 96% for seedlings 41 days after the freeze event. The top portions of seedlings were more indicative of freeze events than the middle and bottom portions, leading to better classification accuracies. Further, variable selection enabled significant reductions in wavelengths while achieving even better accuracies of up to 97% than full-spectrum SVM modeling. This study demonstrates that hyperspectral imaging can provide tree breeders with a valuable tool for improved efficiency and objectivity in the characterization and screening of freeze tolerance for loblolly pine.

Transactions of the ASABE 64(6): 2045-2059