Evaluating drought stress response of poplar seedlings using a proximal sensing platform via multi-parameter phenotyping and two-stage machine learning

Drought has become a major climate threat affecting the growth and yield of agricultural and forestry crops. Rapid evaluation of drought tolerance, response, and recovery plays an important role in the cultivation and management of forestry seedlings. In this study, the response of poplar ( Populus L.) seedlings under different drought stress levels was analyzed using two-stage machine learning. Two varieties of poplars differing in their drought tolerance were used for experiment. Three groups of phenotypic traits were measured. The first group was the morphological traits of plant height, ground diameter, crown width, and leaf number, collected via manual measurement. The second group was the physiological and biochemical traits of chlorophyll content, leaf water content, specific leaf weight, and equivalent water thickness, measured by destructive leaf sampling. The third group was the nondestructive spectral traits captured by a RedEdge-MX multispectral camera mounted on a custom-made phenotyping platform, including B (blue), G (green), R (red), NIR (near-infrared), RedEdge (red edge), RVI (ratio vegetation index), NDVI (normalized difference vegetation index), SIPI (structure insensitive pigment index), GI (green index), VDVI (visible-band difference vegetation index), GNDVI (green NDVI), and SRI (simple ratio index). Random forest (RF) was employed in a two-stage modeling scheme, with the first stage to classify poplar varieties, and the predicted variety information was added to the second stage for drought tolerance classification. The results showed that the accuracy of variety classification (Stage 1) reached 100%. Moreover, the classification of drought stress (Stage 2) was more accurate with the addition of the predicted variety information. The average accuracy, recall, and precision of the best drought classification model were 95.6%, 90.2%, and 92.1%, respectively. This study constructed a drought stress detection and grading system for poplar seedlings, which would be further applied for accurate and rapid evaluation of drought-tolerance and forestry irrigation management.