Nondestructive detection of carotenoids in chrysanthemum and identification of drought using hyperspectral and machine learning

Carotenoids, a class of naturally occurring pigments ubiquitously distributed across plant and animal kingdoms, serve vital biological functions in photoprotection and antioxidant defense mechanisms. Within plant systems, these compounds critically modulate photosynthetic efficiency under drought stress conditions. While carotenoid content fluctuations offer a reliable biomarker for evaluating plant drought responses, conventional chemical quantification methods remain labor-intensive and unsuitable for large-scale agricultural monitoring. This investigation employs hyperspectral imaging technology (operating in the 859–1705 nm spectral range) to acquire spectral signatures from Chrysanthemum × morifolium (Ramat.) Hemsl. samples, subsequently applying five spectral preprocessing techniques to optimize data quality. Three feature selection algorithms (Successive Projections Algorithm, Max-Relevance Min-Redundancy, and Competitive Adaptive Reweighted Sampling) were employed to identify diagnostically significant wavelengths, while three regression models (Random Forest Regression, Partial Least Squares Regression, and Gradient Boosting Machine) were constructed for carotenoid prediction. SHapley Additive exPlanations analysis was performed to assess the interpretability and relative contributions of different feature selection approaches. Experimental results demonstrate that the SPA-Random Forest Regression hybrid model achieved optimal prediction performance (R C  = 0.9545, RMSE C  = 0.0096, R P  = 0.9228, RSME P  = 0.0134). Notably, the Random Forest Regression algorithm attained 99.03% classification accuracy in drought stress discrimination. This methodology establishes the viability of integrating hyperspectral imaging with machine learning as a non-destructive analytical framework for simultaneous carotenoid quantification and plant drought stress assessment.