Disentangling Verticillium Wilt and Nitrogen Deficiency in Cotton Through Hyperspectral Temporal Signatures and Cross-Stage Generalization

Cotton Verticillium wilt and nitrogen deficiency can produce similar chlorotic symptoms and spectral responses, while variation among growth stages may further limit model transferability. We used leaf-level hyperspectral and SPAD data from FS, BS, and BO to classify three operational classes—healthy-reference (H), nitrogen-deficiency (N), and Verticillium-wilt (W)—and assess preprocessing, feature selection, classification, and cross-stage transfer. N and W were defined by reduced nitrogen supply and characteristic field symptoms, respectively, without independent tissue-nitrogen or pathogen confirmation. Random forest–recursive feature elimination (RF-RFE) and a mutual-information-based particle swarm optimization filter (Filter-PSO) each selected 20 spectral bands, and support vector machine (SVM), Light Gradient Boosting Machine (LightGBM), and Extreme Gradient Boosting (XGBoost) were integrated into an Opt-SLX probability-fusion ensemble. Cross-stage analyses compared direct transfer with correlation alignment (CORAL) and mean–variance alignment. Repeated cross-validation on the development set showed that SG-MSC increased mean overall accuracy (OA) relative to raw spectra by 28.58, 27.78, and 35.71 percentage points at FS, BS, and BO, respectively. Under the fixed 20-band setting, classifier and ensemble performance depended on both growth stage and feature-selection pipeline. RF-RFE–Opt-SLX achieved held-out test OA values of 0.9206, 0.9286, and 0.9127 at FS, BS, and BO, respectively, whereas RF-RFE–SVM achieved the highest OA at BS (0.9365). Direct cross-stage transfer declined markedly, with mean OA ranging from 0.4206 to 0.5556 under RF-RFE and from 0.3968 to 0.5159 under Filter-PSO; transfer performance showed no simple monotonic relationship with the calendar interval between stages. Domain adaptation increased performance in some transfer tasks, but the magnitude and direction of the gains depended on the specific transfer task and feature representation. These results demonstrate pronounced growth-stage dependence in spectral discrimination among the H, N, and W operational classes and show that high within-stage classification accuracy does not necessarily translate into robust cross-stage transferability. Independent validation incorporating tissue nitrogen measurements, pathogen confirmation, and multi-environment datasets is needed to establish model robustness and generalizability.