Construction and application of a drought classification model for tea plantations based on multi-source remote sensing

• Multisource data fusion outperforms single-source data. Among the combinations of multisource data, the combination of MS and TIR input features yields the best results, with an accuracy of 0.983. The ranking of accuracy for single-source data models is: MS > TIR > RGB. The accuracies of the five ML models range from 0.854 to 0.950. • The RSDCM model (combining the GA-BP algorithm with MS + TIR data) surpasses the other four models. The ranking of accuracy for the models is: RSDCM > BP > SVM > RF > ELM. The RSDCM model achieves the best performance (accuracy = 0.983). • Improvements in image acquisition and processing methods can enhance the classification of drought severity in tea plantations. In the backdrop of global climate change, drought is identified as a major natural hazard, posing a severe threat to tea production. Traditional methods for assessing drought stress in tea plants rely on manual investigation. However, this approach is time-consuming and labor-intensive. While unmanned aerial vehicle (UAV) remote sensing offers efficient alternatives, existing studies predominantly rely on single-sensor data (e.g., multispectral (MS) or thermal infrared (TIR)), overlooking the potential of multi-source fusion—especially for tea plantations. To address this gap, we propose RSDCM (Remote Sensing-based Drought Classification Model), an improved Genetic Algorithm-Backpropagation (GA-BP) combined with MS + TIR framework that optimizes initial weights and thresholds via GA's global search (hidden layer=1, neurons in the hidden layers=5, 50 generations, population size=5, NonUnifMutation operators) to escape local minima and accelerate convergence. A UAV platform equipped with MS, RGB, and TIR sensors collected multi-source data from drought-stressed tea plantations in Eastern China. The RSDCM model was benchmarked against single BP and three classical machine learning models (SVM, RF, ELM). The study found that: (1) Multi-source data fusion outperformed single-source data, with MS + TIR achieving optimal performance (Accuracy: 0.983, Precision: 0.967-1.000, Recall: 0.967-1.000, F1-score: 0.967-1.000)—surpassing MS (Accuracy: 0.950, Precision: 0.894-1.000, Recall: 0.917-0.983, F1-score: 0.924-0.983), TIR (Accuracy: 0.925, Precision: 0.862-0.982, Recall: 0.867-0.983, F1-score: 0.889-0.967), and RGB (Accuracy: 0.904, Precision: 0.824-0.950, Recall: 0.783-0.950, F1-score: 0.847-0.950) alone. (2) The RSDCM model (accuracy: 0.983) performed better than the other four models, with high generalizability across all drought levels (F1-scores: 0.967–1.000 for severe/moderate/light/normal classes). (3) The RSDCM model could accurately classify drought stress levels in tea plantations. Thus, RSDCM provides a novel, robust solution for UAV-based drought assessment in tea plantations, combining multi-sensor fusion and deep learning.