Q-TriLSTM-Vision: a quantum-interference- augmented tri-stream LSTM for multi-label plant stress recognition on the OLID-I benchmark.

Abstract

Plant stress recognition plays a vital role in precision agriculture by enabling the early detection of diseases, insect infestations, and nutrient deficiencies that adversely affect crop productivity. Although deep learning models have achieved promising results, existing CNN-, Transformer-, and hybrid architectures often struggle to capture complex spatial dependencies, distinguish visually similar stress symptoms, and handle class imbalance in multi-label classification. To address these challenges, this paper proposes Q-TriLSTM-Vision, a novel hybrid deep learning framework that integrates an EfficientNet-B0 visual encoder, a tri-stream long short-term memory (TriLSTM) network, and a lightweight quantum-inspired interference gate. Unlike quantum computing-based approaches, the proposed interference mechanism is implemented entirely using classical neural operations to enhance feature representation without requiring quantum hardware. The model further employs an entanglement-inspired attention fusion module, focal binary cross-entropy loss with label smoothing, weighted sampling, and per-class threshold calibration to improve discriminative learning and minority-class recognition. The proposed framework was evaluated on the OLID-I dataset and further validated on the PlantVillage and PlantDoc benchmark datasets. Experimental results demonstrate that Q-TriLSTM-Vision achieved Macro-F1 scores of 0.9127, 0.9624, and 0.8975 on OLID-I, PlantVillage, and PlantDoc, respectively, outperforming representative CNN-, Transformer-, and hybrid deep learning models while achieving lower Hamming loss and improved recall. Cross-validation, ablation studies, and statistical significance analysis further confirm the robustness and effectiveness of the proposed framework. Overall, Q-TriLSTM-Vision provides an accurate, computationally efficient, and reliable solution for intelligent plant stress recognition in precision agriculture.

Code and data availability

The paper's plant-phenotyping measurements are based on the public OLID-I plant stress image dataset hosted on Kaggle, which the authors explicitly identify as the data source for training and evaluating Q-TriLSTM-Vision. No author analysis code, trained models, or supplementary assets with explicit public availability