An environment-guided visual-temporal deep learning framework for early disease detection in greenhouse horticultural crops.

Abstract

Introduction In protected horticultural production, early disease identification and precise intervention are critical for safeguarding crop yield and quality while reducing chemical pesticide inputs. However, early-stage greenhouse diseases often exhibit extremely subtle visual symptoms, and their occurrence and progression are highly dependent on environmental condition variations, making stable and reliable early warning difficult to achieve using conventional methods based on single visual information or simple multimodal fusion. Methods To address this challenge, a visual-environment joint early disease perception framework for greenhouse horticultural crops is proposed. Through an environment-guided visual attention mechanism and a spatial-temporal joint modeling strategy, environmental variables such as temperature, humidity, vapor pressure deficit, and CO 2 concentration are transformed from passive features into active priors, thereby guiding visual feature learning and enhancing sensitivity to weak disease signals. The proposed method is systematically validated on a real-world greenhouse multimodal temporal dataset. Results Experimental results demonstrate that the proposed approach achieves an accuracy of 91.3%, a recall of 88.9%, and an F1-score of 89.8% in overall disease detection tasks, significantly outperforming multiple baseline models based on convolutional neural networks (CNNs), Transformers, and existing multimodal fusion strategies. In early-stage disease detection scenarios, early precision and early recall reach 88.5% and 86.1%, respectively, with the lead time extended to 2.7 days, indicating a clear advantage in early warning capability. Ablation studies further verify the critical roles of environment-guided attention, spatial-temporal joint modeling, and the joint loss function in improving early detection performance and stability. Discussion This study provides a practically valuable technical pathway for early intelligent warning and precise regulation of greenhouse crop diseases. By integrating environmental dynamics with visual perception, the proposed framework improves the sensitivity and robustness of early disease detection in complex greenhouse conditions, showing strong potential for practical deployment in protected horticulture.

Plant phenotyping relevance

環境情報と画像・時系列情報を統合して、植物病徴を早期検出する深層学習フレームワークの開発と実データでの系統的検証が中心であり、植物の病害状態を直接推定するため。

Code and data availability

The paper describes a greenhouse multimodal disease dataset and deep learning framework, but no public dataset, code, model, or supplement URL is provided. The data availability statement only offers raw data from the authors upon request.

No evidence-backed public reproduction asset is currently recorded.