Work

Estimation of coffee leaf physiological indicators based on VAG-ViT: A lightweight framework fusing multi-view RGB images and dual-task learning

Year: 2026

Type: article

Abstract: Nitrogen and chlorophyll are two key indicators of coffee nutrient status, yet their field estimation remains constrained by single-view imaging and single-task modeling, which overlook multi-view spatial heterogeneity and the physiological correlation between the two traits. To address this limitation, this paper proposes VAG-ViT, a lightweight dual-task Vision Transformer that jointly estimates leaf nitrogen (LNC) and chlorophyll (SPAD) from three-view (In-situ, Adaxial, Abaxial) RGB images, together with CCN-2025, a field phenotyping dataset of 571 sample groups spanning two phenological stages. VAG-ViT extracts cross-view commonalities via a shared shallow encoder, reconstructs high-order semantics via view-specific deep branches, and dynamically recalibrates view contributions via a view-adaptive gating (VAG) module that suppresses field noise; a dual-task regression head then exploits the physiological correlation between nitrogen and chlorophyll. Under five-fold stratified cross-validation, VAG-ViT attains R 2 of 89.02 ± 1.82 % (LNC) and 89.81 ± 1.89 % (SPAD), outperforming the best CNN baseline by roughly 4 percentage points, and the VAG module reduces RMSE by up to 45.7 % in the nitrogen-deficient bin. These results are achieved with only 3.98 M parameters and 1.14G FLOPs, about 76 % fewer parameters than a concatenation-based ViT-Tiny. Deployed in an Android application, the model delivers real-time end-to-end field inference, offering an efficient and promising route to precision nutrient monitoring for coffee and, potentially, other crops in complex field conditions.

Source: Computers and Electronics in Agriculture

DOI: 10.1016/j.compag.2026.112452