TITLE:
MAF-Net: A Multi-Stream Collaborative Attention Fusion Network for Fine Grained Tomato Leaf Disease Classification
AUTHORS:
Emad Badry, Abeer Abdulaziz Alarfaj, Hosam E. Refaat, Sherif Fathy Ibrahim Nafea, Eslam Mohamed Khairy, Ahmed Awad Mohamed
KEYWORDS:
Tomato Leaf Disease Classification, Multi-Stream Feature Learning, Cross-Attention Fusion, Collaborative Attention, Lightweight Deep Learning, Precision Agriculture, Edge Intelligence
JOURNAL NAME:
Open Journal of Applied Sciences,
Vol.16 No.9,
September
29,
2026
ABSTRACT: Timely and accurate detection of plant diseases is crucial for sustainable smart farming and precision agriculture. However, existing deep learning techniques generally use single-stream feature extraction or static feature fusion strategies that limit their capability of effectively utilizing complementary visual patterns and maintaining computational efficiency. To overcome these limitations, in this paper, we propose a new framework named MAF-NET (Multi-Stream Attention Fusion Network) for tomato leaf disease classification. The proposed architecture presents a collaborative feature learning strategy where four complementary feature streams, namely texture, color, multi-scale spatial, and morphological representations, interact through a Collaborative Cross-Attention Fusion Module (CCAFM). In contrast to conventional feature concatenation methods, the proposed mechanism enables adaptive information exchange and feature refinement among heterogeneous representations, leading to a more discriminative feature representation for disease recognition. The proposed framework was evaluated using Tomato Leaf Disease Dataset consisting of large number of labelled images covering ten tomato leaf categories. The experimental results demonstrate the overall classification accuracy of MAF-NET reaches 97.6%, superior to the representative deep learning models such as ResNet50 (75.0%), MobileNetV2 (78.0%), EfficientNetB0 (81.2%) and Vision Transformer (93.4%). The proposed framework also balances the accuracy and computational efficiency well with only 6.84 million parameters and 0.72 GFLOPs, which is suitable for resource-constrained agricultural applications. Moreover, statistical analysis and ablation experiments are conducted to verify the effectiveness of the proposed collaborative multi-stream learning strategy and attention-based fusion mechanism. The obtained results demonstrate that MAF-NET is an accurate, lightweight and computationally efficient solution for intelligent tomato disease diagnosis and has a strong potential to be used in the future for precision agriculture and edge-based plant health monitoring systems.