Federated learning has emerged as a privacy-preserving and efficient approach for deploying intelligent agricultural solutions. Accurate edge-based diagnosis across geographically dispersed farms is crucial for recognising tomato diseases in sustainable farming. Traditional centralised training aggregates raw data on a central server, leading to communication overhead, privacy risks and latency. Meanwhile, edge devices require lightweight networks to operate effectively within limited resources. In this paper, we propose U-FedTomAtt, an ultra-lightweight federated learning framework with attention for tomato disease recognition in resource-constrained and distributed environments. The model comprises only 245.34K parameters and 71.41 MFLOPS. First, we propose an ultra-lightweight neural network with dilated bottleneck (DBNeck) modules and a linear transformer to minimise computational and memory overhead. To mitigate potential accuracy loss, a novel local-global residual attention (LoGRA) module is incorporated. Second, we propose the federated dual adaptive weight aggregation (FedDAWA) algorithm that enhances global model accuracy. Third, our framework is validated using three benchmark datasets for tomato diseases under simulated federated settings. Experimental results show that the proposed method achieves 0.9910% and 0.9915% Top-1 accuracy and 0.9923% and 0.9897% F1-scores on SLIF-Tomato and PlantVillage tomato datasets, respectively.
F EDERATED learning (FL) and lightweight neural net- works have independently achieved success across diverse domains. FL is widely applied in healthcare [1], finance [2] and mobile systems [3] for privacy-preserving collaborative modelling. Similarly, lightweight neural networks perform effectively in resource-constrained settings, where computational efficiency is critical [4]. Unlike other domains, agriculture uniquely supports the combined use of FL and lightweight neural networks. In automated plant disease recognition, data from geographically dispersed farms render centralised aggregation impractical and privacy-sensitive. In fact, this is particularly relevant for tomato, a globally cultivated crop with disease data available from multiple regions [5]. Moreover, deploying plant disease recognition models in rural and edge environments requires lightweight architectures optimised for limited computational and energy resources. The integration of FL with lightweight neural networks, therefore, presents a promising solution for achieving accurate, privacy-preserving and scalable plant disease recognition in agricultural settings.
In the agriculture domain, an ideal plant disease recognition system should (i) collaboratively learn from distributed farms while preserving data privacy and (ii) operate efficiently in terms of computation and memory on resource-constrained edge devices. However, existing plant disease recognition approaches typically focus on either centralised deep learning, which demands extensive data transfer and computation [6], or lightweight models that cannot fully exploit distributed data across farms [7]. Consequently, existing plant disease recognition approaches exhibit a notable gap in simultaneously addressing data privacy, distributed learning and computational efficiency. The integration of lightweight neural networks with FL has shown promising results in other domains and can be applied to plant disease recognition for globally cultivated crops, such as tomato.
To address these limitations, this study proposes an ultralightweight FL framework for tomato leaf disease recognition. In most cases, reducing the architectural complexity of deep neural networks to enhance computational efficiency often leads to a degradation in recognition accuracy [8]. Several approaches have been investigated to improve the performance of lightweight neural networks. Among these, attention mechanisms have consistently achieved superior results across diverse domains [9]. In the agricultural domain, the incorporation of attention modules into lightweight architectures has similarly yielded substantial gains in classification accuracy and generalisation performance [10], [11]. These improvements are primarily characterised by the ability of attention mechanisms to selectively emphasise informative features while suppressing irrelevant background noise, which leads to more discriminative representations.
Integrating lightweight architectures within FL enables privacy-preserving and computationally efficient training using distributed data. Nevertheless, deploying such models in FL environments presents significant challenges, such as data heterogeneity, statistical imbalance and variability in client performance. In-field plant disease datasets, particularly those for tomato, inherently exhibit a non-independent and identically distributed (non-iid) structure due to variations in environmental conditions, imaging angles and disease severity. These issues are further compounded by the limitations of arXiv:2602.16749v1 [q-bio.QM] 18 Feb 2026 conventional aggregation strategies, such as FedAvg [12], FedProx [13], FedNova [14], Scaffold [15], MOON [16] and Per-FedAvg [17], which are unable to effectively account for both differences in dataset size and local model performance.
In fact, none of these approaches effectively identifies the trade-off between dataset size and local model performance, a factor that significantly influences the overall accuracy of the global model.
To address the shortcomings of existing lightweight FL frameworks in managing highly non-iid and distributed data related to tomato leaf disease, we propose a set of targeted enhancements. First, we propose an attention-enhanced ultralightweight neural network designed for seamless integration within FL environments. Second, we develop a novel aggregation algorithm that adaptively balances client contributions by accounting for both dataset quality and local model performance. In summary, our study has three main contributions: 1) We propose an ultra-lightweight neural network (shown in Fig. 1) incorporating a novel attention mechanism, namely local-global residual attention (LoGRA), which effectively captures informative features while suppressing irrelevant background noise. 2) Through the proposed Federated Dual Adaptive Weight Aggregation (FedDAWA), as elaborated in Subsection III-B, we introduce a principled FL aggregat
This content is AI-processed based on open access ArXiv data.