Lightweight crop disease identification network based on frequency domain and channel mixing attention and...

Accurate identification of crop diseases is essential for enhancing agricultural productivity; however, it encounters challenges arising from complex field conditions and the constraints of deploying on resource-limited devices. This study aims to develop a lightweight yet accurate framework, referred to as FCDRNet, which integrates feature enhancement and compression techniques to facilitate practical deployment in the field.

FCDRNet introduces three key innovations: 1) a frequency-channel mixing attention (FCMA) module that integrates median-enhanced channel pooling with wavelet-based frequency attention to effectively capture both local and global features; 2) a cross-scale semantic fusion (CSF) module that facilitates adaptive multiscale lesion recognition; and 3) a DepGraph-RKD compression strategy that reduces parameters by 70.6% (from 4.32 M to 1.27 M) and FLOPs by 56.98% (from 256.73 M to 110.43 M). Evaluations on the Peanut Leaf Disease Dataset (PLDD) and PlantVillage Dataset (PD) datasets demonstrate that FCDRNet achieves accuracies of 96.60% and 99.67%, respectively, surpassing baseline models by 3.31% and 2.23%. Notably, the compression method maintains robustness with an accuracy degradation of ≤0.14%, enabling real-time inference at 14.84 ms on embedded devices.

FCDRNet offers scalable solutions for smart agriculture by synergistically integrating attention mechanisms, semantic fusion and dependency-aware compression. It achieves a balanced performance in terms of accuracy and efficiency, with accuracy rates of 96.60%, 99.67% and 97.77% on three datasets: the PLDD, PD and PlantDoc, respectively. This performance has propelled the development of practical, field-deployable crop disease monitoring systems, effectively addressing critical gaps in identification accuracy and the limitations associated with edge deployment. © 2025 Society of Chemical Industry.