A Novel Hardware-Aware Multi-Objective Evolutionary Neural Architecture for Crop-Disease Classification on Edge Devices

Abstract Crop disease detection plays a pivotal role in improving yields and quality, thereby supporting the efficient utilization of agricultural resources. Conventional detection methods suffer from limited scalability, subjective detection, and reduced reliability under varying environmental conditions. Machine learning and deep learning-based approaches have demonstrated their strong capabilities for efficient crop disease classification from versatile field and plant images. However, non-scalable architectures of these existing models consume large amounts of parameters, computational overhead, memory, and high latency. To address the aforementioned challenge, this research paper proposes hardware-aware, multi-objective, evolutionary search-based neural architectures for efficient crop detection on low-power edge devices. The proposed framework utilizes an EfficientNet-Lite model as the main component, whose parameters are optimized by the Non-dominated Sorting Genetic Algorithm II (NSGA-II) to achieve high classification performance, low inference latency, and limited computational parameters. In the proposed optimization, the framework predicts the performance of learning networks and estimates their execution time on the target edge hardware without testing each model directly. It then identifies a set of efficient architectures with varying accuracy, latency, and energy trade-offs, enabling the selection of the most suitable model based on the target device's requirements. Extensive Experimentation is carried out using PlantVillage datasets, and performance metrics such as accuracy, precision, recall, F1-score, latency, and computational parameters are calculated. Results illustrates the suggested framework has achieved the accuracy of 0.98, precision of 0.9734, recall of 0.973 with latency and parameters on different edge architectures. Compared with EfficientNet-B0 and MobileNetV3, it is [A]× faster and [B]× smaller, with accuracy within [C]%. These results show that hardware-aware evolutionary search can produce compact and accurate models for real-time plant disease monitoring in the field