Optimizing Lemongrass Disease Detection: A Comparative Analysis of Neural Architecture Search (NAS) with convolutional neural network (CNN) and Transfer Learning models

Abstract

Lemongrass is an economically important crop that faces significant disease challenges, making timely and accurate detection crucial for enhancing crop yield and quality. Traditional detection methods rely on manual observation, which can be subjective and inefficient. This study uses convolutional neural network (CNN) models to investigate automated detection methods for lemongrass diseases. We mainly utilize Neural Architecture Search (NAS) methods involving Efficient Neural Architecture Search (ENAS) and Partial Channel-Differentiable Architecture Search (PC-DARTS) to enhance structural models. Then, we compare the execution of these NAS-optimized models with transfer learning models such as VGG16, AlexNet, and Inception. We applied data augmentation and preprocessing methods to improve the model's performance. PC-DARTS achieved 92.19% accuracy with a validation set, with a significant reduction in computational resources, while the accuracy of ENAS was 83.33%. This paper explains the ability of NAS to detect lemongrass disease in real time, comparing it with other traditional approaches.