Advancing real-time plant disease detection: A lightweight deep learning approach and novel dataset for pigeon pea crop

摘要

Plant disease detection and early disease treatment are essential for sustainable crop production. Computer vision for crop science is overgrowing with the advancement in deep learning. Real time plant disease detection poses a challenge due to the unpredictable spread of diseases within the plant, environmental factors, and the scarcity of real field datasets. The proposed work systematically addresses these issues through three key components: (a) Collaboratively generating the novel pigeon pea image dataset from agricultural fields, in partnership with 20 Agricultural Research Centers (ARS) and governmental agencies spanning 18 Indian states. (b) The design of lightweight and high-performance models for real-time plant disease detection in resource-constrained devices. (c) The extraction of multiscale feature of plant diseases using Multi-kernel Depthwise separable Convolutions. The proposed lightweight Lite-MDC architecture uses the Multi-kernel Depthwise separable Convolutions (MDsConv). The MDsConv module captures spatial features across various scales while maintaining a lightweight design. It effectively extract multi-scale information to characterize plant diseases, accommodating their diverse scale. Proposed architectural approach significantly reduces computational complexity, employing only 2.2 million parameters, which is a 62% reduction compared to the standard VGG16 architecture. The proposed method outperforms the state-of-the-art networks such as InceptionV3, VGG16, ResNet50, DenseNet, MobileNet, MobileNetV3, NASNet, and EfficieNetB0 on the proposed pigeon pea dataset with 94.14% accuracy. Notably, the method achieves a 34 Frames Per Second (FPS) inference on an NVIDIA P100 GPU. Furthermore, its performance is validated across publicly available datasets, including the plant village dataset, Cassava, and apple leaf datasets, yielding 99.78%, 86.4%, and 97.2% accuracy, respectively. The Lite-MDC model exhibits the potential for real-time plant disease detection on resource-constrained edge devices such as Agriculture robots and drones.