Abstrakt: |
Leaf diseases are hazardous to the yield and quality of food crops. There are many deep learning algorithms developed for their detection that may require large computational resources to train a single model with voluminous amounts of data. In the agriculture domain, various plant diseases are found across the country. Accumulating such a huge dataset from various regions is a tedious task, and training them with a single model can also be challenging. This paper proposes Federated Deep Learning (FDL) for Plant Leaf Disease Detection. This concept allows multiple local models to get trained with their region-based datasets and share their knowledge with siblings through the parent, instead of sharing complete datasets. Knowledge transfer significantly reduces the computational costs. This paper formulates the federated dataset using PlantVillage, to simulate the configuration of the FDL. A lightweight and efficient Hierarchical Convolutional Neural Network (H-CNN) is proposed for parent model and child models with 0.09 million parameters and 0.35MB model size. The simulation results show that the proposed FDL attains 93% testing accuracy, outperforming state-of-the-art methods namely, FedAdam and FedAvg with 86.8% and 87.7% respectively. In addition to this, the proposed FDL achieves 95.7%, 95.4% and 95.3% of weighted precision, weighted recall and weighted F1-score respectively for first local model and 92.1%, 90.8% and 91.2% respectively of weighted precision, weighted recall and weighted F1-score respectively for second local model. [ABSTRACT FROM AUTHOR] |