Enhanced Crop Disease Detection With EfficientNet Convolutional Group-Wise Transformer

Crop diseases, as one of the major problems in global agricultural production, lead to crop yield reduction, death, and even total extinction, with serious impacts on farmers and the food supply. Traditionally, crop diseases are identified by visual inspection and based on the experience of farmers and agricultural experts, a method that not only consumes human resources but also has a certain degree of subjectivity and inaccuracy. The development of artificial intelligence technology successfully achieves real-time monitoring, automatic identification, and intelligent decision by combining the Internet of Things (IoT) technology and cloud computing technology. Herein, we proposed an EfficientNet Convolutional Group-Wise Transformer (EGWT) architecture. The local features of crop leaf images are extracted by EfficientNet convolution and then input into a group-wise transformer architecture. In the group-wise transformer process, the input features are divided into multiple groups. An attention mechanism is used within each group to calculate correlations between features. After calculating the intra-group attention, the output features of each group are stitched together to form the final output features. Our proposed model achieves 99.8% accuracy on the PlantVillage dataset, 86.9% accuracy on the cassava dataset, and 99.4% accuracy on the Tomato leaves dataset, with the least number of parameters 23.04M in the state-of-the-art convolutional combinatorial transformer hybrid model. The experimental results indicate that the proposed model has the best accuracy and optimal model complexity so far compared to other neural networks based on CNN, transformer, and the hybrid structure of CNN and transformer.