Incremental Cotton Diseases Detection Based on Aerial Remote Sensing Image and Blockchain Sharding

Abstract

The healthy development of cotton industry is of great significance to the economy of Xinjiang, and the effective management of pests and diseases is the key to ensure the stable development of cotton industry. How to improve the efficiency of cotton pest and disease model detection and get better training effect is a key issue in the task of cotton pest and disease management. Based on the incremental detection model, this article combines the UAV and blockchain sharding technology to create a new cotton pest and disease detection framework, UAV-IFOD-shard. First, the backbone network of YOLOv5n is replaced with ShuffleNetV2, and the squeeze and excitation module is introduced to maintain accuracy and speed. Optimize the neck network using deeply separable convolution to reduce parameters and computation. Improve path aggregation network fusion by replacing concatenation with additive fusion to reduce the number of parameters. Then, an incremental learning method based on knowledge distillation for cotton pest and disease targets is proposed on the basis of the lightweight model to realize parameter updating and memory retention for new and old targets. In addition, the blockchain is further partitioned and a reputation evaluation mechanism is added to the process of federated learning model aggregation to optimize the whole federated learning process. Finally, pest and disease images were collected from cotton fields in several surrounding areas by UAV to construct a dataset on which distributed federation learning was trained. The experimental results show that our model achieves better results than some existing methods, with a reduction of about 69.95% in model parameters, 60% in training time, and only a loss of 5.7% in accuracy. The UAV-IFOD- shard framework improves the system throughput of federated learning and the quality of the aggregated model, and also shows better performance in the face of malicious node attacks, and it is a good choice to use this framework for cotton pest and disease detection in Xinjiang.