Aprendizado federado para classificação da praga do milho em campo de agricultura

Abstract

Corn planting is one of the most common and economically relevant crops globally. However, despite its great economic and nutritional importance, corn plantations are still affected by the pest Spodoptera frugiperda (J. E. SMITH, 1797) (LEPIDOPTERA: NOCTUIDAE). To mitigate the problems caused by this pest, monitoring and control strategies are employed, including the use of camera-equipped traps for image capture and the application of biological or chemical control methods. With advances in technology, new monitoring methods for these pests have been proposed, such as deep learning. Nevertheless, these techniques do not address data privacy issues, as the data may contain confidential and market-sensitive information from the farmer’s field, nor do they address the challenges of distributed scenarios, such as limited connectivity and non-homogeneous data distribution among users. Federated learning proposes decentralized training, in which data is not shared, remaining on the user’s device and ensuring greater privacy. This work proposes an approach using federated learning with various aggregations, ensuring greater privacy and reduced data transmission, while minimizing network resource utilization. Additionally, a simulation of agricultural field scenarios is used to evaluate which aggregation methods are promising for distributed scenarios. We used a database containing images captured by cameras integrated into Raspberry Pi in traps. We divided these images into two classes: with the presence of S. frugiperda and without the presence of the insect. This dataset was distributed in two distinct ways: vegetative (characterized by the number of visible leaf collars) and reproductive (marked by grain development), representing the corn growth stage, and both non-independent and identically distributed. We tested the two scenarios with four aggregation algorithms: FedAvg, FedAvgM, FedProx, and FedYogi. Furthermore, we analyzed CPU and memory consumption. The results show that, across all scenarios, FedProx achieved the best performance in most model evaluation metrics, though with reservations about computational cost. In the vegetative scenario, FedYogi proved to be a good option, consuming fewer computational resources and presenting good results for evaluation metrics.