Mobile application for the detection of diseased apples: Comparison between SOM and CNN networks

Abstract

Disease detection in apple trees represents a critical challenge for modern agriculture, involving production losses and excessive use of agrochemicals. This study developed a mobile application that compares two artificial intelligence approaches: convolutional neural networks (CNN) and self-organizing maps (SOM), to identify diseases in apples. Through a comparative analysis of more than 5,000 field images, it was shown that the SOM neural network achieved 94% accuracy vs. 92% accuracy of CNN with processing times of less than 10 seconds, while maintaining greater robustness in varying conditions. Although CNN showed advantages in computational efficiency, its lower accuracy determined the final selection of the SOM architecture for the application. The implemented solution operates on standard mobile devices, offering small producers an accurate tool that reduces post-harvest losses, which represents a significant advance in the democratization of low-cost agricultural technologies.