Optimizing crop classification in precision agriculture using AlexNet and high resolution UAV imagery

Abstract

The rapid advancement of artificial intelligence, coupled with the utilization of aerial images from Unmanned Aerial Vehicles (UAVs), presents a significant opportunity to enhance precision agriculture for crop classification. This is vital to meet the rising global food demand. In this study, the effectiveness of an 8-layer AlexNet, a Convolutional Neural Network (CNN) variant was investigated for automatic crop classification. A DJI Mavic UAV was employed to capture high-resolution images of a mixed-crop farm while adopting an iterative training approach for both AlexNet and the conventional CNN model. Comparison based on performance was done between these models across various training epochs to assess the impact of training epochs on the model's performance. Findings from this study consistently demonstrated that AlexNet outperformed the conventional CNN throughout all epochs. The conventional CNN achieved its highest performance at 60 epochs, with training and validation accuracies of 62.83% and 46.98%, respectively. In contrast, AlexNet reached peak training and validation accuracies of 99.25% and 71.81% at 50 epochs but exhibited a slight drop at 60 epochs due to overfitting. Remarkably, a strong positive correlation between AlexNet's training and validation accuracies was observed, unlike in the conventional CNN. The research also highlighted AlexNet's potential to generalize its crop classification accuracy to datasets beyond its training domain, with a caution to implement early stopping mechanisms to prevent overfitting. The findings of this study reinforces the role of deep learning and remotely sensed data in precision agriculture.