Abstract: Intelligent agriculture has become a major trend in the development of modern agriculture. Low-altitude UAV remote sensing image analysis is the key point in precision agriculture, which guides UAV to carry out precise job by studying on hyperspectral remote sensing images taken by UAV. However, small and medium-sized farms have the disadvantage of insufficient computational resource. To deal with the problems, a classification method based on lightweight Convolutional Neural Networks (CNN) is proposed. This method aims to reduce the training cost and maintain an acceptable classification accuracy, so as to help farms avoid replacing expensive high-performance equipment. PCA is used to reduce the image spectral dimension and some data augmentation methods are used to enlarge the dataset. The CNN structure is optimized to accelerate the training process and the model used to extract Spectral-Spatial feature to improve classification accuracy. Experiments are conducted on three benchmark datasets, which shows that our lightweight CNN model can guarantee a satisfying accuracy while having a lower training cost compared with both traditional methods and deep learning method.

Key words: intelligent agriculture, lightweight, hyperspectral remote sensing, convolutional neural network