High-Resolution UAV Image Classification of Land Use and Land Cover Based on CNN Architecture Optimization

Unmanned aerial vehicle (UAV) images have high spatial resolution and are cost-effective to acquire. UAV platforms are easy to control, and the prevalence of UAVs has led to an emerging field of remote sensing technologies. However, the details of high-resolution images often lead to fragmented classification results and significant scale differences between objects. Additionally, distinguishing between objects on the basis of shape or textural characteristics can be difficult. Conventional classification methods based on pixels and objects can indeed be ineffective at detecting complex and fine-scale land use and land cover (LULC) features. Therefore, in this study, the efficiency of UAVs in image acquisition was integrated with the capability of convolutional neural networks (CNNs) to extract high-level abstract features from images. Optimized detection models were developed by tuning CNN architectures and hyperparameters, using a UAV image dataset for training and testing. The performance of the models was then evaluated in classifying four types of LULC. All six models achieved training accuracies exceeding 97%. Model 6 achieved the highest accuracy on the testing set (89.4%) and the least overfitting.