UAV-Assisted Wind Turbine Counting With an Image-Level Supervised Deep Learning Approach

Abstract

Unmanned aerial vehicle (UAV)-based autonomous equipment is increasingly employed by the Internet of Things (IoT) digital infrastructure of wind farms. Counting the number of wind turbines (WTs) of UAV-captured images can significantly improve the effectiveness of UAV inspection and the efficiency of wind farm operation and maintenance. However, existing counting methods generally require expensive object position annotations for instance-level supervision as well as a huge number of images to train models. In this article, we propose a two-stage algorithm that combines vision Transformer (ViT) and ensemble learning models to estimate the number of WTs of UAV-taken images. At the first stage, a ViT-based deep neural network is developed to automatically extract high-level features of input UAV images based on the self-attention mechanism. Next, at the second stage, an ensemble learning model, incorporating the deep forest and hist gradient boosting algorithms, is utilized to estimate the counts based on the extracted features. Experimental results show that the proposed algorithm can significantly improve the accuracy compared with the commonly considered and recently reported benchmarks.