# Estimation of Wind Turbine Heights with Shadows Using Gaofen-2 Satellite Imagery

## TL;DR

This paper presents a method to estimate wind turbine heights using satellite imagery and deep learning, which helps monitor turbines after natural disasters.

## Contribution

A novel method combining deep learning and spatial geometry for wind turbine height estimation using GF-2 satellite imagery.

## Key findings

- YOLOv5-CBAM and MSASDNet achieved 96% identification accuracy and 82.53% shadow extraction accuracy.
- The method achieved an average absolute error of 2.2 m in height estimation.
- The technique effectively detects post-disaster status of wind turbines.

## Abstract

Using high-resolution remote sensing imagery to obtain the wind turbine height is a fast and effective method for monitoring the status of wind turbines after natural disasters such as earthquakes, landslides, and typhoons. A height estimation method tailored for wind turbines is proposed using high-resolution satellite images. First, deep learning techniques are employed to identify wind turbines and extract their shadow information from GaoFen-2 (GF-2) satellite imagery. Specifically, YOLOv5-CBAM and MSASDNet are used for target recognition and shadow extraction, achieving an identification accuracy of 96% and a shadow extraction accuracy of 82.53%. Next, the line-by-line scanning method is applied to remove blade shadow from the whole wind turbine shadow. By calculating the number of pixels occupied by the shadow length of the wind turbine after removing the blade shadow and multiplying by the image resolution, the wind turbine shadow length is obtained. Finally, a spatial geometry model involving the satellite angles, solar angles, and wind turbine shadow length is constructed to retrieve the wind turbine height. An experiment was conducted using GF-2 satellite remote sensing data from a wind farm in Huailai County of China. The actual heights of wind turbines in the estimation area were measured by the field experiment, and the average absolute error was verified to be 2.2 m, demonstrating the effectiveness of the proposed method. The experimental results show that this method can detect the post-disaster status of wind turbines.