First Dataset of Wind Turbine Data Created at National Level With Deep Learning Techniques From Aerial Orthophotographs With a Spatial Resolution of 0.5 M/Pixel

Abstract

Deep learning applied to feature extraction and mapping from high-resolution images is demonstrating the potential of this branch of data-intensive Artificial Intelligence to improve terrain mapping processes. The documented experiences have been applied on a small scale and there is a great expectation about its applicability on a country scale. For example, when extracting wind turbines using semantic segmentation models from a region of 28 km × 19 km containing unseen data, we obtained a commission rate of 1.4% and an omission rate of 0.38%. In this article, we present a methodology generated on the basis of two iterations. In these iterations, processing and postprocessing time, energy consumption, and finally results have been optimised to map wind turbines for the first time throughout the Spanish peninsular territory. In addition to adding a binary classification neural network prior to the semantic segmentation that extracts the turbines, a third multiclass recognition network has been used to classify the turbines by their power capacity complementing the features extracted with attributes. The proposed methodology can be adapted in the vectorisation phase and applied to other types of features with linear or polygon representation to achieve a large-scale efficient extraction of geospatial elements using automated procedures.

Wind turbine extraction utilizing high-resolution satellite imagery

“…Manso et al proposed a wind turbine extraction method that combines image classification and segmentation. They introduced a binary classification neural network before the semantic segmentation network to classify wind turbines based on their power [6] . Schulz et al presented a wind farm identification model called DetEEktor, which is based on the Mask R-CNN instance segmentation network.…”