Deep Learning and Climate Change
Deep Learning and Climate Change
How can deep learning solve the problem of climate change?
Join the DZone community and get the full member experience.Join For Free
The most visionary programmers today dream of what a robot could do, just like their counterparts in 1976 dreamed of what personal computers could do. Read more on MistyRobotics.com and enter to win your own Misty.
As a pioneer in the fight against global climate change, Germany is investing more and more in renewable energy, especially in wind energy. With about 300 new turbines from 2016 to 2017, North Rhine Westphalia is one of the major federal states in the construction of new wind turbines. To assess the potential of wind power and the planning of new turbines, it is essential to track the spatial location of wind turbines along with their type and to combine them with information on the characteristics of average wind speed.
In the present case of use, a methodology is described that can locate and segment wind turbines in satellite images. The implemented neural network architecture is called U-Net and is the leading standard for image segmentation. The output is a pixel-level prediction of the likelihood of a pixel belonging to a wind turbine. The deep learning structure was trained to predict wind turbine polygons in 280,000 satellite images covering the entire region of North Rhine-Westphalia. The output has been transferred to the ArcGIS Geographic Information System and can be accessed online through multiple devices. The wind turbine layer can be used for comprehensive analysis of wind power potentials and spatial planning of turbines.
In view of the growing scarcity of fossil fuels, renewable energy sources are becoming increasingly important economically, socially, and politically as an efficient and ecological way of generating electricity.
The objective of the project was to support the Federal Ministry of North Rhine-Westphalia in generating regional registers of sites and types of wind turbines to guide national energy producers in the process of spatial planning of new power plants.
A convolutional neural network (CNN) was formed to identify and segment wind turbines and wind farms on the ground, based on satellite images. Output wind turbine polygons can be fed into Geographic Information Systems (GIS) and can be enriched with current wind data to efficiently monitor current wind energy.
The satellite imagery used in this project contained 280,000 mosaics of images provided by Esri (the market leader in geoinformation systems) from World Imagery. The images cover the border area of the federal state of North Rhine-Westphalia and have an area of 1 km2 each. For each mosaic of images, the corresponding geographic metadata was crawled. For the training data set, 500 images were selected and 200 for the validation data set, including 200 wind turbines of different types and in different terrestrial coverage situations. For both sets, wind turbine polygons were created.
First, an image pre-processing was performed to normalize the satellite images for different levels of brightness, saturation, and contrast. Then the training and validation data had to be generated by the visual location of wind farms and turbines in ArcGIS Pro (Esri’s professional GIS tool). The localized turbines were marked and converted into dereferenced polygons. After combining the resulting polygons with the corresponding mosaic of images, they became picture masks. The mask classifies whether or not an image pixel belongs to a wind turbine. This serves as the desired classification scheme for the developed artificial neural network. The deep learning framework is based on a U-Net architecture, which has proven to work very well for segmentation tasks with a low amount of training data. Segmentation performance was tracked using the Jaccard index, which is an intersection in a union measure. The training was calibrated to achieve maximum accuracy in the validation set to avoid model overfitting. The final layer of the neural network generates an image mask with a pixel-level prediction of the likelihood of a pixel belonging to a wind turbine.
The first challenge was to create training functions and generate wind turbine polygons within the 700 satellite images. The unsupervised collation application, specifically a K-Means colour grouping, has aided in pattern recognition and polygon shape extraction. Another time-consuming challenge was to detect false positives in the network, ie recognized image segments that were falsely identified as wind turbines, such as roads or aircraft with branches. More training periods were required to train the neural network to differentiate between objects with a similar appearance.
Using the deep learning model developed, a regional register of wind turbines was successfully created for the state of North Rhine-Westphalia. In total, about 3,300 wind turbines have been identified in satellite imagery.
This record was also captured as a layer in ArcGIS Pro and is now available as map material within the software, showing the location of all wind turbines identified as polygons of their shapes. As a next step, the model can also be applied to other German states or even create a global register of wind turbines. The wind turbine layer can be combined with current wind speed data to monitor wind power generation and the average wind speed layers to support spatial planning of new wind turbine installations.
The deep learning model developed has already been used in other projects within the field of satellite image segmentation. Based on the satellite images provided, a stylized map material was created. The neural network was trained to detect different objects and types of ground cover in satellite images, such as roads, trees, forests, vehicles, buildings, rivers and agricultural fields.
Satellite imagery has multiplicative fields of application and can be used to gain a better understanding of other domains, eg to help identify natural resources more easily, to visualize and monitor climate or vegetative changes or to represent the impacts of natural disasters with accuracy. However, these things were mostly achieved through manual or semi-automatic methods. Detection of artificial intelligence functions and satellite images can contribute significantly to these geographic applications.
Opinions expressed by DZone contributors are their own.