'DeepSolar' that precisely identifies the position and scale of solar panels from satellite photographs through machine learning

by Wayne National Forest

Solar panels that can operate photovoltaic power generation in various places, from the rooftop of buildings to large outdoor facilities, are used throughout the world. In order to accurately grasp the output of such solar power generation, Stanford University is advancing " DeepSolar " to accurately grasp the position and scale of the solar panel by analyzing the image taken from the satellite by AI

DeepSolar: A Machine Learning Framework to Efficiently Construct a Solar Deployment Database in the United States - ScienceDirect
https://www.sciencedirect.com/science/article/pii/S2542435118305701

There are way more solar panels in the US than we thought
https://www.fastcompany.com/90284523/there-are-way-more-solar-panels-in-the-us-than-we-thought

Solar panels can produce electricity from fossil fuels and renewable energy that is not nuclear power, and it will also be possible to save household electricity bills. From the households installed on the rooftop to the industrial use installed on the vast premises on the outskirts, the position and scale of the solar panels are diverse and it is very difficult to grasp the whole of the solar panels, so the solar panels Currently it is not known exactly how much power generation is done by photovoltaic generation.

DeepSolar is a deep learning framework that analyzes satellite images and identifies the GPS coordinates and sizes of solar panels. Researchers at Stanford University have analyzed high resolution satellite imagery using the Kush algorithm and the machines have identified nearly all solar panels in 48 states in the USA. The research team insists that its finding accuracy is about 93%, which is more accurate than previous models.

As a result of DeepSolar's analysis of more than 1 billion satellite images, 1.47 million solar panels were discovered. The installed number of 1.47 million vehicles seems to be much higher than the value of 670,000 by Google's Project Sunroof which similarly verifies the installation of solar panels using AI and satellites.

By knowing how many solar panels are present, we can make renewable energy power generation more popular. Ram Rajagopal, associate professor of civil engineering and environmental engineering at Stanford University, says, "Electric utilities and system operators need to grasp the surrounding areas where sunlight is located and conduct a new survey on the project. Then we can decide which region we will invest in. "

In addition, it became clear that the higher the population density or the more household income per year, the higher the introduction rate of solar panels will be. However, when it exceeds 150 thousand dollars (about 16.8 million yen) the increase is stopped at about 80%, it turned out to be rather low. The research team thinks that there are things that are difficult to introduce in the low-income group because of the initial cost and maintenance cost of installing solar panels. In addition, there was a correlation between introduction rate and educational level.

Rajagopal said, "To analyze this number of satellite imagery, we must be able not only to accurately detect whether a system exists, but also to be able to process at high speed. "It was the first challenge we had to overcome," he said, using DeepSolar in countries other than the US to clarify the intention to update databases to the public every year.