Customizing Topographical Maps
BGT maps are detailed digital maps of the Dutch landscape that have important objects segmented and labeled such as buildings, roads, water bodies and more. Land and neighborhoods often change and performing all of these updates manually is an arduous process. By enhancing or supporting the BGT charts’ customization process with Computer Vision (CV) models, the cataloguing can be made quicker.
It is becoming cheaper, faster, and easier to create CV models and the value added to data processing has been documented in case after case. The Dutch government is already adopting them, for example, in municipalities and water boards. We have already enjoyed multiple partnerships and collaborations to build and innovate Computer Vision models that can identify (changes to) certain objects or landscapes from aerial photos. The improved efficiency helped these organizations automate processes or improve analyses that had previously been done by hand.
Help Organizations Make Sustainable Impact
One of the core principles for these organizations was not just sustainability of data management, but also shifting towards a sustainable clean energy paradigm supporting climate stability and human resilience. Given our expertise in the area of Computer Vision, we created an internal project, and deployed it as part of our data expertise training, to help these organizations make an impact on the most urgent issues facing us. We posited that including solar panels as a geographic label in the BGT was a great way to harness data technology to work towards sustainability. Our proof of concept was to determine how to easily identify solar panels using aerial, including infrared, photos, and height maps; our goal was to provide robust reliable data that could lead to a systematic understanding and promotion of solar energy improvement or growth.
First, we needed a reliable open-sourced dataset. The municipality of Amsterdam has had a long history of organizing public information, and there has never been more data and better access to it. The city offers high resolution aerial photos (10x10cm per pixel) of the entire area, and an older, still very relevant, dataset from 2017 locates all of the roofs with solar panels (but not the amount, nor the exact location of each individual panel). The other requirements (infrared photos and height maps) were taken from PDOK, an open-source dataset that includes aerial and remote sensing data of the Netherlands.
Labeling with a Mechanical Helping Hand
We still had to label all solar panels in Amsterdam manually, but we had a mechanical helping hand. We used QGIS, an open-source geographic information software to speed up the process. Below is an example of the labeling results. On the left, there is a regular aerial photo capturing four buildings in Amsterdam. Labels for solar panels were added and displayed in red. Labeling all solar panels took roughly 30 hours, which sounds tedious, but work was made lighter by many hands; cataloguing was divided between multiple colleagues and spread out over weeks. It was easy enough to do while listening to a good playlist or podcast and helped give the team a mental break from programming.
Model as a Road Map for New Areas
After we completed labeling, we got to the fun part: training the segmentation model! We used a program based on the DeepLabV3+ architecture with a Resnet50 backbone. More parameters are involved in this model, which makes it very energy and memory intensive to run. Few machines can keep up with this. Thankfully, since the start of 2021, Young Mavericks assembled and installed their own server, which can be deployed whenever a project requires a boost in extra computing power.
Throughout the training, we used 70% of the labeled dataset, and a subset of the other 30%
was used for validation and testing. The results looked promising, as you can see in the image below, as did the metrics (recall≈0.89, MIoU≈0.77). It should be noted that the data used for this project is unique to Amsterdam and the city’s quality and abundance of source data. Therefore, there is no guarantee that this model is so easily replicated in different areas. However, this model can serve as a backbone or road map for whenever data from a new area is added to the training set.
Identifying and Expanding Solar Power
This proof of concept showed that a computer vision model can learn to quickly recognize solar panels from aerial images. This capability can help give organizations a better overview of the current solar panel coverage in their area of interest. Furthermore, with images available from previous years, a comparative analysis can be done to see how and where solar energy has developed over the last few years. This same information can help to understand where solar panels currently don’t exist to help guide policies, programs, and efforts to assist those areas.
Currently, we only segment the surface area of solar panels, but with a few extra features down the road, we will be able to translate surface area to nominal panels or create a new model that recognizes each panel as an individual entity. This will lead to more complex, precise, and workable data sets. Our internal effort to innovate data engineering and model data that matters continues!