Deep Block lands its first PoC with the city of Seoul.

Deep Block recently completed a Proof of Concept (PoC) project with the city of Seoul to develop AI models for analyzing remote sensing imagery. The project aimed to address the challenges faced by the city of Seoul in tracking illegal building construction and changes in land use, which are crucial for tax purposes and building safety. Previously, Seoul city spent around a million USD a year to analyze aerial photos for illegal use of land and rooftops, but this work was outsourced and done manually due to the difficulty of detecting small, illegal artifacts.

For the manager of the geoinformatics division of Seoul City, the current process is time-consuming and inefficient: "The problem is, most of the illegal artifacts are too small to recognize. Therefore, we have to use the highest resolution in aerial imagery we can find, which leads to very large files that are difficult to process. As a result, we spend a lot of money every year just to analyze aerial photos manually.”

Deep Block's AI technology provides a potential solution to the challenges faced by Seoul City. The company's algorithms can detect illegal construction and changes in land use automatically, saving time and money compared to manual analysis. Deep Block uses state-of-the-art multi-temporal change detection models to track changes in ground truth.

For Gwihwan Moon, CEO of Omnis Labs and creator of Deep Block, this new model brought significant improvements compared to the previous experimentations led by the division: "The Seoul Metropolitan Government explained about the failure of the research project they funded in 2022, where they tried to develop a similar change detection model with another company. They asked us to create a model with increased precision and performance. After investigation, we found that the previously developed model was over 8 years old and had many false positive cases. We also found that the mAP calculation formula made by the previous company was incorrect. We fixed the calculation formula and obtained a validation dataset standard mAP of about 0.89."

"However, our Bi-temporal change detection model is more sensitive than expected", continued Gwihwan Moon, "and we were confronted with a problem when finding changes in a very small area. So, we had to remove these small ignorable change areas through post-processing."

For Gwihwan Moon, the geoinformatics division's requirements also proved to be an interesting challenge: "They asked us to show the changed area a little larger so that people could easily see it. Using another post-processing technique, we marked the change area slightly larger than the actual change area, and these post-processing results are shown in the video below."

Video of Change area extracted from remote sensing imagery of the city of Seoul, data set available here.

"Our analysis process can now be significantly shortened, thanks to the AI model developed by Deep Block,” said the manager of the geoinformatics division. “By accurately identifying changed areas in a short amount of time, we can now complete the initial stage of the analysis much faster. Then, our analysts can focus their efforts on examining the remaining areas to identify any changes that may have been missed by the model. With this approach, we can cut our analysis costs by at least half."

The next step will be to combine drone imagery and Machine Learning for inspection and maintenance to investigate bridges to find cracks and other defects. The use of cutting-edge algorithms such as line segmentation, would greatly fit this use case.

The technology can also be used to track changes in water sources and forested areas, making it useful for a wide range of other applications. Deep Block plans to approach other local governments and agencies in Korea that are facing similar challenges, including water management organizations, forest departments, different ministries, and defense agencies. "With Deep Block's technology, local governments can save time and money while improving the accuracy of their analyses," said Gwihwan Moon.

Deep Block's success with Seoul City demonstrates the potential of AI in geospatial analysis. As more governments and organizations recognize the benefits of this technology, we can expect to see more partnerships and projects like this in the future.