How does artificial intelligence understand spatial data?
For artificial intelligence, data must be made machine-readable. Artificial intelligence does not see images in the same way as the human eye, which is able to recognize objects in an image by their color, shape, or context. For the model, an image is a collection of pixels in a specific order. Each pixel has its own value, such as a color code. The model learns that certain pixel values together form an object, such as a traffic sign. Instead of images, the model can also be trained to read video. Have you ever wondered how your car knows what the speed limit is on the road? While navigation apps (such as Waze) know the speed limit for each section of road based on location coordinates, many newer car models have a built-in traffic sign recognition (TSR) driver assistance system based on machine learning. In this case, a camera records the trajectory ahead, and a machine learning model analyzes the signs on the side of the road while driving and informs the driver of restrictions and hazards. This system is more sensitive because it operates in real time and is capable of "reading" temporary speed limit signs that are not marked on the map while driving. Similar technology is also used to train self-driving cars and delivery robots.
In addition to images or videos, the model also systematically reads spatial data in vector or table form. During training, the model learns the rules that apply to the objects being searched for, for example, that water bodies always have an attribute called "shore type" in their metadata. When searching for bog pools, for example, the shore type attribute is one of many indicators for the model that this may be the object being searched for. In addition, the model looks at, for example, the terrain and the color of the object's pixels in the orthophoto and finally decides whether to identify the object or ignore it. Through the combined effect of several factors, the model can also identify groups of objects with a regular layout, such as a row of windows on a building facade.
How does the model learn?
The model's decisions are closely related to the data it has been trained on, i.e., the data the model is able to read. For example, a model trained solely on satellite photos will not be able to identify the same objects on orthophotos because the resolution (pixel size) of the images is so different. While the identification of cats or people from images has been solved worldwide, there is still no universal ready-made solution for creating maps from aerial photographs, because the landscapes of different countries or even regions are different, as are the spatial data from which objects are identified. Hundreds of machine learning models that recognize buildings may exist in parallel, because a model that recognizes Manhattan skyscrapers as buildings will not be able to recognize a farm complex in Saaremaa.
In order for artificial intelligence to be able to identify the desired objects from data, such as images, the model must be trained. The irony of applying machine learning is that developing automated workflows usually requires enormous resources during the model training phase. A large amount of data must be produced for training, testing, and validation, and the model must be analyzed and adjusted according to the results until the desired quality is achieved. Therefore, it is wise to carefully consider the need for machine learning before starting training. Machine learning is best suited to tasks that are data-intensive in terms of processing, yet routine and clear-cut. Model analysis and tuning are usually performed by machine learning engineers and data scientists.
Why is it needed?
Land and Spatial Development Board updates maps for ¼ of Estonia each year, because that is how much territory the Board can cover in one season. This update cycle should ensure that the data is up to date in every location in Estonia. In reality, updating the map is such a time-consuming and labor-intensive process that the existing mappers are unable to completely update a quarter of Estonia's territory each year. Either more mappers or more optimal work processes are needed. Machine learning simplifies the work of mappers by feeding them changes in the territory under investigation. This allows mappers to focus on updating the map instead of searching for changes, significantly speeding up their work. However, this requires that the models are sufficiently well trained, i.e., reliable.
It is also important to understand that artificial intelligence does not produce maps itself. No model makes changes on behalf of the mapper or publishes data itself. Human control is and will remain an important component of the mapping workflow, as no model is 100% accurate. The mapper reviews the data presented by the model and decides for each object whether the data is correct or whether it is a false positive that needs to be discarded.
What have we already done with MaRus?
MaRus is a working model for identifying solar parks. As soon as new aerial images are ready, they can be fed into a deep learning-based model that has been trained to detect solar parks from orthophotos. Running the model results in a new map layer with all objects that the model considers to be solar parks. The results are then compared with previous verified results, making it possible to see what has changed in the meantime without mappers having to visually review the entire territory of Estonia.
However, it is normal to encounter obstacles when training a model. For example, when training the solar panel model, it was unexpected that the model had to be taught to distinguish solar parks from greenhouses. This was very easy for humans, but for the model, these were surprisingly similar objects.
Future plans
Land and Spatial Development Board has received funding from the State Chancellery's innovation fund to test deep learning opportunities in the map production process. Deep learning is a form of machine learning that uses artificial neural networks. In deep learning, the machine learning engineer does not give the model rules for identifying objects; instead, the model creates the rules and connections itself through the analysis of large amounts of data. With this solution, humans can influence the model by selecting training data and model architecture parameters, but the artificial neural network learns on its own.
The project will start this year and will focus on the automatic detection of watercourses, roads, and buildings from data of the Land and Spatial Development Board (e.g., orthophotos, LIDAR point clouds, oblique images, registers). In addition, the project will investigate whether and by what means it is possible to further determine the number of floors in buildings and the location of access points.
Author: Ann Rehemaa, Chief Specialist, Spatial Data Office, Land and Spatial Development Board
Loomise kuupäev: 14.11.2025
