Anyone who works with remote sensing knows that it is not entirely easy to produce truly high-quality data with a high degree of accuracy. Despite major advances in machine learning and AI, we still see many examples of products where the results are of low quality, especially when global models are applied to Swedish conditions.
Often, the problems aren’t related to the technology itself, but rather to shortcomings in the data, methodology, and understanding of the geographic context. The question, therefore, isn’t whether AI works or not. The question is under what conditions it works.
When AI-Generated Data Meets Reality
In recent years, a range of global and European datasets has become available. Many of them are based on AI-driven classification of satellite data and offer quick access to information on, for example, land cover, vegetation, or changes over time.
The problem arises when these products are used without their quality being verified against local conditions. In Sweden, we see many examples of European data where features such as wetlands, clear-cut areas, forests, mountain heathlands, and low-growing vegetation have been misclassified or overestimated/underestimated. For certain applications, such errors are of little consequence, but when the data is used for urban planning, environmental monitoring, climate analysis, or investments, the consequences can be significant.
It is important to understand that an AI-generated dataset is not automatically a quality-controlled dataset.
Some common pitfalls
Pitfall 1: The model was trained on a different reality
One of the most common causes of poor quality is that models are trained on data that does not represent the environments where they will later be used.
A model that works well in Central Europe does not necessarily work as well in Norrbotten. Differences in vegetation, topography, climate, and land use affect what the satellite data looks like and, consequently, how the model interprets the landscape.
It is easy to underestimate just how significant these differences actually are.
Sweden encompasses everything from agricultural landscapes in the south to mountain environments, wetlands, and boreal forests in the north. Creating robust models for such conditions requires training data that represents the full range of geographical variation.
If the training data is limited, the results will be as well.
Pitfall 2: Scaling up is harder than many people think
It is relatively easy to create a model that works well in a test area. The real challenge arises when the same model is to be used nationally or internationally. When a project is scaled up, the model encounters new conditions, such as different soil types, habitat types, climatic conditions, and seasons.
A model that demonstrates high accuracy in a research study may therefore perform significantly worse when used operationally across large geographic areas. This is often where the difference between a promising prototype and a production-ready solution becomes clear.
Pitfall 3: Data Quality Is Underestimated
The discussion about AI often centers on algorithms. In practice, however, it is data quality that determines the final result. If the input data contains errors, gaps, or deficiencies, the model will learn incorrect correlations. The same applies if the training data is inconsistent or has not been adequately quality-assured. Therefore, it is crucial to devote significant effort to developing or producing your own training data.
In remote sensing, results are also influenced by factors such as clouds, atmospheric disturbances, snow cover, shadows, sensor variations, geometric errors, and variations between different data collection times. Addressing these challenges requires significantly more than simply training a model.
Pitfall 4: Lack of Domain Knowledge
A fourth challenge is that AI development often takes place without sufficient knowledge of the reality being described. Machine learning experts are often highly skilled in models and algorithms. But to create useful geographic products, expertise is also needed in remote sensing, geodata, and the operations that the results are intended to support.
For example, it is not enough to know that a pixel is classified as forest. One must often also understand what type of forest it is, how it differs from surrounding land types, and what consequences a misclassification has for the user.
When domain knowledge is lacking, models risk being optimized for statistical metrics rather than actual usability.
AI Requires More Than High Accuracy
A common problem in AI projects is that the focus ends up being on the model’s accuracy rather than on the data quality in the final product. A model can achieve impressive results in technical evaluations and still be difficult to use in practice.
That is why it is very important to validate the results in other ways.
For organizations working in urban planning, environmental analysis, or natural resources, it’s important to ask these questions when faced with a new dataset:
Without answers to these questions, it becomes difficult to build trust in the results.
Metria’s approach: from AI model to a functional information product
At Metria, we view AI as a powerful tool. Our experience with national surveys, geographic analyses, and remote sensing projects shows that it is the interplay between technology, data, and domain expertise that creates value.
That is why we place great emphasis on:
AI will continue to transform how geographic information is produced. The possibilities are vast, and developments are moving quickly.
But experience shows that the most valuable solutions are rarely based on the most advanced model. They are based on the right data, the right methodology, and a deep understanding of the problem to be solved.
For users of geographic information, therefore, the most important question is not whether a product is AI-generated. The most important question is whether the product meets the quality standards required for the specific purpose for which you intend to use it.
Metria offers digital solutions and services in GIS, geodata, real estate, and business information. Our offering covers the entire process—from identifying a customer’s needs for geodata, real estate, and business information to collecting, analyzing, and visualizing data to generate insights that lead to smarter, safer, and greener decisions.
Since spring 2022, Metria has been part of Spir Group, a Nordic group with approximately 260 employees in Norway and Sweden.
Spir Group is a Nordic company that simplifies unnecessarily complex processes by collecting and making information accessible to consumers, the public sector, and the private sector.
Spir Group is the parent company and the visionary force behind several software subsidiaries, all dedicated to delivering business-critical technology that sustains and develops society.
Our customers range from real estate agents, banks, insurance companies, appraisers, real estate developers, media companies, builders, property owners, engineers, and energy companies to manufacturers of building materials.
We are a team of more than 260 colleagues with deep domain and technology expertise. Spir Group does more than just collect and share information. We enable innovation and growth that strengthen trust among people, businesses, and society.
For more information about Spir Group, visitwww.spirgroup.com.