What are the most common challenges and pitfalls of machine learning in geomatics?

The old garbage in -> garbage out still is very true for the geospatial field. And though I do think that ML could be able to fill gaps and filter errors, I fear for an enormous number of faulty applications before this will be the case. Every GIS analysis of a decent size and complexity that I've seen is still being checked and checked again for methodical errors, data errors, bias, etc. I wouldn't trust ML applications yet to do work that goes slightly beyond the most basic of things. The "yet" is a big thing here btw. It will come.

Spatial data, the lifeblood of Geomatics, stands out for its intricacy and significant cost. Though data like remote sensing offers substantial resource savings, meticulous radiometric and geometric corrections are crucial for accuracy and consistency. Moreover, most of it carries a hefty price tag. Standardization is another hurdle. While advancements in standards are ongoing, their application remains intricate due to the sheer diversity of geomatics datasets, spanning fields like politics, tourism, and smart cities. However, open-source data access is steadily expanding, especially for public institutions. While often demanding thorough understanding and potential cleaning, this data represents a valuable resource.

Spatial data is inherently temporal, as the physical Earth and human-made structures are in a constant state of flux. This dynamic nature is further intensified in today's world, where people's movements generate an infinite stream of spatial data, adding a layer of complexity. Moreover, geographical features are interconnected not only in space but also in time. Machine learning algorithms, however, often struggle with these conditions. To address this challenge, one solution involves employing more advanced ML algorithms that are less sensitive to spatial dependencies. While this approach can mitigate some issues, it also introduces additional complexity, highlighting the inherent challenges of working with spatial data.

In machine learning, the golden rule of interpretability states that as model complexity increases, interpretability often decreases. This challenge becomes even more pronounced when analyzing spatial data, which demands a spatial mindset. Therefore, a fair approach involves introducing complexity incrementally. Accompanying this process with effective visualizations is crucial. To create impactful visualizations, begin with a clear goal and a well-defined message tailored to your audience. Ensure your data is tidy. Actively seek feedback. Prioritize clarity and success over endless tuning. Remember, as a professional, you must be prepared to defend your results with rational arguments backed by strong evidence.

Actually, the thing that worries me most is that quite a few people will have full confidence in the AI models. Especially the managers and finance people will be drooling over the possible cost savings. People who still understand (traditional) geospatial stuff will be direly needed for the checks and balances involved, and they won't be around anymore. This would make the perfect ethical nightmare scenario. All issues with the unfair, inaccurate or malicious models will damage society as a whole. So yes, there are indeed ethical and social implications. Not as much in the Artificial Intelligence, but in the Natural Stupidity in humans.

One of the key ethical challenges in geomatics concerns land price management. As the cost of living rises globally, ensuring fair and equitable land pricing becomes increasingly crucial. Geomatics engineers and professionals carry a significant responsibility in this regard. They must develop and implement models that prioritize fairness for all, not just specific groups like investors or public institutions, but also research centers and the general public.

One of the biggest misconceptions I've encountered as a data expert is the assumption that anyone can analyze any type of data. While basic insights are attainable by many, just like a doctor can't specialize in every medical field or an athlete can't excel at every sport, true data proficiency requires a deep understanding of your specific dataset. In the realm of geomatics, arguably the most crucial skill is spatial awareness. All data in this field relates to objects in space or geographical features. Developing this key mindset is fundamental to ensuring meaningful results.