Fine-tuning the models by using Ukrainian data sets with ground
truth obtained through the usage of the Dropla platform, open
collaboration, and data-sharing initiatives will allow us to solve the
problems that stand in the way of “All-in-one AI solution for UXO
detection”

- Lack of large, labeled datasets containing real-world
geophysical sensor data combined with known locations of UXO
threats buried in the soil.
- The limited and imperfect real-world datasets available lead to
challenges in developing robust AI models for UXO detection
that can perform effectively in the field.
- The datasets that do exist often contain class imbalances, with
far more examples of non-UXO clutter than actual UXO threats.
This can make it difficult to train models that are sensitive to
UXO anomalies.

There are differences in sensor types, soil conditions, and other
environmental factors between available datasets, which impacts
the ability of - AI models to generalize to new scenarios.

The limited and non-standardized datasets have hampered the
development of benchmarks and performance metrics to properly
evaluate and compare different UXO detection AI approaches

The main challenge here is that collecting large amounts of
real-world data with known ground truth is costly,
time-consuming, and technically challenging.
This is why sensing platform unification would play a vital role in
speeding up the development of Unified UXO detection AI.