Within the broader computer vision community, the issue of dataset size has received surprisingly little attention. Most analyses simply use all available data and focus on model architecture, with scant attention given to whether the dataset size is appropriate for the task and architecture’s complexity.

Many different variables determine the ultimate mission impact of satellite imagery, a concept CosmiQ has referred to as the Satellite Utility Manifold. Previous CosmiQ studies have explored such variables as sensor resolution (0.3 meter to 2.4 meter), super-resolution techniques, and the number of imaging bands (grayscale versus multispectral).

Expanding on this work, the Machine Learning Robustness Study focuses on training dataset size and diversity on building detection performance in the SpaceNet data. The recent availability of this extensive dataset and model-building capability will make it possible to address dependence on geography and dataset size at the leading edge of geospatial machine learning.

At the inception of this research, the interplay between super-resolution techniques and object detection frameworks remained largely unexplored, particularly in the context of satellite or overhead imagery. Intuitively, super-resolution methods should increase object detection performance, as an increase in resolution should add more distinguishable features that an object detection algorithm can use for discrimination.

This trade study strived to answer these foundational questions:

• Does the application of a super-resolution (SR) technique affect the ability to detect small objects in satellite imagery?
• Across what resolutions are these SR techniques effective?
• What is an ideal or minimum viable resolution for object detection?
• Can one artificially double or even quadruple the native resolution of coarser imagery to make the data more useful and increase the ability to detect fine objects?

Our results showed that the application of SR techniques as a pre-processing step provided a statistically significant improvement in object detection performance in the finest resolutions.  For both object detection frameworks, the greatest benefit is achieved at the highest resolutions, as super-resolving native 30 cm imagery to 15 cm yields a 13−36% improvement in mean average precision. Furthermore, when using YOLT, we find that enhancing imagery from 60 cm to 15 cm provides a significant boost in performance over both the native 30 cm imagery (+13%) and native 60 cm imagery (+20%).