Is it time for the human data analyst to retire in the era of artificial intelligence (AI)? A critical examination in scientific data analysis.

On the way to robotic exploration on the Moon AI is going to play an increasingly important role in geomorphological studies. With the fast-growing amount of available data to be processed a direct human analysis is becoming steadily more difficult to achieves. Machine learning is a branch of AI and computer science which through the use of statistical methods utilize data and algorithms to simulate human learning behavior, characterized by the ability to automatically improve accuracy through experience. The success of these methods together with the readily available machine learning codes has resulted in an increasing deployment of this approach in all fields of geoscience, in particular for image recognition and classification tasks.  Independent verification of the gained results from these studies is often difficult to achieve, partly due to the enormous amount of data to be handled, partly because the AI methods have features that make them hard to check (Szegedy et al., 2014). The effect of data bias is well known (Torralba et al., 2011) but other factors can also play a role. To exemplify and investigate the question of the trustworthiness of AI-based results in remote sensing we confront results from a recent AI-driven research study in planetary geomorphological science which analyzed a staggering data set of high-resolution lunar images with the help of Convolutional Neural Networks (CNNs) to construct a global lunar boulder map (Bickel et al. 2020) with a human-based analysis approach. We show which factors are crucial when preparing such studies and discuss the implications.

 

References:

Bickel et al., 2020, Impacts drive lunar rockfalls over billions of years, Nature Communications, https://doi.org/10.1038/s41467-020-16653-3

Szegedy et al., 2014, https://doi.org/10.48550/arXiv.1312.6199

Torralba et al., 2011, Unbiased Look at Dataset Bias, IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2011