A Machine Learning Approach to Cave and Sinkhole Prediction in the Cradle of Humankind: Implications for Fossil Prospecting in Paleoanthropology

Author

Margaret J. Furtner, Louisiana State University and Agricultural and Mechanical CollegeFollow

Degree

Doctor of Philosophy (PhD)

Department

Geography and Anthropology

Document Type

Dissertation

Abstract

In paleoanthropology, fossils are key to understanding hominin evolutionary relationships and paleoenvironments. Surveying an area for new fossil sites, however, is a labor-intensive and resource-draining activity. Machine learning prediction models trained on remotely sensed landscape imagery can be employed to alleviate difficulties associated with traditional survey practices by identifying areas of interest for fossil prospecting. The purpose of this dissertation is to compare and apply different machine learning frameworks that predict the location of cave entrances and sinkholes as a proxy for fossil sites in the Cradle of Humankind, South Africa. Machine learning models were trained using geomorphological landscape characteristics derived from multispectral satellite images, Digital Elevation Models (DEMs), and geologic maps in association with known cave and sinkhole localities in the region, in order to identify other areas with a similar suite of characteristics for survey.

This dissertation is divided into three main studies. The first utilizes a spatial 10-fold cross-validation method to evaluate Random Forest (RF) model efficacy on test datasets withheld from model training. The RF performed with an average 81.6% accuracy and an Area Under the Curve (AUC) score of 0.912. Variable permutation importance found that fault proximity, location within the Chuniespoort geologic Group, dolomite presence, chert presence, and elevation exhibited the highest importances for model accuracy. The second study compares five different machine learning frameworks for Cradle site prediction: RF, Support Vector Machine (SVM), Maximum Entropy (MaxEnt), 1D Convolutional Neural Network (CNN-1D), and 2D Convolutional Neural Network (CNN-2D). The CNN-1D outclasses the other models in average accuracy, recall, and F1 score but is prone to false positives. Conversely, MaxEnt’s high precision and low recall indicates it is cautious and prone to false negatives. The third study applies the CNN-1D and MaxEnt model in an ensemble approach to specific areas in and around the Cradle for ground-truthing surveys. Sixty-nine cave and sinkhole sites were recorded within or close to the boundaries of model survey zones across approximately 184 hectares. These results demonstrate the models’ ability to efficiently and successfully identify areas of interest for survey, which could greatly expedite future paleoanthropological prospecting efforts.

Date

7-20-2025

Recommended Citation

Furtner, Margaret J., "A Machine Learning Approach to Cave and Sinkhole Prediction in the Cradle of Humankind: Implications for Fossil Prospecting in Paleoanthropology" (2025). LSU Doctoral Dissertations. 6867.
https://repository.lsu.edu/gradschool_dissertations/6867

Committee Chair

Brophy, Juliet K

DOI

10.31390/gradschool_dissertations.6867