Curriculum learning-based strategy for low-density archaeological mound detection from historical maps in India and...

Area of Science:

• Archaeological survey
• Computer vision
• Historical geography

Background:

• Historical maps are crucial for reconstructing past landscapes altered by modern development.
• Many archaeological features, like mounds, are depicted on historical maps but often not recognized as such.
• Automated survey methods face challenges with low feature density and scarce training data.

Purpose of the Study:

• To develop and evaluate two algorithms for large-scale automatic detection and instance segmentation of potential archaeological mounds on historical maps.
• To address the limitations of low archaeological feature density and limited training data in automated surveys.
• To improve the accuracy and efficiency of identifying archaeological settlements from historical map data.

Main Methods:

• Designed algorithms to detect mound features depicted through hachures and form-lines on historical maps.
• Applied a Curriculum Learning strategy with synthetic data to enhance feature recognition.
• Incorporated filters for topographic setting, form, and size to refine detection accuracy.
• Tested algorithms on the historic 1″ to 1-mile map series, covering 470,500 km².

Main Results:

• Achieved recall of 52.61% and precision of 82.31% for hachure mounds.
• Achieved recall of 70.80% and precision of 70.29% for form-line mounds.
• Successfully detected nearly 6000 mound features, representing the largest-scale application of such an approach.
• Demonstrated recall >60% and precision >90% on maps similar to training data.

Conclusions:

• The developed algorithms offer a robust solution for large-scale archaeological mound detection on historical maps.
• The approach effectively handles challenges of low feature density and limited training data.
• The adaptive potential allows for improved detection with minimal retraining on new map data.