AI maps the Andean enclosures that generations have kept changing

Drone photographs around Cusi Cusi, in Argentina’s Andes, revealed enclosures enlarged or merged since older satellite images. A July study combines machine learning and fieldwork across 1,750 square kilometres of this landscape, which Indigenous communities continue to change.[1]

Recognising an enclosure

The detector uses Mask R-CNN, an architecture introduced in 2017. It first proposes regions likely to contain objects. Separate branches then classify the object, refine its bounding box and predict a mask marking its pixels. Keeping those jobs separate lets the system represent both an object’s category and its outline.[2]

Instance segmentation also distinguishes individual objects that share a category. A scene can contain several separately outlined objects, rather than one undifferentiated area labelled with their common name. The model’s training examples determine which categories it learns to recognise.[2]

The original architecture also addresses a boundary problem. Rounding an object’s proposed box onto a coarse feature grid can displace the shape being extracted. Its RoIAlign operation preserves alignment by interpolating between grid locations. Finding an object and tracing its edge place different demands on the representation.[2]

Building the corrected map

The satellite model generated 572 candidates. Review removed 182 false positives and added 162 missed enclosures, producing 552 mapped features. Natural rock formations sometimes resembled walls.[1]

On geographically separate test data, 43% of detections matched enclosures and 77% of known enclosures were found. The selected 0.90 score threshold was not a measured 90% success rate.[1]

Drone imagery supplied finer detail; walking surveys helped interpret use and chronology. Only part of the region was surveyed on foot, so the 552 features were not all individually ground-checked or newly dated.[1]

Report an error or suggest a correction