Mountains of Confusion: Evaluating Image Enhancement to Improve AI Landscape Classification
Date
2024
Authors
Bron, Larissa
Journal Title
Journal ISSN
Volume Title
Publisher
University of Victoria
Abstract
With over 120,000 historic images of the Canadian Rockies, the Mountain Legacy Project (MLP) created the Python Landscape Classification tool (PyLC) to expedite the process of pixel-level identification of ecosystems used in quantifying ecosystem change from repeat photography. PyLC is a set of semantic segmentation artificial intelligence models that have been trained to discern ecosystems in historical and repeat images of mountain landscapes. This research project coincides with multiple efforts towards improving PyLC to better generalize ecosystem mapping to the large geographic region the image collection captures.
The experimental approach was to trial image enhancements using Python functions in an attempt to improve PyLC’s differentiation of ecosystems by separately changing the contrast, noise, and sharpness of images. Following this, comparison of the median accuracy between trials revealed either no change or a decrease to accuracy, though each image enhancement did change how PyLC interpreted and mapped the ecosystems in the images. These results lend support toward the limited generalizability of PyLC to new images while providing a review of image enhancement methods that could be incorporated into improving training through data augmentation.
Description
Keywords
Canadian Rockies, mountains, ecosystem change, image enhancement, artificial intelligence, semantic segmentation