ABSTRACT:

This project investigates the ability of GRIME AI and the embedded Segment Anything Model 2 (SAM2) to detect shade regions at a United States Geological Survey river site in Wisconsin, U.S.A. Our aim was to identify the deeply shadowed portions of water and snow under the bridge over time as shadows moved. We sparsely annotated five images from February 2026 (the 11th through 23rd) with the GRIME AI SAGE module and using two classes: "shade" and "non-shade". The annotations were used to fine-tune and validate SAM2 in the "Deep Learning" toolkit in GRIME AI. The toolkit was also used to segment images from all days in February 2025 between the hours of 10 a.m. and 2 p.m. CST. Output "panel" files that include the raw image, overlay mask, binary mask, and probability heatmap were used to create a GIF and video of the predicted shade regions over the month. We found that with the small training set and sparse annotations, the model only detects the darkest shade areas and fails to detect shade areas closer to the shade/non-shade boundary. Further refinement (more training data and richer annotations) are necessary to improve the segmentation performance.