Detecting Landcover from satellite images using AI and Ellmer
Author:
Jean Marie Ngabonziza ’26Co-Authors:
Odilon Ligan'27, Grayson White, Kelly McConville, Andrew ListerFaculty Mentor(s):
Kelly McConville – Director, Dominguez Centre for Data ScienceFunding Source:
USDA, Forest Service, Rocky Mountain Research StationAbstract
The Forest Inventory and Analysis (FIA) program plays a critical role in monitoring the health and status of forest ecosystems across the United States. As the demand for timely, accurate, and scalable forest data increases, FIA is exploring innovative ways to integrate Artificial Intelligence (AI) into its workflow to enhance efficiency and precision.
The objective of this project is to explore how Large Language Models (LLMs) might assist in processing satellite images for land cover classification and landscape change detection. A key component of this effort is not only to test new AI tools but also to rigorously quantify their performance using spatial data and ground-truth comparisons.
After exploring many AI tools, we opted for Elmer, an R package that allows a user to interact with LLMs directly in our R environment. We developed a workflow for loading images, classifying them with an LLM, and then comparing those classifications to the National Land Cover Database’s (NLCD) classifications. Though AI was not perfect, our results show that it was able to detect some land cover correctly as described by NLCD. However, there is more work to be done. In the future, we envision improving accuracy by revising prompts given to AI to describe in detail the sub-classifications and using a more sophisticated generative AI to increase accuracy.