{"id":1892,"date":"2025-01-07T18:30:26","date_gmt":"2025-01-07T23:30:26","guid":{"rendered":"https:\/\/www.clarku.edu\/geospatial-analytics\/land-cover-change-modeling-using-genai-collaboration-with-washu\/"},"modified":"2025-12-15T12:06:53","modified_gmt":"2025-12-15T17:06:53","slug":"land-cover-change-modeling-using-genai-collaboration-with-washu","status":"publish","type":"post","link":"https:\/\/www.clarku.edu\/geospatial-analytics\/uncategorized\/land-cover-change-modeling-using-genai-collaboration-with-washu\/","title":{"rendered":"Land Cover Change Modeling using GenAI (Collaboration with WashU)"},"content":{"rendered":"\n
\"\"<\/figure>\n\n\n\n


This project leverages generative AI models to project future land cover changes using historical satellite imagery and land cover maps. By exploring three novel AI modeling approaches, we aim to enhance the accuracy in predicting environmental transformations. The project addresses critical geospatial challenges by supporting sustainable land-use planning and climate resilience strategies. Anticipated outcomes include scalable methodologies for integrating geospatial data and generative AI, fostering cross-sectoral applications in conservation, urban development, and policymaking, ultimately advancing the understanding and management of dynamic landscapes.<\/p>\n\n\n\n

Background<\/h3>\n\n\n\n

Traditionally, land cover (LC) change prediction and forecasting has been modeled using Markov chains and cellular automata models at pixel-level. These approaches rely heavily on historical transition matrices to capture the likelihood of changes between land cover classes over time and often simplify spatio-temporal complexities. However, the advancements in deep learning and generative AI (GenAI) provide a valuable opportunity to build more sophisticated approaches to capture the spatio-temporal complexity of these transitions.<\/p>\n\n\n\n

In this project, we are designing and training novel GenAI models that can synthesize future satellite images and forecast future LC maps. These models learn from historical satellite imagery, and their corresponding LC maps as well as auxiliary variables (e.g. climate data and socio-economic variables) and be able to predict the future imagery and LC map.<\/p>\n\n\n\n

Impact<\/h3>\n\n\n\n

This project advances research in geospatial AI by pioneering the use of GenAI for spatio-temporal modeling. By adapting GenAI models for geospatial data, primarily multispectral satellite imagery, this project advances frameworks and datasets for interdisciplinary research in the area of land change modeling. The output of this project provides tools for land-use planning, risk assessment, and environmental monitoring, enabling data-driven decisions in sectors like carbon credit market, agriculture, real estate, and urban development.<\/p>\n\n\n\n

Moreover, by providing realistic future LC maps, this project supports UN Sustainable Development Goals (SDGs) by facilitating climate resilience, biodiversity conservation, and resource management. It also empowers policymakers at local and national levels with actionable insights to address environmental and urban challenges.<\/p>\n\n\n\n

Methodology<\/h3>\n\n\n\n

This project leverages a rich combination of satellite imagery, LC maps, and auxiliary variables to predict future LC changes and synthesize satellite images. Three different GenAI models are designed to predict future satellite images and LC maps concurrently, leveraging a shared multi-task framework to enhance both outputs. This approach combines multi-task learning with robust error handling and guided refinement to produce realistic satellite images and accurate LC predictions.<\/p>\n\n\n\n

The three model architectures are based on Generative Adversarial Networks (GAN), masked auto-encoder (MAE), and diffusion. The satellite image is used as a condition in the training and generation phase to ensure the model predicts realistic LC maps. In addition, various error mitigation strategies are being explored to address noisy LC labels and to mitigate overfitting to noisy labels, ensuring more robust predictions.<\/p>\n\n\n\n

Clark CGA is kickstarting this project in winter 2025 using an award from Taylor Geospatial Institute and AWS.<\/p>\n\n\n\n