Geospatial Diffusion for Land Cover Imperviousness Change Forecasting

Debvrat Varshney, Vibhas Vats, Bhartendu Pandey, Christa Brelsford, Philipe DiasSigspatial/gis 2025 · DOI

TL;DR

This paper introduces a novel approach to forecasting land cover imperviousness changes using generative AI, specifically diffusion models. By conditioning these models on historical land cover and imperviousness data, they achieve sub-kilometer accuracy in predicting future changes, outperforming traditional methods and a simple "no change" baseline.

What Problem Does This Paper Address?

Accurately forecasting land cover change, particularly the increase in impervious surfaces (like roads and buildings), is crucial for understanding and mitigating environmental impacts such as urban heat islands and increased flood risk. While Earth system models are improving, their input of land-use and land-cover (LULC) change forecasts has lagged. Traditional LULC forecasting methods often rely on simplified assumptions about spatial connectivity and require manual parameterization, limiting their ability to capture complex, human-driven changes. This paper aims to bridge this gap by leveraging the power of generative AI to synthesize more accurate and nuanced LULC change forecasts.

Methods

The core of this work is a geospatial diffusion model adapted for imperviousness forecasting. Diffusion models are a class of generative AI that learn to synthesize data by gradually reversing a process of adding noise. Here, the model is trained to predict the noise added to an input image, effectively learning to reconstruct the original data.

Key methodological contributions include:

  • Conditional Diffusion: The diffusion model is conditioned on historical imperviousness maps and LULC maps from previous years. This allows the model to learn from past patterns and trends.
  • Spatially-Adaptive Normalization (SPADE): Inspired by previous work, they adapt SPADE to integrate multiple conditioning inputs (past imperviousness and LULC) into the diffusion process, allowing for more sophisticated guidance of the synthesis.
  • LULC-based Imperviousness Likelihood Maps: They create maps that estimate the probability of LULC transitions leading to imperviousness. This is derived from historical NLCD data and used as an additional conditioning input.
  • Addressing Data Imbalance: Recognizing that most of the US has little land cover change, they use temporal clustering (via Dynamic Time Warping) and a reverse weighting scheme to ensure the model learns from both high-change and low-change areas, preventing it from simply predicting persistence.

Datasets

The primary dataset used is the National Land Cover Database (NLCD) for the conterminous United States (CONUS). This dataset provides raster land cover and imperviousness maps at 30m/px resolution across multiple epochs from 2001 to 2021. For training, data from 2001-2019 was used, with 2021 held out for testing. The evaluation was conducted across 12 diverse Metropolitan Statistical Areas (MSAs).

Key Insights and Ideas

  • Generative AI for Geospatial Forecasting: The paper demonstrates that generative AI, specifically diffusion models, can be effectively applied to complex geospatial forecasting tasks like land cover change, moving beyond traditional rule-based or statistical models.
  • Capturing Complex Spatiotemporal Patterns: Diffusion models can learn intricate, non-linear relationships and multi-temporal associations in the data, which are crucial for understanding human-driven land cover changes that are not strictly spatially contiguous.
  • Data Imbalance is a Major Hurdle: The significant imbalance between areas with high and low land cover change poses a substantial challenge. Simply training on raw data leads to models that favor persistence. Addressing this imbalance through clustering and weighting is critical for meaningful forecasts.
  • Sub-kilometer Accuracy is Achievable: The proposed model achieves a "null resolution" (the point where its accuracy matches a "no change" baseline) at sub-kilometer scales for many areas, indicating a significant improvement over simpler models.

Key Results

  • The GEODIFF-NATIONAL model, trained on data across the CONUS, achieved a null resolution of approximately 0.7km x 0.7km across 12 test areas.
  • For high-change regions (e.g., Las Vegas, Denver, Miami), the model achieved very fine null resolutions, as low as 0.2km x 0.2km.
  • The diffusion model outperformed a CA-Markov baseline in a qualitative case study, generating more plausible and spatially coherent clusters of change rather than just relying on strict spatial connectivity.
  • The model demonstrated a tendency to be more sensitive to areas undergoing rapid change, achieving better performance in high-change clusters.

Limitations and Future Work

  • Assumption of Past Patterns: The model's forecasts are based on historical trends. Future land cover change can be influenced by new drivers (e.g., policy changes, economic shifts) not captured in past data, which could lead to deviations.
  • Categorical LULC Forecasting: This work focused on imperviousness (a continuous variable). Extending the framework to forecast discrete LULC categories is a significant area for future research.
  • Incorporating More Drivers: The authors suggest incorporating additional auxiliary information like population dynamics, land use policies, and terrain data to further constrain and improve forecasts.
  • Evaluation Metrics: Developing more sophisticated evaluation metrics that capture spatial heterogeneity and consistency would be beneficial.

Summary updated

BibTeX
@inproceedings{varshney2025geospatial,
  author = {Debvrat Varshney and Vibhas Vats and Bhartendu Pandey and Christa Brelsford and Philipe Dias},
  title = {{Geospatial Diffusion for Land Cover Imperviousness Change Forecasting}},
  booktitle = {Sigspatial/gis},
  year = {2025},
  doi = {10.1145/3748636.3762730},
}
Figure