Spatiotemporal analysis of rural-urban land conversion

Bo Huang, Li Zhang, Bo WuInt. J. Geogr. Inf. Sci 2009

TL;DR

This paper introduces a novel multi-temporal spatial regression model to better understand and predict rural-urban land conversion. By integrating logistic regression with spatial sampling and exponential smoothing, the model effectively captures both spatial and temporal complexities of urban expansion, offering a more accurate and interpretable approach than previous methods.

What Problem Does This Paper Address?

Urban expansion is a complex phenomenon driven by numerous social, economic, and environmental factors, leading to significant land-use changes. While existing models can predict spatial patterns, they often struggle to account for the temporal dynamics of this conversion. This gap makes it difficult to formulate effective policies for urban development. The authors aim to address this by developing a model that can simultaneously analyze and predict land-use change across multiple time periods, considering both spatial relationships and temporal trends.

Methods

The core of the study is a multi-temporal spatial regression model. This model builds upon logistic regression, a standard technique for modeling binary outcomes (in this case, land conversion or no conversion). To tackle the spatial dependencies inherent in geographic data, the authors employ a spatial sampling scheme. This involves a non-overlapping moving window approach to reduce spatial autocorrelation in the sampled data, aiming to achieve similar accuracy to more complex spatial autoregressive models but with a more manageable process.

To integrate data from multiple time periods, the study utilizes exponential smoothing. This technique, adapted for land-use change, allows for the combination of bi-temporal models into a single prognostic model. It assigns exponentially decreasing weights to older observations, giving more importance to recent trends. The smoothing is applied at the "utility function level" of the regression models, acknowledging the non-linear nature of the underlying land conversion process.

Datasets

The study uses multi-temporal land-use data for New Castle County, Delaware. The data spans three periods: 1984-1992, 1992-1997, and 1997-2002. Land-use data was derived from digital orthophotos and rasterized at a 50m x 50m resolution. Demographic data (population density) for 1984, 1992, and 1997 were estimated using interpolation from 1990 and 2000 data. Transportation network data from 2001 was used as a proxy for earlier periods, with the assumption that major network changes were negligible.

Key Insights and Ideas

  • Spatial sampling is a viable alternative to complex spatial models: The study demonstrates that a well-designed spatial sampling strategy can effectively mitigate spatial autocorrelation, achieving comparable accuracy to more computationally intensive spatial autoregressive models. This offers a more practical approach for large datasets.
  • Temporal dynamics are crucial for accurate prediction: Simply analyzing bi-temporal changes misses the nuances of urban expansion over longer periods. Exponential smoothing helps to capture these temporal trends and create a more robust predictive model.
  • Urban expansion exhibits both agglomerated and dispersed patterns: The results show that while urban development tends to occur near existing urban areas and commercial sites (agglomeration), it also spreads outwards, influenced by factors like transportation infrastructure (dispersion).
  • Predictor variable significance evolves over time: The influence of factors like population density and proximity to commercial sites on land conversion changes across different time periods, highlighting the dynamic nature of urban growth.

Key Results

  • The proposed multi-temporal spatial regression model achieved overall correctness percentages (PCP) of 75.62% (1984-1992), 71.13% (1992-1997), and 68.52% (1997-2002).
  • The model's accuracy was comparable to a Spatial AutoLogit (SAL) model, which explicitly accounts for spatial autocorrelation, with overall PCPs of 76.20%, 72.58%, and 70.03% for the respective periods.
  • The exponential smoothing approach resulted in a higher overall PCP (70.36%) compared to a direct bi-temporal model (68.87%) for the entire 1984-2002 period.
  • Key predictors like population density and urbanized neighborhood were consistently significant drivers of land conversion across all periods.
  • The study found that slope was not statistically significant for land conversion across all periods and was removed from the regression model.

Limitations and Future Work

  • Data constraints: The study acknowledges limitations due to data scantiness, which prevented the incorporation of all potential variables. This necessitated simplifying the model and assuming socioeconomic effects were reflected in spatial factors, potentially leading to approximations.
  • Spatial inconsistency and estimation issues: Problems with spatial inconsistency, estimation (e.g., time extrapolation of population densities), and potential positional errors in older data were noted.
  • Generalizability: Modeling based on a single case study (New Castle County) may have limitations, as urbanization dynamics can vary significantly across different socioeconomic contexts.
  • Future work: The authors suggest that future research should focus on modeling land-use transformation by integrating time dependency and key dimensions of urban spatial expansion. Further investigation into more efficient methods for deriving smoothing models, beyond enumerating alpha values, is also recommended.

Summary updated

BibTeX
@article{huang2009spatiotemporal,
  author = {Bo Huang and Li Zhang and Bo Wu},
  title = {{Spatiotemporal analysis of rural-urban land conversion}},
  journal = {Int. J. Geogr. Inf. Sci.},
  year = {2009},
}
Figure