Abstract: Regional climate is a critical factor in public health research, adaptation studies, climate change burden analysis, and decision support frameworks. Existing climate regionalization schemes are not well suited for these tasks as they rarely take population density into account. In this work, we are extending our recently developed method for automated climate regionalization (LKN-method) to incorporate the spatial features of target population. The LKN method consists of the data limiting step (L-step) to reduce dimensionality by applying principal component analysis, a classification step (K-step) to produce hierarchical candidate regions using k-means unsupervised classification algorithm, and a nomination step (N-step) to determine the number of candidate climate regions using cluster validity indexes. LKN method uses a comprehensive set of multiple satellite data streams, arranged as time series, and allows us to define homogeneous climate regions. The proposed approach extends the LKN method to include regularization terms reflecting the spatial distribution of target population. Such tailoring allows us to determine the optimal number and spatial distribution of climate regions and thus, to ensure more uniform population coverage across selected climate categories. We demonstrate how the extended LKN method produces climate regionalization can be better tailored to epidemiological research in the context of decision support framework.
Introduction
Climate regionalization—the process of grouping geographic areas with similar climatic conditions—is essential for understanding how environmental factors impact human health. However, traditional systems like the Köppen-Geiger (KG) classification often fail to capture the complex, high-resolution patterns available through modern satellite data. Furthermore, standard regionalization is often "population blind," ignoring where people actually live. For an epidemiologist, a climate region covering a vast, uninhabited desert is less useful than a region covering a densely populated urban center. If one climate zone contains the vast majority of a country's population while others are nearly empty, it becomes difficult to perform balanced statistical comparisons of disease burden across those zones.
Methodology: The Regularized LKN Approach
The researchers build upon the LKN method, which consists of three main steps:
- L-step (Data Limiting): Uses Principal Component Analysis (PCA) to reduce the dimensionality of massive satellite time-series datasets while retaining the most important variance.
- K-step (Classification): Employs the k-means unsupervised clustering algorithm to group pixels into hierarchical candidate regions.
- N-step (Nomination): Uses cluster validity indexes (like Calinski-Harabasz) to determine the "optimal" number of regions ($k$).
The Innovation: The researchers added a spatial regularization term to the validation criteria. This term acts as a penalty that discourages choosing too few regions if those regions result in highly uneven population distributions. This mathematical "nudge" pushes the algorithm toward a number of clusters that better serves epidemiological research by balancing climate homogeneity with uniform population distribution.
Data Sources
The study utilized two primary data sources to test their method in Ghana, a region chosen for its diverse climate and significant variability in population density:
- MODIS NDVI (Normalized Difference Vegetation Index): 15 years of satellite data (2000–2013) from NASA’s Terra and Aqua satellites. NDVI measures "greenness" and serves as a high-resolution proxy for phenology and climate, effectively integrating temperature and moisture conditions over time.
- WorldPop: A high-resolution age-structured population distribution dataset used to calculate the demographic weight of each proposed climate region.
Results and Discussion
The researchers compared the standard LKN method against their regularized version using Ghana as a case study. Purely mathematical clustering (without regularization) suggested that Ghana should be divided into only 3 climate regions. While mathematically "clean," this was practically useless for health research because over 56% of the population was concentrated in a single region, resulting in a standard deviation of population counts across regions of 7.8 million.
By adjusting the regularization constant, the modified criteria suggested that 8 or 15 regions were superior for this application. At $k=15$, the population distribution became significantly more uniform, and the standard deviation of population counts dropped sharply to 1.4 million. This more even "weighting" of regions allows for more reliable comparisons of health outcomes across different climates.
Conclusion
The extended LKN method provides a robust framework for tailoring climate regions to the specific needs of epidemiological research. While the choice of regularization constants currently requires manual tuning, the approach ensures that climate-based regions are statistically meaningful for public health analysis. Future work will focus on automating hyper-parameter selection and scaling this methodology to larger geographic areas, such as the Continental United States, to study the effects of extreme weather on vulnerable populations.
Summary updated