Adaptive RBF Interpolation for Estimating Missing Values in Geographical Data

Kaifeng Gao, Gang Mei, Salvatore Cuomo, Francesco Piccialli, Nengxiong Xu

Introduction

Geographical datasets, particularly Digital Elevation Models (DEMs), serve as the backbone for critical applications such as urban planning, hydrological modeling, and environmental monitoring. However, these datasets are frequently plagued by "holes"—missing values caused by sensor malfunctions, signal interference, or data corruption during transmission. To maintain the integrity of spatial analyses, researchers rely on interpolation techniques to estimate these missing values based on surrounding known data.

While traditional methods like $k$-Nearest Neighbors ($k$NN) and Inverse Distance Weighting (IDW) are computationally efficient, they often fail to capture the complex, non-linear variations inherent in natural terrain. Radial Basis Functions (RBFs) offer a more mathematically sophisticated approach but introduce a significant parameterization challenge: the "shape factor." A globally fixed shape factor rarely performs optimally across an entire map because geographical data density and terrain complexity vary significantly by region.

Methodology: The Adaptive RBF Approach

To address the limitations of static interpolation, a new Adaptive RBF Interpolation algorithm has been proposed. This method dynamically adjusts the RBF shape factor based on the local distribution of data points.

Flowchart of the Adaptive RBF Interpolation algorithm.

The algorithm follows a structured four-stage workflow:

  1. Local Windowing: To handle large-scale geospatial "big data," the algorithm avoids global calculations. Instead, it identifies the 20 nearest known points for each missing value using a $k$NN search.
  2. Density Analysis: It calculates the local density of these neighboring points and compares it to the expected density of the overall dataset.
  3. Fuzzy Logic for Shape Factors: The core innovation lies in using fuzzy membership functions to automatically determine an optimal shape factor ($c$) for the Multi-quadric RBF ($\phi(r) = \sqrt{r^2 + c^2}$). The algorithm selects a factor that "stretches" or "shrinks" the basis function to best fit the local point distribution.
  4. Imputation: Finally, it solves a localized linear system to estimate the missing elevation value with high precision.

Experimental Setup and Datasets

The performance of the Adaptive RBF algorithm was evaluated using high-resolution DEM data from three Chinese cities with diverse topographical profiles: Beijing, Chongqing, and Longyan.

Digital Elevation Models (DEM) of the three study areas: (a) Beijing, (b) Chongqing, and (c) Longyan.

The datasets were substantial, containing over one million points each. To simulate real-world data loss, the researchers randomly removed 10% of the known points from each dataset. These points were then treated as "missing," and the algorithm was tasked with reconstructing them. The accuracy of the reconstruction was measured using the Root Mean Square Error (RMSE) against the original ground-truth values.

Results and Discussion

The experimental results demonstrated a clear trade-off between interpolation accuracy and computational overhead.

Accuracy Performance

In all three city datasets, the Adaptive RBF method consistently outperformed both $k$NN and Adaptive Inverse Distance Weighted (AIDW) methods. By tailoring the shape factor to local density, the RBF was able to model terrain features more accurately than methods that rely on simpler distance-based weighting. The RMSE for the Adaptive RBF was significantly lower, indicating a more faithful reconstruction of the original topography.

Computational Efficiency

However, this superior accuracy comes at a computational cost. The Adaptive RBF method was found to be significantly slower than its counterparts. For example, in the Beijing dataset, the RBF method required over 3 seconds to process a data segment, whereas $k$NN and AIDW completed the same task in approximately 1 second. This is because the adaptive approach requires solving a unique system of equations and performing fuzzy logic calculations for every single missing point.

Conclusion

The Adaptive RBF interpolation method represents a significant advancement in geospatial data processing, offering a robust solution for filling gaps in complex terrain models. While its current computational demands make it less suitable for real-time processing of massive datasets, its accuracy makes it ideal for high-precision offline modeling. Future research will focus on optimizing the algorithm through parallel computing and GPU acceleration to bridge the gap between accuracy and efficiency.

Summary updated

BibTeX
@inproceedings{gao2019adaptive,
  author = {Kaifeng Gao and Gang Mei and Salvatore Cuomo and Francesco Piccialli and Nengxiong Xu},
  title = {{Adaptive RBF Interpolation for Estimating Missing Values in Geographical Data}},
  booktitle = {NUMTA},
  year = {2019},
  doi = {10.1007/978-3-030-39081-5\_12},
}
Figure