Abstract
Spatial soil sampling is an integral part of a soil survey aimed at creating a soil map. We propose considering the soil sampling procedure as a task of optimal design. In practical terms, optimal experiments can reduce experimentation costs, as they allow the researcher to obtain one optimal set of points. We present a sampling design, based on the fundamental idea of selecting sample locations by performing an optimal design method called the maxvol algorithm. It is shown that the maxvol-base algorithm has a high potential for practical usage. Our method outperforms popular sampling methods in soil taxa prediction based on topographical features of the site and deals with massive agricultural datasets in a reasonable time.
Introduction
Creating accurate soil maps is a cornerstone of precision agriculture and environmental management. However, the physical collection of soil samples remains a significant bottleneck due to high labor and laboratory costs. The primary challenge lies in identifying a "Goldilocks" set of sampling points: a collection large enough to capture the complex variability of a landscape, yet small enough to remain economically viable.
Traditional methods, such as Conditioned Latin Hypercube Sampling (cLHS), typically rely on probabilistic distributions. While effective, these methods can be computationally intensive on large datasets and often require numerous iterations to converge on a suitable layout. This research proposes a shift in perspective, treating spatial sampling as an optimal design problem. By utilizing the maxvol algorithm—a tool from linear algebra and machine learning—researchers can identify the most informative locations by treating them as the most significant rows in a massive matrix of landscape features.
Methodology: The Maxvol Approach
The core of this methodology is the application of the maxvol algorithm (and its rectangular variant, rect_maxvol) to spatial data. The process follows a structured pipeline:
- Feature Extraction: Using a Digital Elevation Model (DEM), topographical features such as slope, aspect, and topographic wetness index are calculated for every pixel using GIS software.
- Matrix Construction: These spatial features are flattened into a large information matrix $A$ of size $m \times n$, where $m$ represents the number of pixels and $n$ the number of features.
- Maxvol Optimization: The algorithm searches for a submatrix $\hat{A}$ with the maximum volume (D-optimality criterion). Geometrically, this selects points that are as dissimilar as possible in the feature space, ensuring the samples represent the full diversity of the terrain.
- Spatial Constraints: To ensure practical utility for human surveyors, a heuristic distance constraint is applied to prevent the selection of adjacent, redundant pixels.
Experimental Evaluation
The method was rigorously tested across three distinct sites in European Russia to evaluate its versatility and scalability:
- Kursk Site: A 0.4 km² virgin forest-steppe reserve used as a primary benchmark for soil taxa prediction accuracy.
- Kshen Site: A 1.21 km² cultivated field where yield maps served as a proxy for soil variability.
- Ozery Site: A massive dataset comprising seven fields and 32 million pixels, testing the algorithm's performance in a "cold start" scenario with only DEM data available.
Key Results and Discussion
Accuracy and Precision
On the Kursk site, the maxvol algorithm consistently outperformed cLHS, Random Sampling, and Kennard-Stone methods in balanced accuracy for soil taxa prediction. This advantage was most pronounced with small sample sizes (7–12 points), making it an ideal choice for low-budget or preliminary surveys. Crucially, maxvol successfully captured "small spots"—rare soil types existing in tiny topographical pockets—that probabilistic methods often missed.
Computational Efficiency
Unlike randomized simulations, maxvol is a deterministic, greedy iterative algorithm. This allows it to scale exceptionally well; it processed the 32-million-pixel Ozery dataset in approximately 6 minutes on a single CPU core. This speed enables researchers to handle massive agricultural datasets that would be computationally prohibitive for traditional iterative statistical methods.
Practical Advantages
Because the algorithm is deterministic, it provides a single, optimized sampling scheme. This eliminates the variability inherent in randomized methods, offering a stable and reliable plan for field expeditions.
Conclusion
The maxvol-based sampling design offers a robust, efficient, and highly accurate alternative to traditional spatial sampling methods. By framing sampling as a D-optimality problem, it ensures maximum information gain from a minimum number of physical samples. Future work will focus on automating the distance constraint parameters and integrating more diverse predictor variables beyond topography to further enhance the algorithm's predictive power.
Summary updated