Modern Neural Networks for Small Tabular Datasets: The New Default for Field-Scale Digital Soil Mapping?

TL;DR

A comprehensive benchmark that evaluates state-of-the-art ANN architectures, including the latest multilayer perceptron (MLP)-based models (TabM, RealMLP), attention-based transformer variants (FT-Transformer, ExcelFormer), retrieval-augmented approaches (TabR, ModernNCA), and an in-context learning foundation model (TabPFN), reveals that modern ANNs consistently outperform classical methods on the majority of tasks.

Abstract

In the field of pedometrics, tabular machine learning is the predominant method for soil property prediction from remote and proximal soil sensing data, forming a central component of Digital Soil Mapping (DSM). At the field-scale, this predictive soil modeling (PSM) task is typically constrained by small training sample sizes and high feature-to-sample ratios in soil spectroscopy. Traditionally, these conditions have proven challenging for conventional deep learning methods. Classical machine learning algorithms, particularly tree-based models like Random Forest and linear models such as Partial Least Squares Regression, have long been the default choice for pedometric modeling within DSM. Recent advances in artificial neural networks (ANN) for tabular data challenge this view, yet their suitability for field-scale DSM has not been proven. We introduce a comprehensive benchmark that evaluates state-of-the-art ANN architectures, including the latest multilayer perceptron (MLP)-based models (TabM, RealMLP), attention-based transformer variants (FT-Transformer, ExcelFormer, T2G-Former, AMFormer), retrieval-augmented approaches (TabR, ModernNCA), and an in-context learning foundation model (TabPFN). Our evaluation encompasses 31 field- and farm-scale datasets containing 30-460 soil samples and three critical soil properties: soil organic matter or soil organic carbon, pH, and clay content. Our results reveal that modern ANNs consistently outperform classical methods on the majority of tasks, demonstrating that deep learning has matured sufficiently to overcome the long-standing dominance of classical machine learning in pedometrics. Notably, TabPFN delivers the strongest overall performance, showing robustness across varying conditions. We therefore recommend the adoption of modern ANNs for field-scale DSM and propose TabPFN as the new default choice in the toolkit of every pedometrician.

Figures (8)

• Figure 1: Global distribution of the 31 field- and farm-scale datasets used in this study. Background map is based on Homolosine projection. The datasets span multiple continents and diverse agricultural contexts, including locations across North and South America, Europe, Asia, and Australia. Base map source: Natural Earth.
• Figure 2: Distribution of datasets by number of features and soil samples (log scale), highlighting feature-to-sample ratios. Each point represents one dataset, with two groups shown: Low-Dimensional datasets (lower feature counts) and High-Dimensional datasets (higher feature counts due to spectral features).
• Figure 3: Performance comparison of ANN architectures versus classical ML baselines on DSM tasks. Models are grouped by architectural approach: MLP-based (MLP, TabM, RealMLP), Attention-based (AutoInt, T2G-Former, ExcelFormer), Retrieval-based (TabR, ModernNCA), and In-Context Learning (TabPFN). Lower rank values indicate better performance. Error bars show $\pm 1$ standard error of the mean of each model's rank across the datasets. The in-context learning model TabPFN consistently achieved the best performance across both Low-Dimensional and High-Dimensional datasets.
• Figure 4: Head-to-head comparison of ANN architectures against the best-performing classical ML baseline for each dataset group. For Low-Dimensional datasets, models are compared against Random Forest; for High-Dimensional datasets, against Ridge Regression. A higher number of wins indicates superior performance. Modern deep learning approaches, particularly TabPFN, substantially outperformed classical baselines on the majority of tasks.
• Figure 5: Model performance rankings grouped by dataset size, demonstrating the relationship between sample size and model effectiveness. Rankings are computed at the architectural group level using a two-stage process: first identifying the best model within each group, then ranking groups by their best model's performance. In-context learning (TabPFN) maintained superior performance across nearly all dataset sizes, with classical ML only prevailing on the smallest High-Dimensional datasets with PCA preprocessing.