Digital Soil Mapping of the Steppe Zone in Northern Kazakhstan: Predicting Agrochemical Properties of Soils Using Multimodal Satellite Data and Machine and Deep Learning Techniques
Aliya Yskak, Gulnaz T. Yermoldina, Almabek B. Nugmanov, Berik S. Rakhimbayev, Zhanna B. Suimenbayeva, Vladimir D. Fominov, Zhassulan B. Irzhanov, Tatiana A. Paramonova, Sergey V. Mamikhin, Aleksandr G. Bulaev
Digital soil mapping (DSM), based on multimodal satellite data, is a crucial tool for the transition to precision agriculture. However, systematic studies using this method and machine and deep learning techniques are lacking for the arid and semi-arid regions of Central Asia, where multimodal satellite data can provide valuable insights into soil conditions. This work provides, for the first time, benchmark metrics for the predictive ability of six soil agrochemical properties (pH, Soil Organic Carbon, NO3, P2O5, K2O, and S) in the dry steppe zone of Central Asia, with a quantitative assessment of the difference between “standard” and “fair” validation strategies. This has methodological significance for the entire field of DSM research. A comprehensive comparison of 11 machine learning (ML) models and four deep learning (DL) architectures was conducted to predict soil agrochemical properties using a set of 530 features extracted from various satellite datasets. These features were extracted from Sentinel-2, Landsat-8, Sentinel-1 SAR, SRTM DEM, and ERA 5-Land using Google Earth Engine (GEE) automated pipeline. All models were evaluated using three spatial validation strategies with increasing stringency: Leave-One-Field-Out CV (LOFO-CV), Leave-One-Farm-Out CV (Farm-LOFO), and an optimized spatial split. We propose a three-level hierarchical validation scheme that allows for the quantitative separation of spatial leakage and feature selection leakage, a methodology that can be applied to any spatial ML problem. Local models have been shown to outperform the global SoilGrids v2.0 product in terms of accuracy, demonstrating the need for high-resolution regional models for precision agriculture. Local models outperformed SoilGrids v2.0 by 3.6× in Spearman ρ for pH (0.750 vs. 0.208), quantitatively confirming the necessity of regional calibration over global soil products. Multi-season ConvNeXt with SE-blocks on 54-channel composites improved R2 for NO3 by 36% (0.422 → 0.575), confirming the value of temporal dynamics for mobile elements; however, it underperformed RF on tabular features for most properties at the available sample size (n = 1085).