Journal Paper Digests 2026 #27
- Transfer and Uncertainty for Soil Spectroscopy: A Parsimonious Approach Using Bayesian Mixed Models and Soil Surveys
- Soil science-informed data-driven modeling of effective saturated hydraulic conductivity
- A hybrid modelling framework for spatiotemporal prediction of soil organic carbon stock in Dutch agricultural soils
- Disentangling spatial extent, spatial grain, and geomorphic context in land-use change and landscape pattern analyses of a subtropical mountain–lake plain region
Disentangling spatial extent, spatial grain, and geomorphic context in land-use change and landscape pattern analyses of a subtropical mountain–lake plain region
Land-use/cover change (LUCC) and landscape metrics are scale sensitive, yet geomorphic context can confound their responses and statistical support can alter explanatory associations. We separated administrative–geomorphic comparisons, replicated extent and grain experiments, and fixed-support GeoDetector analyses in Hunan Province, China. Using the 30 m China Land Cover Dataset (1985–2025), we descriptively compared LUCC across four unequal administrative–geomorphic units. Separately, extent and grain responses were tested at five independent anchors in each of the two county contexts, and GeoDetector associations were assessed for three LUCC periods at a fixed 1000 m support. Forest and cropland dominated regionally; Anhua retained a relatively stable forest matrix, whereas Yuanjiang underwent active cropland–water turnover and impervious-surface expansion. Within the tested design, a significant spatial-extent response was detected only for the landscape shape index (LSI) across 5.01–15.03 km windows; no context × extent interaction was detected. Grain effects were significant for all five metrics; context × grain interactions occurred for patch density (PD), LSI, Shannon’s diversity index (SHDI), and Simpson’s diversity index (SIDI), but not for the aggregation index (AI). PD and LSI responded most strongly from 30 to 60 m. Slope and distance to built-up land had comparatively large q values for total change and cropland–forest turnover in Anhua; in Yuanjiang, distance to water bodies, surface-water frequency, and elevation ranked highly for total change and cropland–water turnover. All eligible interactions had joint q values exceeding both single-factor q values, and factor rankings were generally robust across temporal alignment, support size, and most discretization settings. These findings distinguish geomorphic context, spatial extent, spatial grain, and statistical support as distinct dimensions to be matched to the landscape attribute and research question.
A hybrid modelling framework for spatiotemporal prediction of soil organic carbon stock in Dutch agricultural soils
Accurate modelling and mapping of soil organic carbon (SOC) stocks across space and time is essential for climate change mitigation and sustainable land management. Machine learning models often achieve high predictive performance for digital soil mapping, but their data-driven nature may limit their ability to represent temporal SOC dynamics in a physically consistent way. To address this, integrating process-based models may provide additional process information to improve temporal consistency, while it may also improve predictive performance. This study focused on investigating whether integrating the Rothamsted carbon model (RothC) with a Random Forest (RF) model can improve spatiotemporal modelling of SOC stocks (0–30 cm) in Dutch agricultural mineral soils between 1986 and 2022. Specifically, two hybrid strategies were developed, where Hybrid 1 incorporates RothC outputs as additional covariates in the RF model, and Hybrid 2 applies RF to the residuals of RothC simulations and then adds the predicted residuals back to the RothC-simulated SOC stocks. Model performance was evaluated using both random and nearest neighbour distance matching (NNDM) cross-validation. The results indicate that Hybrid 1 achieved the best performance (ME = −1.19 t C ha−1, RMSE = 26.8 t C ha−1, MEC = 0.47), improving the MEC by 20.5% compared to RF under NNDM cross-validation. Spatiotemporal analysis showed that spatial and temporal patterns predicted by Hybrid 1 closely aligned with those of the RF model, whereas Hybrid 2 showed greater similarity to RothC. Hence, integrating process-based knowledge into digital soil mapping frameworks improved the accuracy of SOC stock predictions and reduced unrealistic temporal fluctuations. This study provides a flexible methodology for SOC stock prediction. The framework allows users to prioritize predictive performance or physical consistency, particularly in intensive agricultural systems.
Soil science-informed data-driven modeling of effective saturated hydraulic conductivity
The translation of readily available soil data into soil hydraulic conductivity parameters, which describe water infiltration at the soil surface, is often limited by complex, non-linear interactions associated with soil structure that are rarely captured by classical predictor sets. We hypothesize that soil taxonomic information, particularly World Reference Base (WRB) soil groups, encodes meaningful pedogenic and structural information that can help bridge this gap. We test this hypothesis using eleven Soil Science-Informed Machine Learning (SoilML) models trained on Europe-wide in-situ observations from the Soil Water Infiltration Global (SWIG) database. Models were developed to predict infiltration-based effective saturated hydraulic conductivity (Ks) using combinations of WRB soil groups, land-use type, and classical soil properties (clay, sand, silt, organic carbon, and bulk density), with additional hydrological state variables included in selected scenarios. Our results showed that ’tree-based’ models, specifically CatBoost, outperform linear regression and kernel-based neural approaches. Importantly, incorporating WRB soil groups and land-use information consistently improves predictive performance, increasing explained variance by up to 20% compared to texture-only models. We demonstrate that WRB soil classification provides useful additional information that reflects pedogenic development and site-specific structural effects not represented by soil texture alone. This highlights the potential of integrating soil taxonomic and land-use information into data-driven pedotransfer approaches for more robust upscaling of continental-scale soil hydraulic properties. However, through the use of spatial cross validation, it is found that although SoilML models perform well in regional interpolation, they fail miserably in terms of generalization across unknown geography.
Transfer and Uncertainty for Soil Spectroscopy: A Parsimonious Approach Using Bayesian Mixed Models and Soil Surveys
Two challenges that limit the adoption of soil spectroscopy are (1) model transfer to settings with limited data, and (2) uncertainty quantification, especially with correlated errors. Linear mixed models (LMMs) have sucessfully addressed these challenges in other fields, but not in soil spectroscopy. Compared to machine learning transfer algorithms, LMMs are more straightforward and interpretable. LMMs can also leverage auxiliary variables, in our case: farm, field, and soil depth variables; and external soil survey (SSURGO) estimates of pH and texture. But instead of using auxiliary variables to rigidly stratify a dataset into groups and fit separate models, LMMs partially pool the entire dataset to balance local and global modeling. In this study, we predict soil carbon from field-moist NIR spectra of 459 samples of three soil layers (0–15, 15–30, 30–60 cm) in 155 cores across 18 agricultural fields of 8 farms in Northeastern USA (data and code provided). Using cross validation, we find that our LMM succeeds in model transfer between farms, outperforming existing approaches such as memory-based learning and Cubist. Moreover, as a probability model, the LMM automatically provides excellent uncertainty quantification, including for aggregated predictions where existing methods fail due to correlated errors. We conclude that LMMs show promise as a parsimonious approach for model transfer and uncertainty quantification in soil spectroscopy.