Performance Variability of Mid-Infrared Spectroscopy–Based Predictions of Soil Radiocaesium Dynamics across Diverse Soil and Land Use Conditions

Following the Fukushima Daiichi Nuclear Power Plant accident, radioactive substances such as radiocaesium (¹³⁷Cs) were widely dispersed and contaminated soils, raising concerns about their transfer from soil to crops. ¹³⁷Cs transfer is primarily regulated by exchangeable potassium (K_Ex), a chemically analogous element, but its effectiveness varies across environmental conditions such as soil type and land-use. Recent studies suggest that soil exchangeable ¹³⁷Cs (¹³⁷Cs_Ex) dynamics and its solid–liquid partitioning play key roles in predicting ¹³⁷Cs transfer irrespective of regional differences. In contrast, current direct methods for measuring ¹³⁷Cs are costly and time-consuming, making them unsuitable for rapid risk assessment. As an alternative approach for risk management, mid-infrared spectroscopy (MIRS) may provide a rapid and cost-effective means of estimating soil properties. Recently, models for predicting soil K_Ex concentrations from spectral data have been reported. However, their applicability to ¹³⁷Cs transfer remains unclear. In this study, we aimed to construct prediction models for the ratio of soil ¹³⁷Cs_Ex to soil total ¹³⁷Cs (¹³⁷Cs_Total) using MIRS spectra and to evaluate the variability of model performance among soil or land-use categories.

1249 soil samples collected in Fukushima Prefecture, Japan, from 2015 to 2020, were analyzed for soil properties, including soil total C, ¹³⁷Cs_Ex, and ¹³⁷Cs_Total, through MAFF and NARO in Japan. Each soil sample was analysed after drying at 37°C for at least 12 hours and being sieved to less than 0.2 mm before measurement. Mid-infrared spectra for these samples were obtained at the FAO/IAEA Soil and Water Management and Crop Nutrition Laboratory over the wavenumber range of 650–4000 cm^–1 using four replicate measurements per sample. Using noise-removed spectral data, partial least squares regression models were developed to predict the ratio of soil ¹³⁷Cs_Ex to ¹³⁷Cs_Total. In addition, prediction models were constructed for different soil types (andosol, brown forest soil, lowland soil, and peat soil) and land-use categories (upland fields and paddy fields), and their differences in model performance were evaluated.

Prediction models were constructed and achieved moderate predictive performance (R² around 0.6). In contrast, by stratifying prediction models by soil type, prediction accuracy improved for all soil types except for peat soil relative to the non-stratified model. In particular, andosol showed the highest prediction accuracy. Comparison of variable importance in projection (VIP) scores among these models showed that the contributions of specific wavenumber ranges to model performance differed among soil types. In andosols, VIP scores were higher in wavenumber ranges associated with carbohydrates, quartz, and clay minerals compared with the model constructed using all data. These results suggest that soil type specific mineralogical composition and carbon content may play roles in improving prediction performance. Furthermore, predictions stratified by land-use showed higher accuracy in upland fields than in paddy fields. Differences of VIP scores between them were also observed in wavenumber ranges associated with carbohydrates and clay minerals. These results suggest that environmental conditions, such as soil redox status, may influence prediction accuracy through their effects on soil minerals and carbon.