Abstract
Field measurements of soil respiration (SR) are often conducted using one-time campaigns with portable chambers rather than continuous automated monitoring. However, SR is highly sensitive to soil temperature and moisture, which vary among measurement campaigns and can mask long-term trends. Although normalizing biological indicators to environmental conditions is standard practice in many fields of ecology, normalizing in situ SR data to both temperature and moisture has not previously been attempted. To address this gap, we normalized SR to reference temperature and moisture using both conventional regression and machine-learning (ML) models. We propose selecting reference values not arbitrarily, but as the temperature and moisture corresponding to predicted maximum SR while remaining close to typical summer field conditions in the study region. Using a nine-year dataset (2011–2024, 36 measurement campaigns) from 60 sample plots along pollution gradients around two copper smelters in the Urals, we tested three hypotheses: (i) ML models outperform conventional regression models in predicting SR, (ii) normalization to common temperature and moisture reduces interannual variation in pollution effect sizes, and (iii) ML-based normalization provides greater precision than conventional regression-based normalization. All three hypotheses were supported. XGBoost showed the highest predictive accuracy among ML algorithms. ML-based normalization substantially reduced interannual variability and narrowed confidence intervals for the pollution effect size compared with conventional regression normalization and raw observations. We conclude that ML-based normalization is a practical tool for reducing weather-related noise in repeated SR field measurements, improving the detection of long-term trends.

This publication has 40 references indexed in Scilit: