Towards site-independent calibration of in situ soil pH sensor data: Relevance of spatial and temporal proximity, sample size and data spread for calibration model performance

• Sample size is the key factor for improving calibration model performance. • Field-wise calibration lowers RMSE by 25% but requires more reference samples. • Neighborhood, farm, universal models cut sample size by 74, 87, 86% at similar RMSE. • pH range > 1 unit is crucial for achieving high quality calibration models. • Spatial and temporal dispersion have minimal impact on model performance. Proximal soil sensing data usually require calibration to extract meaningful information, e.g. for precision agriculture. For optimal application, calibration modeling is typically based on reference data collected from field-wise sampling. However, this approach is neither time- nor cost-effective and hinders the adoption of proximal soil sensing in practical agriculture. Thus, minimizing calibration sampling is essential. For the case of pH mapping in precision agriculture, we have investigated whether calibration efforts can be reduced by reference sampling, which is less dependent on location and timing. This study utilized 612 sensor and standard lab pH measurements from 62 fields of two farms in Northeast Germany. We assess the effects of spatial and temporal proximity, sample size and data spread on the performance of univariate linear regression models for pH sensor data calibration. Moreover, the performance of site-independent calibration was analyzed for five domain levels: (i) regional, (ii) farm, (iii) spatial neighborhood (NH), (iv) temporal NH and (v) field. This multi-level approach aims at assessing the generalizability of calibration models. Partial correlation analysis revealed that sample size is the most crucial factor for optimizing model performance. Field-wise calibration yielded a 34 % lower mean RMSE than regional, farm, or neighborhood models but required the most reference samples. In contrast, neighborhood models reduced sample size by approximately 74 %, farm models by 87 % and the regional model by 86 %, while achieving similar RMSEs. Thus, farm or regional models can significantly reduce sampling effort and costs in practical precision agriculture. The spread of the pH value was the next important factor influencing model performance. The pH range in the sensor measurements selected for calibration should be larger than 1 unit to obtain good quality calibration models. Finally, spatial and temporal dispersion had the least effect on calibration performance.