Valuation Impact of Data Quality Using Geospatial Machine Learning: Application to Soil Data in Prairie Canada

This paper develops a Machine Learning (ML) approach to quantify the impact of geospatial data quality on farmland valuation accuracy. The approach is useful when both the set of quality attributes and the predictor set are high-dimensional relative to the sample size, making standard non-market valuation methods difficult to implement. Conceptually, we draw on value of information (VoI) theory, which defines value as the difference in expected payoff under higherquality versus imperfect information. The VoI measure can be approximated as the difference between predicted farmland value under full-quality data and predicted value under incomplete data. We apply the approach to soil data and farmland values in the Canadian Prairies. Baseline and counterfactual values are predicted using a Random Forest model with hyperparameters tuned on spatially blocked folds and evaluated with blocked spatial cross-validation to limit spatial leakage. Counterfactual quality scenarios are implemented as random perturbations of soil features. Specifically, we simulate six scenarios: spatially localized bias, limited geographic coverage, coarse spatial resolution, measurement error, numeric rounding, and categorical misclassification. The results show that coarse spatial resolution generates the largest average valuation distortion (238.46 CAD/ha, 2006 CAD), followed by limited coverage (104.49 CAD/ha), while the remaining degradations have small effects. Quality–value curves traced over degradation intensities are nonlinear and concave, consistent with diminishing marginal returns to information improvements. The findings have policy implications for prioritizing investments in public geospatial data: budgets may yield higher returns by shifting from incremental precision upgrades toward improving spatial coverage and resolution.