A Data-Centric Approach to Water Quality Prediction: Sample Size, Augmentation, and Model Performance with a Focus on Ammonium in a Tropical Wetland

Framed within data-centric artificial intelligence, this study integrates statistics, geotechnologies and AI to improve water quality prediction. The primary objective was to identify the minimum sample size required to train robust and accurate machine learning models. Based on 30 sampling points in a tropical wetland in northern Colombia, ammonium concentration was selected as the target variable, and total dissolved solids, suspended solids, phosphate, dissolved oxygen, nitrate and chemical oxygen demand were chosen as predictors. Because 30 observations are insufficient to train robust models, data augmentation was performed using ordinary kriging (OK) and empirical Bayesian kriging (EBK). From the kriging-interpolated surfaces, 1000 synthetic points (randomly and spatially distributed while preserving the estimated spatial structure) were sampled; from this expanded dataset, subsamples of varying sizes were drawn to train six algorithms: multiple linear regression (MLR), random forest (RF), k-nearest neighbours (k-NN), gradient boosting machines (GBM), multilayer perceptron (MLP) and radial basis function neural network (RBF-NN). The RF, k-NN, MLP, RBF-NN and GBM models trained on the interpolated data exhibited excellent performance: in the testing phase, they achieved adjusted coefficients of determination > 0.95 and symmetric mean absolute percentage errors (SMAPEs) < 10%, and the resulting predictive surfaces showed comparable performance under external validation. According to the criteria of stability, goodness of fit, and external validation, the optimal minimum sample size for most algorithms was 104 observations. These results represent a significant advance in mitigating data scarcity in water quality modelling. The identification of effective data augmentation methods and the determination of appropriate sample sizes, as demonstrated here, support the robust application of AI techniques in water quality prediction. The proposed strategy is transferable to other quantitative, spatially continuous environmental variables and thus contributes to the development of the emerging subdiscipline of geospatial artificial intelligence (GeoAI).