Ordinary kriging as a data augmentation technique for enhanced total dissolved solids prediction in tropical wetlands using Sentinel-2 and machine learning

This study aimed to demonstrate the efficacy of the ordinary kriging interpolation method as a data augmentation technique to enhance the accuracy of models that predict the water quality parameter total dissolved solids (TDS). For this purpose, the total dissolved solids parameter was utilized as a case study. The models employed seven spectral indices derived from Sentinel-2 images as predictor variables, recognized as effective in predicting TDS: Green, Ferdous, SI-3, SI-5, TDS-1, and TDS-5. Three types of models were compared: (1) simple linear regression models trained with the original field samples; (2) simple linear regression models trained with an augmented dataset of 1000 data points, generated from a geostatistical kriging surface derived from the in-situ data; and (3) multilayer neural network models trained with the same augmented dataset of 1000 data points. It was concluded that the ordinary kriging method is an effective data augmentation technique, as there is a statistically significant difference between the models trained with field samples and those trained with augmented data, with the latter demonstrating greater explanatory power. Additionally, in the validation of the prediction surfaces derived from the models trained with the augmented dataset, it was observed that the mean absolute error (MAE) and the root mean square error (RMSE) were below the natural variability of the original in-situ data, thereby confirming the high predictive capacity of the models trained with the augmented dataset. This research is of significant importance, as it innovatively integrates geostatistical techniques, satellite remote sensing, and artificial intelligence to effectively address the critical issue of data scarcity in water quality modelling, taking TDS as a test parameter. The results of this research establish a robust foundation for the application of artificial intelligence in water quality modelling, facilitating precise predictions at specific locations and times.