Revealing spatial variability of groundwater level in typical ecosystems of the Tarim Basin through ensemble algorithms and limited observations

Abstract

Owing to the lack of hydrological monitoring stations and groundwater observations, particularly for natural habitats, the distribution of groundwater along the Mountainous (alluvial fan)-Oasis-Desert System (MODS) in the middle reaches of the Tarim River has not yet been systematically investigated. The performance of ensemble algorithms with multi-source/multi-scale (30 m, 90 m, 250 m, 500 m, and 1000 m) remote sensing data to predict the groundwater levels of the MODS of the Tarim Basin should be investigated. To improve the current knowledge on the spatial distribution of groundwater in this region, we conducted a study on the applicability of ensemble learning algorithms for groundwater assessments in typical regions of dryland, namely the Oasis-Taklimakan Desert (OTD), Nature land in Oasis system (NLOS), Irrigation area (IA), and Oasis system (OS). The results showed that the R2 values of the groundwater level prediction accuracy for the four ecosystems were 0.92, 0.96, 0.86, and 0.50, respectively, and their corresponding optimal ensemble algorithms were AR_GRBFN, AR_RF, RotationForest_MLP, and AR_ GRBFN. The scales corresponding to the optimal prediction accuracy of OTD, NLOS, IA, and OS were 250 m, 30 m, 250 m, and 90 m, respectively, and the effect of the scale on their respective groundwater level prediction accuracies (maximum value compared to minimum value) were 11.48%, 21.63%, 26.73%, and 10.59%. Random Subspace, Rotation Forest, and Additive Regression improved the prediction accuracy of the base learning method by 76%, 72%, and 64%, respectively, followed by Bagging, whereas Dagging greatly increased prediction errors. Rotation Forest and Random Subspace showed stable performances and guaranteed relatively low prediction errors. Among all ensemble algorithms, Additive Regression helped the base learners obtain relatively optimal prediction accuracies with high probabilities. Except for the OS system, groundwater level could be predicted with much greater accuracy at depths of >8 m than at <8 m. The contribution analysis showed that topography and land-use patterns controlled the spatial distribution of groundwater across MODS. The ensemble learning algorithm showed good performance using multi-source and multi-scale data.