Reconstructing aquifer dynamics with machine learning: Linking irrigation expansion to groundwater decline in a data-scarce hyper-arid region

The Caplina aquifer, in the hyper-arid Peru–Chile border, sustains irrigation in a region with scarce surface water but is increasingly threatened by over-extraction and seawater intrusion. Effective management requires reliable characterization of groundwater level dynamics; however, records from the National Water Authority of Peru are highly dispersed and discontinuous, complicating spatiotemporal analysis. To solve this problem, we addressed data gaps by imputing missing water-level observations using four approaches: back-propagation neural networks (BPNN), multivariate imputation by chained equations (MICE), missForest, and k-nearest neighbors (KNN). BPNN achieved the lowest errors and most accurate spatiotemporal reconstructions. To synthesize temporal behavior across wells, we applied K-means and hierarchical clustering, separating signals into trend/shape and amplitude/size groups, revealing three dominant behaviors (decreasing, stable, and mixed), whose spatial distribution aligns with hydrogeologic settings. Over the last two decades, the central aquifer declined by 0.6 m yr−1, while near-coastal levels remained comparatively stable due to seawater intrusion. To examine potential drivers, we combined supervised RF analysis with Landsat imagery to map the expansion of irrigated areas through time, which shows a direct association with groundwater extraction. These findings highlight the urgent need to strengthen sustainable groundwater management to protect water availability and quality in this hyper-arid region.