Synthesizing regional irrigation data using machine learning – towards global upscaling via metamodeling

Knowledge on irrigation is key to sustainable water resource management, but spatio-temporal irrigation data are rarely available. Recent advances are based upon satellite remote sensing data to quantify irrigation at high spatial resolution, and this study utilizes published irrigation datasets at regional scale to develop a metamodel approach to synthesize the available irrigation knowledge. We investigate the potentials and limitations of a Random Forest-based metamodeling approach that predicts irrigation at monthly timescale using only globally available and easily accessible features related to hydroclimatic and vegetation variables. The training dataset consists of three irrigation water use datasets derived from the soil moisture-based inversion framework and covers a variety of climatic conditions and irrigation practices in Spain, Italy, and Australia. Further, the study includes irrigation predictions from three test sites representing major global hot spots for unsustainable irrigation management: the North China Plain, Indus, and Ganges Basins. Our study aims to test the model transferability in space and time based on a series of split-sample experiments. We quantify and outline model transferability based on the area of applicability analysis, showing that although the feature space was mostly well represented, the magnitude of the target variable was equally important for assessing model transferability. A comprehensive feature importance analysis reveals that ranking of the most important input features depends on geographical extent of the training dataset. We find that model transferability was more robust across space than time within the small study areas, mainly because of the small geographical extents of the training datasets. The developed metamodel demonstrates satisfying performance with less than 10% bias and 3 mm/month mean error for a successful model transferability outside the training study areas and predicted spatial patterns of irrigation closely linked to climate and vegetation features. Given the increase in published regional irrigation datasets, we see great potential for further developing metamodel approaches for synthesizing existing knowledge and work towards global upscaling opportunities.