AUTHOR=Zeynoddin Mohammad , Gumiere Silvio José , Bonakdari Hossein TITLE=Enhancing water use efficiency in precision irrigation: data-driven approaches for addressing data gaps in time series JOURNAL=Frontiers in Water VOLUME=5 YEAR=2023 URL=https://www.frontiersin.org/journals/water/articles/10.3389/frwa.2023.1237592 DOI=10.3389/frwa.2023.1237592 ISSN=2624-9375 ABSTRACT=

Real-time soil matric potential measurements for determining potato production's water availability are currently used in precision irrigation. It is well known that managing irrigation based on soil matric potential (SMP) helps increase water use efficiency and reduce crop environmental impact. Yet, SMP monitoring presents challenges and sometimes leads to gaps in the collected data. This research sought to address these data gaps in the SMP time series. Using meteorological and field measurements, we developed a filtering and imputation algorithm by implementing three prominent predictive models in the algorithm to estimate missing values. Over 2 months, we gathered hourly SMP values from a field north of the Péribonka River in Lac-Saint-Jean, Québec, Canada. Our study evaluated various data input combinations, including only meteorological data, SMP measurements, or a mix of both. The Extreme Learning Machine (ELM) model proved the most effective among the tested models. It outperformed the k-Nearest Neighbors (kNN) model and the Evolutionary Optimized Inverse Distance Method (gaIDW). The ELM model, with five inputs comprising SMP measurements, achieved a correlation coefficient of 0.992, a root-mean-square error of 0.164 cm, a mean absolute error of 0.122 cm, and a Nash-Sutcliffe efficiency of 0.983. The ELM model requires at least five inputs to achieve the best results in the study context. These can be meteorological inputs like relative humidity, dew temperature, land inputs, or a combination of both. The results were within 5% of the best-performing input combination we identified earlier. To mitigate the computational demands of these models, a quicker baseline model can be used for initial input filtering. With this method, we expect the output from simpler models such as gaIDW and kNN to vary by no more than 20%. Nevertheless, this discrepancy can be efficiently managed by leveraging more sophisticated models.