Area wise high resolution water availability estimation using heterogeneous remote sensing and ensemble machine learning

In this paper a novel remote sensing data integration framework has been developed using ensemble machine learning to estimate large area wise ground water balance. Heterogeneous spatio-temporal database including `Australian Water Availability Project (AWAP) database´, `Australian Digital Elevation data (ADED)´, and `NASA MODIS Vegetation Index (VI) data´ were processed and integrated. An irrigated farming area (covering 20km × 20km) in Tasmania described by S 42°36 Latitude and E 147°29 Longitude, where weekly data from the period Jan 2007 - Dec 2013 (total 320 weeks) were studied. An ensemble machine learning framework combining Sugano type Adaptive Neuro Fuzzy Inference System (ANFIS), Elman (ENN), Cascade Feed Forward (CFFNN), and Function fitting neural networks (FFNN) were trained with combined training inputs of VI and ADED demographic data against the AWAP based water balance estimations as training targets. Based on the spatial distribution of the training performance, different trained estimators were selected to estimate water balance at various spatial locations purely based on VI and ADED inputs, where no AWAP data were available. A high-resolution (250m) water availability map was created for the whole area on a weekly temporal scale, which could potentially provide accurate irrigation management support over a very large area.