Efficient analysis of hydrological connectivity using 1D and 2D Convolutional Neural Networks

Understanding hydrological connectivity is essential to investigate ecological processes in river catchments and floodplains. Assessing flooding behavior, including flooded areas and connection times, is required to analyze hydrological connectivity in river floodplains. Deep learning, especially Convolutional Neural Networks (CNNs), is an attractive alternative to hydrodynamic modeling, which is more computationally expensive. This paper aims to develop a methodology to analyze the functional connectivity in remote and field measurement data-scarce areas using remote sensing data, CNN models, and connectivity metrics. The northern Lakes of the Narran River catchment, located in the Condamine-Balonne River floodplain in New South Wales, Australia, is the showcase for this method. One-dimensional CNN and two-dimensional U-Net configurations were applied and yielded comparable flood extents to the satellite images with Hit Rate values of 0.853 and 0.873, respectively. Two algorithms for determining hydrological connectivity were investigated, including the geostatistical Connectivity Function (CF) and the newly proposed Potential Connection Length (PCL). It was found that the connection along the main Narran River stream was more substantial than between the river and the floodplain lakes. The analysis using the PCL shows that the connectivity patterns in different stages of a flood event can vary depending on the initial condition of the floodplain. The overall conclusion from this work is that hydrological connectivity can be assessed computationally efficiently using only remote sensing, discharge data, and CNN models.