Hāpua are freshwater river mouths that are parallel to the coastline with a dynamic gravel or mixed sand and gravel barrier intermittently separating them from the marine environment. Our understanding of hāpua processes are limited by a lack of high frequency and high quality data. To address this lack of data, deep learning models are tested across hāpua located along the Canterbury Bight. The resulting models produce 2-dimensional masks of wet/dry areas. Delineations of water and land provide insight into hāpua morphology and water storage changes.

Field work at these lagoons can be challenging, which means hāpua studies tend to focus on individual sites. There are few studies that report synoptic information on the morphology of multiple hāpua and interrogate their changing form. To address this data gap, this thesis presents water mask results using a deep learning framework to delineate hāpua using high-resolution and near daily frequency imagery from the Planet satellite constellation. In this approach a convolution neural network is built and trained to automate the delineation of binary (wet/dry) masks to represent lagoon, barrier and tidally influenced river areas a system wide scale. These masks are used to conduct a long term study of the Rangitata hāpua morphology.

The results highlight that the convolution neural network model trained with 90 scenes at a tile mosaic of 10 by 10 pixels (T90M10) produced the most reliable water masks both through time and across study sites. A comparison between manual shore line delineation, a semi-automatic classification tool, and the T90M10 model demonstrated a close fit. However, the CNN model demonstrated slight under fitting of the water line and is therefore conservative in its water estimates.

This thesis also undertook an analysis of hāpua morphology over time. Through water masks from the T90M10 model, it was found that mouth width and orientation fluctuate in response to river flow, where wider openings correlating with straight mouth positions. Mouth width also correlates inversely to hāpua water volume, whereby wide mouth openings result in lower observed volumes.

With increasing pressures of water abstraction coupled with changing river regimes and sea-level rise under a changing climate, this approach offers important insights that may help to benchmark current conditions.