Identifying coastal land cover types using a hybrid approach of optical and SAR satellite data in the Auckland region

Abstract

Satellite remote sensing provides low-cost data appropriate for assessing coastal change at large scales. To apply such techniques at regional and national scales, initial identification of coastal land cover types using satellite data is required with high degrees of accuracy. In New Zealand the physical diversity of the coast presents challenges for large-scale application of remote sensing techniques. This paper aims to identify coastal land cover types in the Auckland region by validating a hybrid rule-based and machine learning methodology developed in Google Earth Engine that utilises freely available satellite data from both optical and synthetic aperture radar (SAR) sensors. Data pre-processing was applied to develop a composite image for 2019 containing data from Sentinel-1 (SAR) and Sentinel-2 (optical) satellites. Hierarchal rules were developed to separate water and vegetation land covers. The remaining land cover types (titanomagnetite volcanic sand, quartz-feldspathic sand, intertidal, artificial surfaces, and rock coasts) were identified using a random forest machine learning classifier trained with high-resolution satellite data and ancillary datasets. The overall accuracy of the approach was 82% with a kappa statistic of 0.79. The overall detection of sandy coast in the Auckland region had a producer’s accuracy of 93.6% and user’s accuracy of 86.4%. SAR data provides valuable information about the physical/textural characteristics of built areas that reduces the misclassification of these areas as quartz-feldspathic sand, leading to greater accuracies compared to using just optical data. The methodology provides a low-cost solution for the identification of coastal landcover types that can be applied at national scale. Such information will be instrumental in efforts to develop remote sensing approaches to accurately detect coastal change beyond local scales.