Classification of estuaries and coastal wetlands from Planet-NICFI imagery based on convolutional neural networks and transfer training

Abstract

The authors consider the importance of monitoring coastal wetland ecosystems, negatively impacted by human activities and climate change. In this context, artificial intelligence neural networks are applied to classify this type of wetland. However, they encounter a task that requires extensive volume of training data to achieve high accuracy results. Within the conducted research, a method of transfer training from neural networks is proposed to overcome the aforementioned problem. The developed model combines multi-temporal Planet-NICFI satellite images for classifying coastal wetlands, especially under tidal conditions. The research results indicate that the model has upgraded its accuracy from 89,2 % to 91,3 % in the wetlands of the Ba Lat estuary. Besides, it has been successfully applied to classify similar lands in the Red River Biosphere Reserve during the period of 2016–2022. This will enable improving the management of this area in the future