Using fully convolutional networks and pléiades satellite imagery with pre-trained VGG-19 model to detect impervious surface in farmland

The rapid growth of artificial impervious surface areas such as buildings and pavements has caused the loss of arable land and pose threat to food production. Periodic monitoring the use of farm land is necessary. It is on an urgent need to develop laborsaving methods since commonly practices of tallying impervious surfaces are based on manual digitization. Emerging deep convolutional neural networks (CNNs)-based techniques have shown a great potential for remote sensing image classification tasks. Transfer learning refers to the technique that transfers pre-learned representation from one domain to another. And, it has been generally applied in CNNs on specific datasets to reduce the data labelling effort and shorten the training process by adopting pre-trained networks. This study aims to explore the potential of detecting impervious surfaces using fully convolutional networks (FCNs) with pre-trained network parameters and high resolution multispectral satellite imagery. Since off-the-shelf pre-trained networks are usually trained with RGB bands, the number of input bands is restricted to three. We propose adding a convolutional layer, which generates three-layer feature maps before the pre-trained network. Experiments on a pansharpened Pléiades satellite image dataset with a pre-trained VGG-19 network were conducted. The classification accuracy F1-score of 94.2% was achieved.