Authors: Feilin Lai*, Florida State University, Xiaojun Yang, Florida State University
Topics: Remote Sensing, Land Use, Urban Geography
Keywords: Deep learning, Urban landscape, Remote sensing, ResNet, CNN
Session Type: Virtual Paper
Start / End Time: 8:00 AM / 9:15 AM
Room: Virtual 33
Presentation File: No File Uploaded
With the global trend of urbanization, it is increasingly important to map and monitor urban land covers for resource management, environmental protection, and sustainability planning. Although remote sensing has long been used to map land cover and land use, its applicability over urban areas can be compromised by the heterogeneity and complexity of urban landscapes. Deep learning is a newly developed branch of neural networks with deep structures that are more efficient in modeling complex dependencies. While a range of deep learning models has been applied to improving remote sensor image classification, there is no case-specific guidance on which deep learning model to choose, and there is no firm conclusion that deep learning models are significantly better than traditional classifiers. This study aims to examine the potential of some deep learning models in land cover mapping in a complex urban area and whether they can significantly outperform traditional classifiers. Specifically, two common deep learning models, i.e. convolutional neural networks and residual neural networks, were implemented and compared in mapping a complex urban area. Their performance will be compared with three conventional classifiers, i.e. random forests, support vector machines (SVM), and multi-layer perceptron neural networks (MLP). The experiment was conducted in Google Colabs with access to Google Earth Engine and advanced deep learning packages.