Deep Learning for Urban and Landscape Mapping from Remotely Sensed Imagery: Fid-Bau Portal

Deep Learning for Urban and Landscape Mapping from Remotely Sensed Imagery

Lai, Feilin / Sharma, Atharva / Liu, Xiuwen / Yang, Xiaojun

Deep learning has recently become a trending research topic in remote sensing because of its potential in generating high performance for pattern classification. Some deep learning models can make effective use of spectral, spatial, and temporal information from remotely sensed data, resulting in improved mapping solutions, especially over complex environments, such as urban areas. This chapter provides an overview of deep learning models for urban and landscape mapping from remotely sensed data focusing on the potential of deep neural networks to address current challenges in land classification. It begins with a brief discussion of the evolution of artificial neural networks and the basic architecture of multi‐layer neural networks deemed as the foundation for developing deep learning models, which is followed by a summary of some major advantages of deep learning. Then, several deep learning models commonly used in remote sensing are introduced, along with a close look at the two most popular models: convolution neural networks (CNNs) and recurrent neural networks (RNNs). Two case studies using CNNs and RNNs for landscape mapping over a complex urbanized coastal area are further presented to demonstrate how deep learning models can be used to generate improved performance in remote sensing. It is believed that these case studies can encourage further thinking over some potential issues (e.g. hyperparameter optimization) challenging the performance of deep learning applications in remote sensing .