Cnns in land cover mapping with remote sensing imagery: a review and meta-analysis

Convolutional neural network (CNN) comprises the most common and extensively used network in the field of deep learning (DL). The design of CNNs was influenced by neurons, like a traditional neural network. CNN has fundamental advantage over earlier works since it can detect in an automatic way critical features without the need for human intervention. CNNs have been widely employed in various applications, including land cover classification. Multiple CNN architectures have been introduced over the previous decade. Applications rely on model architecture to improve their performance. The CNN architecture has undergone several alterations up to this day. Several CNN architectures have been introduced in the literature, depicting strong and weak points. This review article presents an overview of the development of convolutional neural networks as they are described in state-of-the-art literature, including remote sensing books, journals, and conferences. Following a thorough assessment of current CNN case studies in land cover mapping through statistical analysis, informative results pertaining to the implemented CNN architecture are presented including relevant findings such as the framework that was utilized, highlighting the most popular choices among the users. It has to be noted that there is not a miraculous CNN model, and the statistical findings reflect the latest developments. Finally, current issues and innovative aspects are addressed.