A review on linear feature extraction using remote sensing approaches

Keywords: Remote sensing, Linear feature, Supervised and unsupervised methods, Deep learning, Mapping, Accuracy assessment
Session Type: Virtual Poster Abstract
Day: Friday
Session Start / End Time: 2/25/2022 05:20 PM (Eastern Time (US & Canada)) - 2/25/2022 06:40 PM (Eastern Time (US & Canada))
Room: Virtual 40

Abstract

Acequias are features of critical importance to various stakeholders, but their exact distribution is unknown. Automated mapping is needed but has not been applied to date. This poster will provide a review of existing extraction approaches of linear features (e.g., roads, rivers, and shorelines) using remote sensing –approaches that may also be used for automatically mapping acequias. The review emphasizes medium to very high-resolution remote sensing data and describes the advantages and disadvantages of available methods as well as their potential usefulness for mapping acequias. This review also focuses on the supervised and unsupervised extraction methods for different types of linear features to make a comprehensive comparison among the methods. It is noticed that most of the researchers used supervised methods for feature extraction and applied those methods for road extraction. The practice of using deep learning and machine learning algorithms has been noticed in this review to extract features, especially for roads where the high spatial resolution was given the main priority. The performance of those models has been evaluated by evaluation matrices where it is found deep learning models showed an average of above 90% accuracy compared to other traditional models. The data-driven methods are found prominent. Moreover, the research analyzes the general study gaps and scope of improvement for future research that can help different professionals to use those models according to their requirements. Finally, potential models for extracting acequias have been discussed by considering context, feature type, data types, and performance of different models.