Remote-sensing Image Segmentation using Convolutional Neural Network and Its Applications to Pan-sharpening and Detection

  • Essa Basaeed

Student thesis: Doctoral Thesis


Essa Ibrahim Basaeed, 'Remote-sensing Image Segmentation using Convolutional Neural Network and Its Applications to Pan-sharpening and Detection', Ph.D. Thesis, Ph.D. in Engineering, Electrical and Computer Engineering Department, Khalifa University of Science, Technology, and Research, United Arab Emirates, Mar. 2016. Remote-sensing images play an important role in different applications such as target detection, urbanization, and change detection. In such applications, there is an increasing demand for autonomous image understanding in order to support intervention-free analysis. Remote-sensing image segmentation is one of the fundamental image processing techniques that facilitates autonomous image understanding. In this thesis, a supervised boundary-based remote-sensing image segmentation method is introduced using a novel multi-band multi-scale fusion framework that exploits a committee of Convolutional Neural Networks (CNNs). The proposed boundary-based image segmentation method begins with the classification of image pixels in individual bands of a multi-spectral remote-sensing image into boundary and nonboundary pixels using a committee of CNNs. The confidence in classification produced by each network forms a confidence map. Confidence maps are then intra-fused in order to combine spectral cues across different bands and produce a fused confidence map. Further, the fused confidence map is inter-fused in order to integrate spatial information at the same plane and provide a hierarchical segmentation. The proposed fusion framework provides a seamless integration of spectral and spatial information while remaining generic to the characteristics of remote-sensing images and the requirements of the end-user application. This thesis also suggests the incorporation of multi-scale analysis within the framework in a novel way that comes with no increase in computational complexity of the CNN compared to the single-scale counterpart and no redesigning burden in terms of the CNN architecture while improving effectiveness. Moreover, the accuracy of classification of the committee of CNNs is further improved through boosting where individual networks within the committee are exposed to increasingly challenging examples from the data set. The proposed framework is both qualitatively and quantitatively compared to baseline methods on publicly available data sets. Finally, the thesis presents two applications for the proposed framework, namely, water classification and region-based pan-sharpening. The first is an important preprocessing step in fields such as environmental monitoring, target detection, and map updating. Through combining a novel spectral water index along with the proposed segmentation method, an accurate discrimination between water and land regions in multi-spectral satellite imagery data can be achieved. The second application, region-based pan-sharpening, aims at incorporating region information into the process of fusing multi-spectral and panchromatic images thus assuring spatial and spectral consistency in the pan-sharpened image. The research work described in this thesis is among the first remote-sensing image segmentation methods that incorporate CNNs. It represents a step forward towards achieving a seamless integration of remote sensing and geographic information system (GIS). Such an integration would allow a better use of remote-sensing images. Indexing Terms: image segmentation; artificial neural networks; multispectral imaging; remote sensing
Date of AwardJun 2016
Original languageAmerican English
SupervisorMohammed Al-Mualla (Supervisor)


  • Image segmentation; artificial neural networks; multispectral imaging; remote sensing.

Cite this