Zhang, Ce and Atkinson, Peter (2018) Deep learning for land cover and land use classification. PhD thesis, Lancaster University.
PhD_Thesis_Ce.pdf - Published Version
Available under License Creative Commons Attribution.
Download (12MB)
Abstract
Recent advances in sensor technologies have witnessed a vast amount of very fine spatial resolution (VFSR) remotely sensed imagery being collected on a daily basis. These VFSR images present fine spatial details that are spectrally and spatially complicated, thus posing huge challenges in automatic land cover (LC) and land use (LU) classification. Deep learning reignited the pursuit of artificial intelligence towards a general purpose machine to be able to perform any human-related tasks in an automated fashion. This is largely driven by the wave of excitement in deep machine learning to model the high-level abstractions through hierarchical feature representations without human-designed features or rules, which demonstrates great potential in identifying and characterising LC and LU patterns from VFSR imagery. In this thesis, a set of novel deep learning methods are developed for LC and LU image classification based on the deep convolutional neural networks (CNN) as an example. Several difficulties, however, are encountered when trying to apply the standard pixel-wise CNN for LC and LU classification using VFSR images, including geometric distortions, boundary uncertainties and huge computational redundancy. These technical challenges for LC classification were solved either using rule-based decision fusion or through uncertainty modelling using rough set theory. For land use, an object-based CNN method was proposed, in which each segmented object (a group of homogeneous pixels) was sampled and predicted by CNN with both within-object and between-object information. LU was, thus, classified with high accuracy and efficiency. Both LC and LU formulate a hierarchical ontology at the same geographical space, and such representations are modelled by their joint distribution, in which LC and LU are classified simultaneously through iteration. These developed deep learning techniques achieved by far the highest classification accuracy for both LC and LU, up to around 90% accuracy, about 5% higher than the existing deep learning methods, and 10% greater than traditional pixel-based and object-based approaches. This research made a significant contribution in LC and LU classification through deep learning based innovations, and has great potential utility in a wide range of geospatial applications.