Abstract/Summary

In LULC (land-use land-cover) studies, a wide range of sensors have visualised a vast number of very high-resolution remote sensing (RS) aerial scenes, gathering images with very fine spatial details that are both spectrally and spatially complicated. As such, a fundamental challenge is that of automating this classification to achieve maximum advantage. Recently, there has been a wave of excitement in deep learning regarding modelling high-level abstractions through hierarchical feature representations, an approach that demonstrates great potential in automating the interpretation of LULC patterns without human-designed features or rules. In this research, a set of novel deep learning classification methods based on deep convolutional neural networks (CNN) are investigated and adapted. The modelling consists of two phases, one-scene classification and object-based segmentation, both applied to remotely sensed images for land cover classification. For one-scene classification, some generated rules are deduced to tackle the limitation of the limited number of RS images. AlexNet was utilised in conjunction with two exploration strategies, fine-tuned and fully trained, thus revealing an effective result. The filter size batches, and epochs and input sizes can be scientifically tuned for better performance. On the other hand, comparing large epochs to batch size reveals a positive result for a smaller number of selected batches. In the second part of this thesis, the most recent Fully Convolutional Neural (FCN) network architectures for object-based segmentation are considered by utilising multiple state-of-the-art architectures which address the limitations of standard CNN classifiers. To improve the effectiveness of such efficient state-of-the-art architectures, an extensive adaption experiment was carried out to evaluate the performance of FCDenseNet, PSPNet and MobileUNet, which proved to be extremely powerful for land cover segmentation, specifically for small objects, by investigating the effect of the number of layers of each Densenblock, increased max pooling layers and skipped connections, respectively. In addition, comprehensive comparison experiments using the most recent state-of-the-art semantic segmentation architectures (FC-DenseNet, GCNet, PSPNet, DeepLab, Encoder-Decoder, AdapNet and RefineNet) were implemented, revealing the superior performance of FC-DenseNet on land cover segmentation, specifically small-object segmentation, due to the successive iterative concatenations of dense blocks and skip connections that capture information from the loss. To further improve the localisation of previously predicted maps, an analytical investigation was applied to measure the effectiveness of a localisation-improving method (FCCRF) in which a Gaussian kernel is used to define the pairwise potentials, which qualitatively improves large object boundaries. Thus, a visual labelling improvement is noticed, where excessive smoothing might harm small, isolated regions. The main contribution of the deep learning techniques is achieved by the development of a customised architecture that is built upon the best-performing deep segmentation architecture (FC-DenseNet), resulting in an improvement of 0.96% in overall average testing accuracy and improvements of 2.7% and 3% in average accuracy scores for cars and trees, respectively, compared to the baseline model where data augmentation achieved an improvement for the low vegetation class of 6.3% in the F1 score and 12% in the average accuracy score. This research makes a significant contribution to LULC classification through deep-learning-based innovations and has potential utility in a wide range of geospatial applications.