Land cover segmentation of high-sesolution images using deep learning

High-resolution remote sensing imagery is widely used for land cover mapping and has found application in several fields, such as urban planning, environmental monitoring, agriculture, and disaster management. Over the last years, the availability of very-high-resolution aerial and satellite images has increased considerably. Nevertheless, accurate land cover classification remains a difficult problem. This is mainly due to the high spatial variability of the scene, similarities in spectral characteristics between different land cover types, strong class imbalance, and the presence of complex object boundaries. These issues are particularly met in urban and heterogeneous areas. For these reasons, traditional pixel-based and object-based classification methods often fail to provide satisfactory results. This thesis investigates the use of deep learning methods for land cover classification in high-resolution aerial imagery, with a specific focus on semantic segmentation approaches based on encoder–decoder architectures. The adopted model consists of a deep residual network used as an encoder and a decoder inspired by the U-Net architecture. This design allows the extraction of semantic information at different spatial scales, while still preserving local spatial details that are important for accurate segmentation. The experiments are conducted on the LandCover.ai dataset, which is a publicly available dataset of high-resolution RGB aerial images. The dataset includes five land cover classes, namely background, building, woodland, water and roads, but a reduced three-class configuration is adopted in this study. Specifically, the building and road categories are merged with the background class to mitigate severe class imbalance, relatively small spatial extent, structural complexity, and their uneven distribution in the dataset. In order to improve the segmentation of object boundaries, a combined loss function based on Focal loss and Dice loss is employed. Additionally, data augmentation techniques and a one-cycle learning rate scheduling strategy are applied during training. The results obtained from the experiments show that the proposed approach performs consistently on both the validation and test datasets. The achieved mean Intersection over Union (IoU) is approximately 87% under the reduced three-class configuration. High segmentation accuracy is observed for background, woodland and water classes. Both quantitative results and visual inspection of the segmentation outputs indicate that the encoder–decoder architecture is able to capture global contextual information as well as local spatial characteristics. Overall, the findings confirm that deep encoder–decoder models provide a reliable solution for high-resolution land cover classification, while also indicating the importance of class distribution considerations in semantic segmentation performance evaluation. At the same time, the results highlight the need for further investigation on issues related to scalability, transferability, and domain adaptation in practical remote sensing applications.