Dottorato in INFRASTRUTTURE E TRASPORTI

Titolo della tesi: Automated 3D change detection for land use evaluation using data science techniques

An increase in urbanization resulting from an increase in economic activities within urban contexts has characterized the last decades. This phenomenon has caused an intensification of overbuilding that must be monitored constantly. One of the most widely used methodologies to carry out this type of monitoring is the analysis of remotely sensed data (RS), as it can provide useful and consistent information on urban morphological structures with different spatial and temporal resolutions, allowing for long-term spatiotemporal analyses of the historical development of cities and thus monitoring the evolution of their urbanisation patterns, an objective closely related to the United Nations Sustainable Development Goals (SDGs) (SDG 11 - Sustainable Cities and Communities). In this context, most of the recently developed Change Detection (CD) methodologies rely on deep learning architectures. Nevertheless, this kind of algorithm mainly focuses on the generation of two-dimensional (2D) change maps, in which the planimetric (2D) extent of the areas affected by changes is identified without providing any information on the corresponding elevation (3D) changes. The aim of this work is therefore to lay the foundations for the development of DL algorithms capable of automatically generating a (3D) elevation CD map together with a standard 2D CD map, using only bitemporal optical images as input and thus without the need to rely directly on elevation data during the inference phase. In particular, our work proposes two innovative networks capable of solving the 2D and 3D CD tasks simultaneously, and a new and freely available dataset, named 3DCD dataset designed precisely for this multitasking. Specifically, the dataset covers the urban area of Valladolid and is composed of: - 472 pairs of images cropped from optical orthophotos acquired through two different aerial surveys (performed respectively in 2010 and 2017); - 472 pairs of DSMs produced from the rasterization of the point clouds acquired through two different LiDAR flights covering the same area and years (2010 and 2017) of the orthophotos; - the corresponding 472 2D CD maps in raster format; - the corresponding 472 3D CD maps -- the elevation variation maps -- in raster format. The proposed architectures consist of a deep convolutional neural network (CNN) (SUNet) and a transformer-based network (MTBIT). Encouraging results are shown on the 3DCD dataset by comparing the proposed architectures with other networks specifically designed to solve the 2D CD task and modified by us to also solve the 3D CD task. In particular, MTBIT achieves a normalized root-mean-square error (nRMSE) of 6.33 m, a change root-mean-square error (cRMSE) of 5.60 m and an F1 score of 62.80 \% - the best performance among the compared architectures - with a limited number of parameters (13.1 M). Conversely, SUNet achieves an nRMSE of 6.48 m, a cRMSE of 5.72 m and an F1 score of 58.80 \% but with a significantly higher number of parameters (35.7 M). The results are promising considering that the proposed network architectures were trained on a few images. Finally, we implemented innovative data augmentation techniques to improve the training phase and thus performance. We only used MTBIT, the network that gave the best results. We developed different data augmentation techniques according to the specific task the network has to perform, and the results obtained are quite encouraging. The quantitative evaluation shows that each new technique has a positive impact on reducing nRMSE and cRMSE values. The most efficient method is the combination of individual strategies. In particular, the best is the Crop or Resize strategy coupled with Gaussian noise on the 3D map, which achieves nRMSE values of 5.88 m, cRMSE values of 5.34 m and an F1 score of 65.16%. The 3DCD code and dataset are available at https://sites.google.com/uniroma1.it/3dchangedetection/home-page.