Inferring 3D change detection from bitemporal optical images

In recent years, change detection (CD) using deep learning (DL) algorithms has been a very active research topic in the field of remote sensing (RS). Nevertheless, the CD algorithms developed so far are mainly focused on generating two-dimensional (2D) change maps where the planimetric extent of the areas affected by changes is identified without providing any information on the corresponding elevation variations. The aim of this work is, hence, to establish the basis for the development of DL algorithms able to automatically generate an elevation (3D) CD map along 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. Specifically, our work proposes a novel network, capable of solving the 2D and 3D CD tasks simultaneously, and a modified version of the 3DCD dataset, a freely available dataset designed precisely for this twofold task. The proposed architecture consists of a Transformer network based on a semantic tokenizer: the MultiTask Bitemporal Images Transformer (MTBIT). Encouraging results, obtained on the modified version of the 3DCD dataset by comparing the proposed architecture with other networks specifically designed to solve the 2D CD task, are shown. In particular, MTBIT achieves a metric accuracy (represented by the changed root mean squared error) equal to 6.46 m - the best performance among the compared architectures - with a limited number of parameters (13,1 M).

Automated summary

In recent years, deep learning algorithms for change detection in remote sensing have been an active research area. However, existing algorithms mainly focus on generating 2D change maps without considering elevation variations. This study proposes a novel network, MTBIT, capable of generating both 2D and 3D change maps using bitemporal optical images as input, without relying on elevation data. The results show that MTBIT achieves the best performance among the compared architectures in terms of metric accuracy (6.46m) with a limited number of parameters (13.1M).