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, CD developed so far are mainly focused on generating two-dimensional (2D) maps where planimetric extent areas affected by changes is identified without providing any information corresponding elevation variations. The aim this work is, hence, to establish basis for development DL able automatically generate an (3D) map along with standard 2D map, only bitemporal optical images as input, and thus need rely directly data during inference phase. Specifically, our proposes novel network, capable solving 3D tasks simultaneously, modified version 3DCD dataset, freely available dataset designed precisely twofold task. proposed architecture consists Transformer network based semantic tokenizer: MultiTask Bitemporal Images (MTBIT). Encouraging results, obtained comparing other networks specifically solve task, shown. particular, MTBIT achieves metric accuracy (represented changed root mean squared error) equal 6.46 m – best performance among compared architectures limited number parameters (13,1 M). code at https://sites.google.com/uniroma1.it/3dchangedetection/home-page.