Restricted supervised Cascade Information Network for remote sensing change captioning with serial sentences

Remote sensing change captioning generally depicts land cover changes with a single sentence, which can hardly achieve throughout descriptions for complicated scenes. Although researchers have employed auxiliary tasks for more comprehensive outputs, the required costs are heavy. To solve this problem, we propose a Cascade Information Network (CI-Net) with low costs to obtain serial sentences. Specifically, we define each caption as a probability case related to deep features, which are updated in a designed Cascade Linguistic Module (CL-Module) by introducing the information theory. Afterwards, CI-Net measures implied information quantities contained in generated captions, while the captioning terminates when accumulated information quantities exceed a threshold. For better evaluation, we create a SEmantic Change capTION dataset (SECTION) with serial sentence annotations for each sample. Experimental results on the SECTION and public dataset validate the theoretical analysis and the effectiveness for CI-Net.