Data Augmentation in Remote Sensing Image Change Captioning

Remote Sensing Image Change Captioning (RSICC) involves the automatic generation of descriptive captions for land cover transformations observed in bitemporal remote sensing images. One of the major challenges in RSICC is the limiting availability of training data. Data augmentation, a technique to artificially increase the size of the training dataset, has been underexplored in RSICC despite its potential to improve model performance. This paper proposes using basic data augmentation methods that are tailored for RSICC. Additionally, a hybrid augmentation approach is explored, combining multiple transformations to create diverse training data. Our experimental evaluation employs the LEVIR-CC dataset, assessing the impact of these augmentation techniques on the performance of the Progressive Scale-Aware Network (PSNet). Results demonstrate that the proposed techniques significantly improve captioning performance, with the hybrid method yielding the most substantial enhancements. Our findings highlight the effectiveness of tailored data augmentation in improving RSICC model performance and suggest that further exploration of augmentation strategies could contribute to the field. The implementation of the data augmentation approaches from this study are published at https://github.com/ChangeCapsInRS/Data-Augmentation-in-RSICC.