Intelli-Change Remote Sensing - A Novel Transformer Approach

Change detection is becoming increasingly popular in contemporary practical applications, examples include monitoring urban growth, calculating the extent of deforestation, and environment assessment.In this research we created a model called IntelliChange-RSCC for remote sensing image change captioning where it primarily consists of two important components. The first involves utilising ResNet as the base to produce features from the images, and the next component is a Transformer model which contains Encoder and Decoder. As part of Encoder Transformer, we are having four main modules:1) Image Change Detector (ICD) which subtracts the two image features to get the image difference; 2) Cross Encoder(CE), It acts as a modified transformer encoder; 3) Conceptual Tokenizer(CT) where it extracts semantic tokens from the bi-temporal images and 4) Sequential Temporal Merging(STM) this focuses on relevant changes rather than irrelevant changes. For this study we built a new dataset named LEVTN-CC Dataset by modifying existing LEVIR-CC Dataset. Our model IntelliChange-RSCC performed better than other competing models in the field of remote sensing image change captioning with $+5\%$ on BLEU-1 and $+15\%$ on CIDEr-D. The results of the attention visualisation also imply that our model can ignore unimportant changes and focuses on those that are of interest.