The current methods to assess damage which has occurred during a disaster are usually manual based assessments. This work talks about developing an automated damage assessment model using computer vision and deep learning techniques. This helps to provide rapid relief during disasters to the affected areas. Satellite images provide real time photos of the affected area and help in visualizing large areas. Integrating satellite images with computer vision and deep learning techniques for damage assessments reduces time and manual labor for damage assessment. The thesis focuses on creating a novel deep learning model which can classify the damage using pre and post satellite images of the disaster. To localize the damage we have used GRAD CAM visualization method. For the model to be deployed on satellites for real time damage detection and classification, pruning and quantization is proposed to compress the size of the model. The models are tested on XBD dataset and compared with existing state of the art models. The thesis proposes basic model, patch based model, vision transformer, cross fusion and cross stitch models to test the damage classification. The XBD dataset is pre processed using image registration techniques. Multi scale inputs are used to train models over various scale of images. Further GRAD CAM is used to detect the damaged areas on post disaster images. Pruning and quantization methods are used to reduce size of models.