부산대학교 도서관

Satellite imagery has revolutionized earth's civilization. It allowed human beings to explore many hidden wonders of earth, it allowed to uncover many unexplored places and discovered Iceland's, volcanoes, forests, natural resources and many more. Satellite imagery has improved exponentially in the millennium. Satellites are used to take images of earth and stitch it to form maps and helped people to navigate, many technologies used in imaging and positioning to improvise and sensors are used to accurate imaging like LISS, however sensors have a hardware limitation. There are many factor which affect the clarity of imaging like mist, light pollution, air pollution, clouds, which masks the actual image, and that's where image enhancement comes into picture. Existing systems use neural networks to enhance the image which uses model VGG19 which has error rate up to 7.5% and it has only 60% accuracy. Satellite image enhancement is divided into 2 segments, upscaling and de-hazing, image is loaded and pixels are segmented to have compatibility and work optimization. Earth image is very vast and complex, dedicated image enhancement for particular place is not possible so neural network is selected which can detect the place automatically and enhance it. Resnet is the model which can be used efficiently in this context, it results in better accuracy 70%-80% and has very less error rate. And for haze removal dark channel prior is used which is efficient and works faster without mat lab. The results shows positive improvement for image enhancement for almost all type of image with more efficiency. This can be done by comparing with the trained datasets present in landsat and present data. The landsat dataset consists of multiple satellite imagery which has been cropped. The result has showed increased accuracy and also haze level has been reduced from 43% to 35% in de-hazing. The proposed system has over 80% efficiency.