Using mask R-CNN in remote sensing images

Thanks to various remote sensing satellites, a very large amount of data is obtained today. It is possible to access free data such as Sentinel and remotely sensed data with very high spatial resolution. It is challenging to classify remotely sensed images with such large volumes and complex earth details using traditional methods. Fast deep learning models such as Mask R-CNN can be trained once with a large amount of training data, allowing images of different regions to be classified within seconds. In this study, classification studies were carried out on Sentinel-2 and Worldview-3 satellite images using Mask R-CNN, one of the newest convolutional neural network models. A total of 1515 images and 18811 labels were created to create the data set. Two different scenarios were carried out on the created data sets. In the first of these scenarios, the region was determined by using Sentinel-2 satellite images. ResNet101 and ResNet50 network architectures were used while detecting built up area in the Sentinel-2 satellite image. 81% and 79% accuracy were obtained from these created models, respectively. Five class labeled data sets were created for land cover classification on the Sentinel-2 satellite image, and as a result, 74% accuracy was obtained. In the second scenario, the building was detected using Worldview-3 satellite images and 83% accuracy was obtained. Obtained results show that Mask R-CNN can be used successfully in classification of satellite images. The use of this method in satellite images is important because it allows instant classification.