Satellite images for roads using transfer learning

Abstract

The challenge of identifying roads in very high resolution (VHR) satellite pictures is a major one in remote sensing. The purpose of this piece was to develop a model for automatically locating and labeling roads on a map as roads or non-roads. As originally conceived, this paradigm was to be a software tool. Hand-sort the satellite images. Automated separation is recommended over manual inspection because of the potential for human error. Together, a convolutional neural network and transfer learning were used to create the model. With a 97.6% success rate for MobileNetv2 and an 81.6% success rate for Alex's Net, the authors were able to identify photos using transfer learning. Given that each point in the spatial feature can now make a reference to any additional contextual data, the accuracy of the road segmentation is enhanced. Our models outperform the state-of-the-art models that have been previously published on the topic in the official Deep Learning Challenge. Our developed model also has a shorter training convergence time than the competing methods. The authors also offer empirical assessments of the best and worst times to deploy non-local blocks in the baseline model. Theoretically, a previously unknown best alternative identification approach was discovered as a result of the methods employed to train the model. The use of transfer learning on a rather sizable dataset allowed for this success. Yet another benefit of the design is that it incorporates data augmentation into its development.