AI Framework for Urban and Natural Scene Classification Using Multiscale Attention for SAR/Multispectral Big Data

Urban and natural scene classification has drawn increasing interest in recent years and has been used in various remote sensing applications, such as forest monitoring, planning for cities and controlling vegetation and land use. In order to make better use of the knowledge that already exists, a novel attention based classification is suggested as a solution to the shortage of annotated data and lack of generalization capacity of existing models. While the conditional-distribution holds similar significance in practical applications, current adaption techniques for cross scenes evaluation solely take into account the marginal distribution. Furthermore, deep learningbased methods align the feature distribution taken from a singlescale structure, which results in information loss. An Attention based deep learning module is suggested for cross-scene classification tasks in order to get around the aforementioned problems. Conditional distributions (CD) is also used with a marginal distribution (MD) in the suggested neural network. Additionally, more information is extracted using the multiscale technique and the attention mechanism. Comparing experimental findings using So2Sat LCZ42 dataset, it is clear that attention mechanism performs noticeably better than the baseline model.