AutoRSISC: Automatic design of neural architecture for remote sensing image scene classification

Abstract

• A remote sensing image scene classification algorithm base on neural architecture search (NAS) is proposed. • A search strategy is proposed to search for the most appropriate neural architecture in the pre-defined search space. • A performance estimation strategy to evaluate the performance of search strategy. • The proposed method saves time and manpower while achieves better accuracy compared with the manually design method. Remote sensing image scene classification (RSISC) is one of the bases of visual question answering for remote sensing data (RSVQA). In the conventional deep learning based solutions of RSISC, domain experts have to manually design the neural architecture, which is costly in both time and manpower. In this paper, AutoRSISC, a high resolution remote sensing image scene classification method based on neural architecture search (NAS) is proposed. AutoRSISC algorithm samples the neural architecture in a certain proportion to reduce the redundancy in the search space which is based on the continuous relaxation of the neural architecture representation. Edge normalization is introduced to make architecture search more efficient. The conventional deep learning approaches take days or more just to manually design the appropriate neural architecture according to the current data set. The experimental results show that our method saves time and manpower compared with the existing manually designed RSISC methods. Our AutoRSISC algorithm took 7.6 h from the automatic design of the most appropriate neural architecture to the end of all experiments according to the UCM data set, and the classification accuracy rate of the final test set reached 97.85%. And our AutoRSISC algorithm took 101.7 h from the automatic design of the most appropriate neural architecture to the end of all experiments according to the NWPU45 data set, and the classification accuracy rate of the final test set reached 94.66%.