Xi'an Institute of Optics and Precision Mechanics, Chinese Academy of Sciences

Scientists Build Novel Methods to Enhance the Accuracy of Remote Sensing Scene Classification

Data：11-03-2020 | 【 A A A 】 | 【Print】【Close】

Remote sensing scene classification (RSSC) refers to inferring semantic labels based on the content of the remote sensing scenes. The development of the RSSC technology can contribute to target detection, change detection. Remote sensing scenes contain multiple land-cover units with different sizes or same land-cover units.

Nevertheless, the intermediate features in RSSC are aggregated by some unsupervised feature encoding methods. Little attention has been paid to explore the information of semantic labels for the feature aggregation.

A researcher group which led by Prof. Dr. LU Xiaoqiang from Xi’an Institute of Optics and Precision Mechanics (XIOPM) of the Chinese Academy of Science (CAS) propose a supervised convolutional feature encoding module to utilize the information of semantic labels to build a convolutional representation for RSSC. It is feasible to embed the supervised convolutional feature encoding module into the CNN to build the end-to-end feature aggregation CNN (FACNN) for the joint training. The results were published in IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING. (https://ieeexplore.ieee.org/document/8730481)

In the research, the VGG-16 is employed as the backbone pipeline to learn the intermediate features by taking advantages of its superiority in the RSSC task.

Convolutional representation is further merged with the second FC feature of the backbone pipeline to generate the discriminative scene representation.

Finally, the classic softmax classifier is employed to obtain the semantic labels from the scene representations. In the training phase, the information of semantic labels can be back-propagated to supervise the feature aggregation and the feature learning.

Furthermore, the fine-tuning strategy is employed to alleviate the problem that CNN cannot be trained from scratch with insufficient training samples on the existing remote sensing scene databases.

This will further promote the accuracy of remote sensing scene classification.

Overall structure of the proposed FACNN ( Image by XIOPM ).