Hash Coding For Classification Of Remote Sensing Images

With fast development of technology, more and more hyperspectral remote sensing data needs to be dealt with. There is an urgent demand for a kind of intelligent algorithm to analyze and compute these image data. On the other hand, in recent years, learning to hash has become a popular machine learning method on large data.This paper proposes a dimension reduction method based on learning to hash and segment hashing methods, and applies them on the classification problem of hyperspectral remote sensing image. This method can obtain acceptable classification precision with a great reduction of calculation time. Further, we propose block hashing method which combines segment hash method and block classification method to improve the classification accuracy. We successfully applied our methods on four different representative hyperspectral remote sensing image data sets to obtain better classification results. On one hand, the two hashing methods can achieve similar classification accuracies as the corresponding methods without hashing, namely naive method and block method. On the other hand, it can greatly accelerate the calculation speed, and the accelerating rate can be even 10 times in some data sets. Further, we systematically compare the four methods in classification accuracy and calculation speed, including naive method, segment hash method, block method, and block hash method. Finally, we discuss the impact of the model parameters and different Hash functions forms on accuracy and calculation speed.In conclusion, based on our results, the hash methods can not only have good performances on hyperspectral remote sensing images, but also can be widely applied on general classification problems.