Object classification of remotely sensed images using deep learning

Abstract

Object classification of remotely sensed images is one of the fundamental tasksto perform various operations such as land planning, traffic monitoring, urbanmonitoring, disaster control and weather forecasting, etc. There are many machinelearning algorithms that use neural networks but still it is a very challengingtask because the learning of large sizes of high resolution remotely sensed imagesis very time consuming. In our proposed approach, we have used reduced sizehand-crafted Speed-Up Robust Features (SURF) as input to deep neural networkrather than an original image as used by other researcher. Feedforward and backpropagationalgorithms which are jointly known as feedforward neural networkalgorithm is used to learn the weights of the four-layer deep neural network accordingto images of different objects applied at the input of the deep neural network.Our approach reduces the computational complexity and also providesbetter classification than the previous approaches using deep neural network ondifferent high resolution imagery.