Convolutional neural network for GF-2 image stand type classification

ObjectiveThe classification of stand type based on remote sensing imagery is an important application in modern forestry. In recent years, many studies have explored this territory using multiple data sources and classification algorithm. Convolutional neural network (CNN), a new neural network algorithm, has higher accuracy in pattern recognition, scene classification and objective detection because of its unique structure and deep learning technology. The purpose of this paper is to propose a convolutional neural network system tailored for GF-2 (a high-resolution multispectral remote sensing data) applied in stand type classification on pixel-level.

MethodWe chose different image patch size (i.e. m = 5, 7, 9 and 11) for building CNN and multilayer perceptrons (MLP) as benchmark in tensorflow (an open source library of machine intelligence), to train and compare classification accuracy of model.

ResultExperimental results showed that the CNN (m = 9) outperformed MLP and CNNs (m = 5, 7 and 11) by 10.91%, 6.55%, 1.3% and 2.54%, respectively, in overall classification accuracy. And the CNN (m = 9) alleviates the effect of salt-and-pepper and boundary uncertainties greatly in visual assessment.

ConclusionCNN can fully exploit the spatial features of images while utilizing the spectral features of high-resolution images to improve classification accuracy. And on remote sensing image classification based on CNN, selecting the appropriate image patch size according to the data source and the features of the objective is the key measure to improve the classification accuracy and classification effect.