Combination of 2D-CNN and random forest models for crop classification with UAV imagery

초록

Recently, a convolutional neural network (CNN) has been regarded as an effective deep learning model that can extract spatial contextual information without user's intervention for classification. However, to extract useful spatial features may be difficult from the CNN model when limited training data are used for supervised learning. In this case, if the simple application of softmax activation functions, the final classification may not lead to satisfactory classification performance due to less informative spatial features. As an alternative, conventional machine learning algorithms can improve the classification performance because more sophisticated algorithms than the softmax operator are applied to the final classification. In this paper, a hybrid model is presented that combines two dimensional CNN (2D-CNN) and random forest (RF). Spatial contextual information extracted from 2D-CNN is used as input features of RF-based classification. To evaluate the potential of the hybrid model for crop classification, a case study of crop classification with unmanned aerial vehicle images was carried out. The classification performance of the hybrid model proposed in this study was superior to those of 2D-CNN and RF classifiers, implying the effectiveness of the proposed model when small training data are used for supervised classification. © 2020 40th Asian Conference on Remote Sensing, ACRS 2019: Progress of Remote Sensing Technology for Smart Future. All rights reserved.

키워드