A method for extracting the leaf litter distribution area in forest using chip feature. Issue 15 (18th August 2018)

• Record Type: Journal Article
• Authors: Zhou, Xun; Sun, Zhongping; Liu, Suhong; Yu, Peixin; Wang, Xikai; Wang, Yuebin
• Is Part Of: International journal of remote sensing. Volume 39:Issue 15/16(2018)
• Page Start: 5310
• Page End: 5329
• Publication Date: 2018-08-18
• Subjects: Remote sensing -- Periodicals; Télédétection -- Périodiques; 621.3678
• DOI: 10.1080/01431161.2018.1484965
• Languages: English
• ISSNs: 0143-1161

Abstract

ABSTRACT: The leaf litter layer is an important surface coverage in the forest ecosystem, which imposes an important influence on soil and water conservation and affects the soil physical and chemical characteristics as well as the ecological environment. It is also a significant factor in estimating vegetation coverage. Currently, most remote-sensing methods to obtain leaf litter target are based on the analysis of the spectral features. However, the leaf litter and the soil background are easily confused. The traditional image processing method based on the pixel spectral information alone fails to make full use of the texture and shape information of the target and diminishes the extraction effect. With the research object focused on the continuous distribution area of leaf litter, whose features are intra-class complicated and inter-class boundary blurred, this article proposes a method to extract the leaf litter distribution area in forest by using the Convolutional Neural Network (CNN) to automatically retrieve the image block (which is also called image chip) features. The ResNet50, VGG16, and VGG19 models based on CNN are, respectively, used to extract and to learn the features of the image chips so that they can automatically obtain the deeper, more abstractive and representative features. The accuracy of this method in leaf litter target extraction has reached 95.1% which demonstrated that this new method has higher classification accuracy and better generalization ability than the traditional method of manually selecting image features. Moreover, the use of pretraining models can solve the problem in deep learning with small sample size. Through the study of scale effects, we also found the appropriate segmentation scale for the chips.