The Method of Extracting Land Classification Information by HRI-RefineNET Model

Abstract

Accurate land use classification results play an important role in scientific research and production practice. The existing neural network structure contains many pooling layers. Although the model incorporated spatial location information when extracting ground object information, many detailed features were also lost. At the same time, due to the great difference between remote sensing image and natural image, the model with good performance of natural image extraction cannot be transplanted directly into the extraction of high-resolution remote sensing image. Based on RefineNET model, this paper proposes an improved HRI-RefineNet (High Resolution Image) model. The model chooses farmland, residential areas, forest, roads and water area as extraction targets, and establishes a new method for extracting land classification information from high resolution remote sensing images. The proposed HRI-RefineNet model adjust the convolution layer number and convolution kernel size respectively according to the texture features of different five kinds of ground objects. An improved pooling layer is designed for the HRI-RefineNet for reducing the loss of location information in feature maps. A decoder group is added to the HRI-RefineNet model. 29 remote sensing images of Tai'an city, China, from 2015 to 2017 were selected for experiment. The experimental results showed that the accuracy of farmland, residential areas, forest, roads and water area obtained by HRI-RefineNET model was 93.9%, 92.3%, 91.7%, 90.1%, 88.7%, respectively, superior to the RefineNET model and some existing advanced models.

Keywords

• Convolution
• image recognition
• land classification
• model structure
• neural networks
• remote sensing