基于U-Net 的甘蔗提取方法

Extraction of sugarcane from Google Earth image based on the U-Net model

关键词:  遥感  卷积神经网络  信息提取  甘蔗

Key words:  remote sensing  convolution neural network  information extraction  sugarcane

摘要:

【目的】基于深度学习的分类方法是使用高分辨率遥感影像快速提取作物种植空间信息的新方法。【方法】以云南省陇川县甘蔗种植园为研究区，收集空间分辨率为0.5m的Google Earth 开放影像进行数据预处理，建立样本数据集，构建U-Net神经网络模型，训练模型参数；使用U-Net 模型提取甘蔗种植空间信息，通过地面样方数据验证甘蔗提取精度。【结果】（1）基于深度学习方法的甘蔗分类总体精度和Kappa系数分别为92.76%和0.848 0，面积总精度为94.41%；平坝区、丘陵区分类精度存在差异，总精度和Kappa系数分别为97.10%、0.922 1 和88.42%、0.767 3；（2）受部分地物RGB 影像特征与甘蔗相似的影响，分类结果存在错分现象。【结论】基于U-Net神经网络模型的方法可用于高分辨率影像的甘蔗提取，更准确的分类精度还有待进一步研究和验证。

Abstract:

［Purpose］To extract the desired information from remote sensing big data requires efficient data processing capabilities. In order to realize the extraction of crop planting spatial information by using high-resolution remote sensing images rapidly，a method of sugarcane classification based on deep learning was preliminary studied.［Method］In this study，part of the sugarcane planting area in Longchuan county which is located in Dehong prefecture was selected as the study area. Subsequently，the classification of sugarcane was performed in ENVI5.5 through the deep learning model which is based on the U-Net architecture. Firstly， the Google Earth image on November 28，2018 was collected and resampled into an image with a 50 cm spatial resolution and an 8 km×8 km area. Secondly，the regions of interest（ROI） about sugarcane were selected and the ground sample verification data was produced by manual interpretation. Training work about U-Net deep learning classification model parameters were completed in ENVI5.5. Finally，sugarcane classification and spatial information extraction were completed by using the trained deep learning classifier，and the accuracy of the method was verified preliminary.［Result］First，the overall accuracy and Kappa coefficient of sugarcane classification based on deep learning method were 92.76% and 0.848 0，and the area accuracy was 94.41%. Second，there were differences in classification accuracy among different regions. The total accuracy and Kappa coefficient of the flat dam area were 97.10% and 0.922 1，much higher than the hilly area which had the results of 88.42% and 0.767 3. In addition，the miss classified pixels in the hilly area were much more than flat dam. Third，partial misclassification were existed in the classification results by using Google Earth images，because it is difficult to distinguish in RGB images for some objects features are similar to sugarcanes. ［Conclusion］This study showed that U-Net neural network model can be used to extract sugarcane spatial information from high resolution images. ［Purpose］To extract the desired information from remote sensing big data requires efficient data processing capabilities. In order to realize the extraction of crop planting spatial information by using high-resolution remote sensing images rapidly，a method of sugarcane classification based on deep learning was preliminary studied.［Method］In this study，part of the sugarcane planting area in Longchuan county which is located in Dehong prefecture was selected as the study area. Subsequently，the classification of sugarcane was performed in ENVI5.5 through the deep learning model which is based on the U-Net architecture. Firstly， the Google Earth image on November 28，2018 was collected and resampled into an image with a 50 cm spatial resolution and an 8 km×8 km area. Secondly，the regions of interest（ROI） about sugarcane were selected and the ground sample verification data was produced by manual interpretation. Training work about U-Net deep learning classification model parameters were completed in ENVI5.5. Finally，sugarcane classification and spatial information extraction were completed by using the trained deep learning classifier，and the accuracy of the method was verified preliminary.［Result］First，the overall accuracy and Kappa coefficient of sugarcane classification based on deep learning method were 92.76% and 0.848 0，and the area accuracy was 94.41%. Second，there were differences in classification accuracy among different regions. The total accuracy and Kappa coefficient of the flat dam area were 97.10% and 0.922 1，much higher than the hilly area which had the results of 88.42% and 0.767 3. In addition，the miss classified pixels in the hilly area were much more than flat dam. Third，partial misclassification were existed in the classification results by using Google Earth images，because it is difficult to distinguish in RGB images for some objects features are similar to sugarcanes. ［Conclusion］This study showed that U-Net neural network model can be used to extract sugarcane spatial information from high resolution images.