CNN影像输入尺寸和分辨率对川西针叶林植被分类精度的影响

Effects of image input size and resolution by CNN on the classification accuracy for coniferous forest vegetation in western Sichuan

川西亚高山针叶林位于中国西南地区,受多云、多雨、多雾的影响,难以通过卫星影像进行植被分类的研究。为了解决这一难题,本研究选取川西亚高山针叶林的典型区域王朗自然保护区作为研究区,使用多旋翼无人机获取研究区域北部高分辨率RGB影像,结合卷积神经网络进行植被分类;为进一步挖掘卷积神经网络在无人机遥感影像上的潜力,选择语义分割方法(U-Net)进行分类,并根据不同分辨率的无人机影像和不同尺寸下的样本集构建植被分类模型,建立森林指纹库。结果表明:(1)结合无人机可见光影像和CNN模型进行分类能够获得高精度分类结果。在空间分辨率为5 cm,尺寸为256×256像素的情况下达到最优,总体精度为93.21%,Kappa系数为0.90;(2)选择合适的尺寸大小能够提高模型的分类精度。在5 cm的空间分辨率下,尺寸为128×128像素的模型总体精度为82.30%,Kappa系数为0.76;尺寸为256×256像素的模型总体精度为93.21%,Kappa系数为0.90;(3)超高空间分辨率的升高对模型精度的提升是有限的。当空间分辨率从10 cm升到5 cm时,模型的总体精度提高了0.02,Kappa系数提高了0.03,模型的分类精度并没有明显提升。(4)对于区域内代表性不足的植被类型来说,受空间分辨率和尺寸大小的影响要远高于区域内优势树种,特别是空间分辨率的影响最大。在20 cm的空间分辨率下落叶灌木的生产者精度和用户精度均低于70%。综上,利用无人机高分辨率RGB影像结合CNN模型对川西亚高山针叶林的植被分类能够取得高精度分类结果,本研究可为该区域植被分类提供一种自动、准确的方法。

The subalpine coniferous forest in west Sichuan is located in southwest China,which is affected by cloudy,rainy,and foggy conditions.Thus,conducting vegetation classification in the area by using satellite images is difficult.(Objective)Therefore,this work selects Wanglang Nature Reserve,which is a typical area of subalpine coniferous forest in western Sichuan,as the study area.A multi-rotor UAV is used to acquire high-resolution RGB images of the northern part of the study area,and it is combined with a convolutional neural network model for vegetation classification.(Method)This study selects the semantic segmentation method(U-Net)for classification,constructs vegetation classification models based on UAV images of different spatial resolutions and sample sets under different tile sizes,and establishes a forest fingerprint library to further explore the potential of convolutional neural networks on UAV remote sensing images.(Result)(1)The combination of UAV visible images and convolutional neural network model for classification can obtain classification results of high accuracy,which reached the optimum at a spatial resolution of 5 cm and a size of 256×256.The overall accuracy was 93.21%,and the kappa coefficient was 0.90.(2)The increase in ultrahigh spatial resolution had limited improvement on the model accuracy.When the spatial resolution was increased from 10 cm to 5 cm,the overall accuracy of the model improved by 0.02 and the Kappa coefficient improved by 0.03,and the classification accuracy of the model did not improve significantly.(3)Choosing the appropriate size can improve the classification accuracy of the model.Under the spatial resolution of 5 cm,the overall accuracy of the model with the size of 128×128 was 82.30%and the kappa coefficient was 0.76,and the overall accuracy of the model with the size of 256×256 was 93.21%and the kappa coefficient was 0.90.(4)For the vegetation types that were underrepresented in the region,the influence of spatial resolution and tile size was much higher than that of the dominant tree species,especially the influence of spatial resolution was the highest.The producer and user accuracies for deciduous shrubs at a spatial resolution of 20 cm were below 70%.(Conclusion)This study shows that vegetation classification of subalpine coniferous forests in western Sichuan using UAV high-resolution RGB images combined with convolutional neural networks can achieve high-precision classification results.The effects of UAV spatial resolution and tile sizes on the accuracy of convolutional neural network models are explored,which further details the potential of convolutional neural networks on UAV high-resolution RGB images to provide an automatic and accurate research method for vegetation classification in this region.