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。因此,本文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,结合卷积神经网络模型进行植被分类。为进一步探究卷积神经网络在无人机遥感影像上的潜力,本文选择语义分割方法(U-net)进行分类,还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,并建立森林指纹库。结果表明：(1)结合无人机可见光影像和卷积神经网络模型进行分类能够获得高精度分类结果；在空间分辨率为5cm,尺寸大小为256×256的情况下达到最优,总体精度为92.35%,平均F1分数为91.99%,其中岷江冷杉的F1分数达到93.92%；(2)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10cm升到5cm时,总体精度和平均F1分数仅提高了0.02,模型的分类精度并没有明显提升。(3)选择合适的尺寸大小能够提高模型的分类精度。在5cm的空间分辨率下,尺寸为128×128的模型总体精度仅为75.59%,尺寸为256×256的模型总体精度为92.35%。这说明较小的裁剪尺度会使模型精度有明显的下降；(4)对于代表性不足的植被类型来说,低空间分辨率对模型精度的影响最大,在20cm的空间分辨率下落叶灌木的F1得分仅为62.45%。该研究表明利用无人机高分辨率RGB影像结合卷积神经网络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,并探讨了无人机空间分辨率和裁剪尺寸对卷积神经网络模型精度的影响,进一步挖掘了卷积神经网络在无人机高分RGB影像上的应用潜力,为该区域植被分类提供一种自动、准确的研究方法。

The subalpine coniferous forest in west Sichuan is located in southwest China, which is affected by cloudy, rainy and foggy conditions, making it difficult to research vegetation classification by satellite images. (Objective) Therefore, this paper selects Wanglang Nature Reserve, a typical area of subalpine coniferous forest in western Sichuan, as the study area, and uses a multi-rotor UAV to acquire high-resolution RGB images of the northern part of the study area, combined with a convolutional neural network model for vegetation classification. (Method) To further explore the potential of convolutional neural networks on UAV remote sensing images, this paper selects the semantic segmentation method (U-Net) for classification, and also 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. (Result) The results show that (1) the combination of UAV visible images and convolutional neural network model for classification can obtain high accuracy classification results, reaching the optimum at a spatial resolution of 5 cm and a size of 256×256, with an overall accuracy of 93.21% and a Kappa coefficient of 0.90. (2) The increase of ultra-high spatial resolution has 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, and explores the effects of UAV spatial resolution and tile sizes on the accuracy of convolutional neural network models, further tapping 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.