基于无人机高光谱的荒漠草原地表微斑块分类研究

Surface Micro-patches Classification on Desert Grasslands Based on UAV Hyperspectral Data

草原荒漠化会严重破坏草原生态平衡,荒漠草原地物分类已成为草原监测管理的关键问题。本文通过构建无人机高光谱遥感系统,解决了原有草原调查方式上效率低与空间分辨率不足问题;构建高分辨率图像卷积神经网络(HR-CNN)解决了荒漠草原地表微斑块精细化分类问题;与ResNet34、GoogLeNet、常规卷积神经网络模型进行对比,总体上HR-CNN模型表现更优,总体分类精度与Kappa系数分别为98.27%、96.63。在相同迭代次数条件下,模型构建速度上,HR-CNN相较其它三类模型分别提升65.88%、65.71%、13.77%。模型内存占有量上,HR-CNN相较其它三类模型分别降低92.11%、79.21%、43.64%。该网络模型是轻量化卷积在荒漠草原地物分类研究中的有效探索,可为后续草原地物分类提供新思路。

The ecological balance of grasslands is seriously affected by grassland desertification,and the classification of desert grassland land cover has become a key issue in grassland monitoring and management.In this study,the problem of low efficiency and insufficient spatial resolution in traditional grassland investigation methods was addressed by building an unmanned aerial vehicle hyperspectral remote sensing system.The issue of fine-grained classification of surface micro-patches in desert grasslands was also tackled by constructing a high-resolution image convolutional neural network(HR-CNN).Compared with ResNet34,GoogLeNet,and conventional convolutional neural network models,the HR-CNN model demonstrated superior overall performance.The overall classification accuracy and Kappa coefficient were 98.27%and 96.63%,respectively.Under the same number of iterations,the HR-CNN model was 65.88%,65.71%,and 13.77%faster in model construction speed than the other three models,respectively.Moreover,HR-CNN required a 92.11%,79.21%,and 43.64%smaller memory footprint than the other three models,respectively.The network model constructed in this study is an effective exploration of lightweight convolution in the study of desert grassland land cover classification,which provides new ideas for subsequent grassland land cover classification research.