【目的】植被检测是城市生态研究的重要手段,然而由于遥感图像中植被存在阴影区域、遮挡区域以及色彩上的畸变等,导致当前的植被检测精度较低。基于遥感卫星影像,采用深度学习技术快速有效地检测出城市中的植被区域,为植被资源统计等相关研究提供依据。【方法】选用深度卷积神经网络模型,对高分辨率遥感影像中的植被区域进行检测。对不同的优化器,通过设置不同的卷积核大小,对精度进行对比分析。最后对网络层数进行研究,对设置合适网络层数进行分析,用构造的深度卷积神经网络在实验数据上进行植被区域检测。【结果】利用卷积神经网络处理二维图像时,无需手动提取特征,进行简单少量的预处理后,直接把图像输入到CNN模型中进行训练,即可实现图片的识别分类功能。降低了预处理的难度,同时局部感知和权值共享大幅度地减少了参数量,加快了计算速度。次抽样还能保证图像处理后的平移、旋转、缩放和拉伸的不变性。解决了传统方法计算量和样本量大、结构复杂以及费时的缺点。在采集到的高分辨率紫金山区域的遥感图像中,通过设计的多层卷积神经网络模型对区域中的植被资源进行分析,对比和研究不同的优化器、卷积核和网络层数,植被检测精度达到95.4%,明显高于当前众多植被检测算法。【结论】在深度学习中,目标检测的精度依赖于网络的结构设置,通过对优化器、卷积核以及网络层数进行设定,可以明显提高目标检测效率和精度。

【Objective】 Vegetation detection is an important aspect of urban ecological research. However, because of shadowed areas, occluded areas and color distortion of vegetation in remote sensing images, the current vegetation detection accuracy is low. The purpose of this study is to efficiently and effectively measure the vegetation area in a city based on remote sensing satellite images and deep learning technology to provide algorithms for relevant research, such as vegetation resource statistics. 【Method】 Initially, a deep convolution neural network (CNN) model was used to detect vegetation areas in high-resolution remote sensing images. Subsequently, different optimizers were analyzed and discussed, and the accuracy was compared by setting different convolution kernel sizes. Finally, the number of network layers was studied, the appropriate number of network layers was analyzed, and the present deep convolution neural network was used to detect the vegetation area in the experimental data. 【Result】 Experimental data were obtained from high-resolution remote sensing images of the Zijin Mountain area. The vegetation resources in the area were analyzed using the designed multilayer convolution neural network model, and different optimizers, convolution cores and network layers were compared and studied. The vegetation detection accuracy reached 95.4%, which is significantly higher than that of many current vegetation detection algorithms. 【Conclusion】 It can be concluded that the accuracy of the target detection depends on the structure of the convolutional neural network. By setting the optimizer, convolution kernel and number of network layers, the efficiency and accuracy of target detection can be significantly improved.