基于BP神经网络的竹林遥感监测研究

Remote sensing image based bamboo forest monitoring with a back propagation(BP) neural network

竹林信息提取对利用遥感技术估算竹林碳储量至关重要,高精度地提取竹林信息将有利于降低碳储量估算误差。借助Matlab神经网络模块,采用BP神经网络(back propagation neural network)对ETM+(enhanced themativ mapper plus)遥感影像提取竹林信息,得到了较高的精度,生产精度和用户精度分别为84.04%和98.75%;同时比较了Levenberg-Marquardt BP算法函数(Trainlm)、自适应学习率BP的梯度递减函数(Traingda)和梯度下降动量BP算法函数(Traingdm)等3种训练函数在分类中的差异。分析表明,Traingda算法函数分类精度最高,而Trainlm算法函数的训练时间最短。

To estimate the carbon content of bamboo forest based on remote sensing, highly accurate data acquisition is necessary to reduce estimation errors. In this study, enhanced thematic mapper plus（ETM＋） remote sensing data was used to extract bamboo forest data using a back propagation （BP） neural network. Matlab program language（Version 7.1） was used to compile the classification algorithm with algorithms of three training functions being compared; namely, Traingda-gradient descent backpropagation with adaptive learning rate backpropagation; Trainhn-levenberg-marquardt backpropagation; and Traingdm-gradient descent with naonaentum backpropagation. Results showed that for bamboo forest the BP neural network had a high classification accuracy with a producer accuracy of 84.0% and a user accuracy of 98.7%. Meanwhile, of the three different training functions Traingda had the highest classification accuracy, whereas Trainhn had the shortest training time.