生成式对抗网络的土壤有机质高光谱估测模型  

Hyperspectral Estimation Model of Soil Organic Matter Content Using Generative Adversarial Networks

作　　者：何少芳[1] 沈陆明[1] 谢红霞[2] HE Shao-fang;SHEN Lu-ming;XIE Hong-xia(College of Information and Intelligence,Hunan Agricultural University,Changsha 410128,China;College of Resources&Environment,Hunan Agricultural University,Changsha 410128,China)

机构地区：[1]湖南农业大学信息与智能科学技术学院,湖南长沙410128 [2]湖南农业大学资源环境学院,湖南长沙410128

出　　处：《光谱学与光谱分析》2021年第6期1905-1911,共7页Spectroscopy and Spectral Analysis

基　　金：国家科技基础性工作专项(2014FY110200);国家自然科学基金项目(61973111);湖南省教育厅重点项目(19A242)资助。

摘　　要：已有的土壤有机质含量估测模型大多以光谱特征波段、线性和非线性模型为基础,较少考虑通过拓展样本数据建模集来提高模型的估测能力。为进一步提高土壤有机质高光谱反演模型估测精度,提出利用生成式对抗网络(GAN)合成伪高光谱数据和有机质含量的动态估测模型。选取湖南省长沙市及周边区域的水稻田为研究对象,采集土样和实测高光谱数据(350~2500 nm),室内化学测定有机质含量。以高光谱数据和有机质含量为基础,利用生成式对抗网络生成等量新数据,结合原始数据建模集组成增强建模集。在GAN正式训练中,每轮训练完成后,设置4个观测点(对应增强建模集中含50,100,150和239个生成样本),动态构建交叉验证岭回归(RCV)、偏最小二乘回归(PLSR)和BP神经网络(BPNN)土壤有机质含量估测模型(分别简称GAN-RCV,GAN-PLSR和GAN-BPNN),并在相同测试集上实施模型评估。实验结果表明:(1)原始数据建模集上拟合的估测模型中,交叉验证岭回归表现最佳,决定系数(R^(2))和均方根误差(RMSE)分别为0.8311和0.1896;(2)GAN的150轮正式训练中,增强建模集上动态构建的GAN-RCV,GAN-PLSR和GAN-BPNN模型性能显著提高,具体表现为:GAN-RCV的R^(2)取得最大值0.8909(RMSE 0.1537)、最小值0.8505(RMSE 0.18)与平均值0.8687(RMSE 0.1686),最大R^(2)比建模集上拟合的RCV提高了7.2%(RMSE降低了18.9%),GAN-PLSR获得R^(2)最大值0.8554(RMSE 0.1769)、最小值0.7270(RMSE 0.2432)与平均值0.7801(RMSE 0.2177),最大R^(2)比建模集上拟合的PLSR提高了20.6%(RMSE降低了29.5%),GAN-BPNN表现最佳,R^(2)取得最大值0.9052(RMSE 0.1433)、最小值0.8017(RMSE 0.2073)与平均值0.8681(RMSE 0.1686),最大R^(2)比建模集上拟合的BPNN提高了30.8%(RMSE降低了44.5%);(3)随着增强建模集中生成样本数量增加,模型精度提升效果呈先升后降趋势,4个观测点中第3个观测点的模型性能提升最显著。充分的实验表明:基于In the previous study of the estimation model of soil organic matter content,most models were based on the feature bands,linear and non-linear empirical models rarely explored the ability promotion using an extended modeling dataset.To further improve the performance of the estimation model,it proposed a dynamic estimation model of soil organic matter content using generative adversarial networks(GAN)to generate the pseudo hyperspectral and organic matter content.Paddy soil samples and hyperspectral data(350~2500 nm)were collected from Changsha and its surrounding areas of Hunan Province,and the organic matter content was monitored chemically.Based on these data,equivalent new samples were generated by GAN and combined with the modeling set to form anenhanced modeling set.After completing each epochformal training of GAN,the prediction models of soil organic matter content were dynamically constructed using cross-validation ridge regression(RCV),partial least squares regression(PLSR)and BP neural network(BPNN)on four observation points(corresponding 50,100,150 and 239 generated samples in enhanced modeling set)(the abbreviation of models were GAN-RCV,GAN-PLSR and GAN-BPNN).The experimental results showed that:(1)Among the estimation models fitted on modeling set of the origin data,RCV was the best-performing model,whose determination coefficient(R^( 2))and root square error(RMSE)were 0.8311 and 0.1896;(2)In the 150 epochs formal training of GAN,the performance of GAN-RCV,GAN-PLSR and GAN-BPNN dynamically constructed on the enhanced modeling set were significantly improved,specific performances:R^( 2) of GAN-RCV obtained the maximum 0.8909(RMSE 0.1537),minimum 0.8505(RMSE 0.18)and mean 0.8687(RMSE 0.1686),the maximum R^( 2) increased by 7.2%(RMSE decreased by 18.9%)compared with RCV fitted on the modeling dataset,R^( 2) of GAN-PLSR had the maximum 0.8554(RMSE 0.1769),minimum 0.7270(RMSE 0.2432)and mean 0.7801(RMSE 0.2177),the maximum R^( 2) increased by 20.6%(RMSE decreased by 29.5%)than PLSR constructed on the mo

关 键 词：有机质 高光谱 生成式对抗网络 交叉验证岭回归 BP神经网络 

分 类 号：O657.3[理学—分析化学]