可迁移的土壤重金属污染高光谱定性分类方法研究
|
摘要
现有基于高光谱遥感的土壤重金属污染定性分类模型,大多采用同一地区室内光谱测定训练样本数据进行模型构建与测试。但室内光谱测定需要复杂的处理过程,成本高,效率低,且无法快速获得目标区域空间上连续的光谱信息。考虑到实际应用需求,模型在相同实验区和不同试验区野外光谱数据是否具有较好的迁移推广能力是目前迫切需要回答的问题。为回答这一问题,选取湖南省郴州市和衡阳市两铅锌矿区作为实验研究区,选用支持向量机(SVM)作为分类器,将郴州实验区室内采样的83个样本数据和衡阳实验区室内采样的46个样本数据分别用于分类器训练,将衡阳地区野外采样的46个样本数据用于分类测试。并首先通过基于联合分布适配(JDA)的迁移学习方法进行光谱变换以缩小两地室内外测定光谱分布差异,然后进行不同区域室内外土壤重金属污染定性分类模型迁移。实验结果表明:(1)由于野外测得的光谱数据会受到太阳辐射、提取的土壤成分差异等因素的干扰导致室内外光谱数据存在显著的分布差异,难以直接将基于室内采样数据训练得到的土壤重金属污染定性分类模型迁移到同一地区测定的野外高光谱数据上。但通过JDA变换缩小室内外分布差异后,模型迁移能力得到显著提升,砷(As)、铅(Pb)和锌(Zn)三种重金属含量是否超标的分类精度都达到了84%以上, Zn元素含量是否超标的分类精度甚至达到了89%以上。(2)由于季节性影响、地区成分的干扰和光谱噪声的增加,不同地区光谱数据存在着更为显著的分布差异,加大了不同地区土壤重金属污染监测的难度,难以将基于室内采样光谱数据所建立的土壤重金属定性分类模型直接迁移到其他地区野外采样数据上(平均分类精度仅在50%左右)。经过JDA迁移学习方法进行室内外光谱变换处理后,模型迁移能力得到保证,因此,室外光谱采样可直接用于研究不同试验区域重金属(As, Pb和Zn)的污染情况。
The current qualitative classification models of soil heavy metal content based hyperspectral remote sensing technology mostly use indoor measured spectral data from the same area for model training and testing. However, the indoor spectrum measurement requires a complicated processing process with high cost and low efficiency, and thus cannot obtain the spatially continuous spectral information in the target area quickly. Moreover, whether this kind of model can be transferred to the outdoor measured spectral data in different test areas is still unclear. In order to answer this question, two lead-zinc mining areas in Chenzhou City and Hengyang City of Hunan Province were selected as research areas. Support Vector Machine was used as classifier. Then 83 sample data from indoor sampling in Zhangzhou experimental area and 46 sample data from indoor sampling in Hengyang experimental area were used for classifier training, and 46 sample data from field sampling in Hengyang area were used for classification testing. The difference of spectral distribution between the indoor and outdoor measured spectral data was reduced by the transfer learning method based on joint distribution adaptation(JDA), and then the domain adaption model for two research areas was constructed. The experimental results show that:(1) The spectral data measured by outdoor samples may be affected by factors such as solar radiation and differences in extracted soil components, leading to the significantly spectral difference for indoor and outdoor samples. As a result, it is difficult to directly transfer the qualitative classification model of soil heavy metal pollution trained by indoor samples to the outdoor samples from the same area. However, after the reduction of indoor and outdoor distribution differences by JDA transformation, the transfer ability of the model has been significantly improved, and the classification accuracy of three heavy metals As, Pb and Zn has reached over 84%. The accuracy of classification of Zn elements exceeding the standard even reached 89%.(2) Due to seasonal influences, regional component interference, and spectral noise, there are even more significant differences in the distribution of spectral data in different areas. This further increases the difficulty of soil heavy metal pollution monitoring in different areas, and it is difficult to directly transfer the qualitative classification model of soil heavy metals based on indoor sampling spectral data to field sampling data in other areas(with an average classification accuracy of about 50%). After the indoor and outdoor spectral transformation processing by JDA, the transfer ability of the model has been greatly improved. Therefore, the outdoor spectral sampled can be directly used to investigate the pollution situation of heavy metals(As, Pb and Zn) in different test areas.
The current qualitative classification models of soil heavy metal content based hyperspectral remote sensing technology mostly use indoor measured spectral data from the same area for model training and testing. However, the indoor spectrum measurement requires a complicated processing process with high cost and low efficiency, and thus cannot obtain the spatially continuous spectral information in the target area quickly. Moreover, whether this kind of model can be transferred to the outdoor measured spectral data in different test areas is still unclear. In order to answer this question, two lead-zinc mining areas in Chenzhou City and Hengyang City of Hunan Province were selected as research areas. Support Vector Machine was used as classifier. Then 83 sample data from indoor sampling in Zhangzhou experimental area and 46 sample data from indoor sampling in Hengyang experimental area were used for classifier training, and 46 sample data from field sampling in Hengyang area were used for classification testing. The difference of spectral distribution between the indoor and outdoor measured spectral data was reduced by the transfer learning method based on joint distribution adaptation(JDA), and then the domain adaption model for two research areas was constructed. The experimental results show that:(1) The spectral data measured by outdoor samples may be affected by factors such as solar radiation and differences in extracted soil components, leading to the significantly spectral difference for indoor and outdoor samples. As a result, it is difficult to directly transfer the qualitative classification model of soil heavy metal pollution trained by indoor samples to the outdoor samples from the same area. However, after the reduction of indoor and outdoor distribution differences by JDA transformation, the transfer ability of the model has been significantly improved, and the classification accuracy of three heavy metals As, Pb and Zn has reached over 84%. The accuracy of classification of Zn elements exceeding the standard even reached 89%.(2) Due to seasonal influences, regional component interference, and spectral noise, there are even more significant differences in the distribution of spectral data in different areas. This further increases the difficulty of soil heavy metal pollution monitoring in different areas, and it is difficult to directly transfer the qualitative classification model of soil heavy metals based on indoor sampling spectral data to field sampling data in other areas(with an average classification accuracy of about 50%). After the indoor and outdoor spectral transformation processing by JDA, the transfer ability of the model has been greatly improved. Therefore, the outdoor spectral sampled can be directly used to investigate the pollution situation of heavy metals(As, Pb and Zn) in different test areas.
引文
[1] ZHANG Yin-ling,YAO Feng,NIU Ting,et al(张银玲,姚峰,牛婷,等).Environmental Protection of Xinjiang(新疆环境保护),2016,38(3):15.
[2] Wang F,Gao J,Zha Y.ISPRS Journal of Photogrammetry & Remote Sensing,2018,136:73.
[3] Shi T,Chen Y,Liu Y,et al.Journal of Hazardous Materials,2014,265(2):166.
[4] XIAO Jie-ying,WANG Yan,ZHANG Qian,et al(肖捷颖,王燕,张倩,等).Hubei Agricultural Sciences(湖北农业科学),2013,52(6):1248.
[5] Sun W,Zhang X,Zou B,et al.Remote Sensing,2017,9(6):632.
[6] Dong J,Dai W,Xu J,et al.International Journal of Environmental Research & Public Health,2016,13(7):640.
[7] Shi T,Liu H,Chen Y,et al.Sensors,2017,17(5):1036.
[8] TAO Chao,WANG Ya-jin,ZOU Bin,et al(陶超,王亚晋,邹滨,等).Spectroscopy and Spectral Analysis(光谱学与光谱分析),2018,38(6):1850.
[9] SHANG Xuan,LI Xi-can,XU You-you,et al(尚璇,李西灿,徐邮邮,等).Scientia Agricultura Sinica(中国农业科学),2017,50(8):1465.
[10] Piekarczyk J,Kazmierowski C,Krolewicz S,et al.IEEE Journal of Selected Topics in Applied Earth Observations & Remote Sensing,2016,9(2):827.
[11] Long M,Wang J,Ding G,et al.Transfer Feature Learning with Joint Distribution Adaptation,2013 IEEE International Conference on Computer Vision,2013.2200.
[2] Wang F,Gao J,Zha Y.ISPRS Journal of Photogrammetry & Remote Sensing,2018,136:73.
[3] Shi T,Chen Y,Liu Y,et al.Journal of Hazardous Materials,2014,265(2):166.
[4] XIAO Jie-ying,WANG Yan,ZHANG Qian,et al(肖捷颖,王燕,张倩,等).Hubei Agricultural Sciences(湖北农业科学),2013,52(6):1248.
[5] Sun W,Zhang X,Zou B,et al.Remote Sensing,2017,9(6):632.
[6] Dong J,Dai W,Xu J,et al.International Journal of Environmental Research & Public Health,2016,13(7):640.
[7] Shi T,Liu H,Chen Y,et al.Sensors,2017,17(5):1036.
[8] TAO Chao,WANG Ya-jin,ZOU Bin,et al(陶超,王亚晋,邹滨,等).Spectroscopy and Spectral Analysis(光谱学与光谱分析),2018,38(6):1850.
[9] SHANG Xuan,LI Xi-can,XU You-you,et al(尚璇,李西灿,徐邮邮,等).Scientia Agricultura Sinica(中国农业科学),2017,50(8):1465.
[10] Piekarczyk J,Kazmierowski C,Krolewicz S,et al.IEEE Journal of Selected Topics in Applied Earth Observations & Remote Sensing,2016,9(2):827.
[11] Long M,Wang J,Ding G,et al.Transfer Feature Learning with Joint Distribution Adaptation,2013 IEEE International Conference on Computer Vision,2013.2200.