农业工程学报
農業工程學報
농업공정학보
2015年
5期
107-114
,共8页
陈红艳%赵庚星%陈敬春%王瑞燕%高明秀
陳紅豔%趙庚星%陳敬春%王瑞燕%高明秀
진홍염%조경성%진경춘%왕서연%고명수
土壤%盐分%遥感%支持向量机%植被指数%反演%黄河口
土壤%鹽分%遙感%支持嚮量機%植被指數%反縯%黃河口
토양%염분%요감%지지향량궤%식피지수%반연%황하구
soils%salts%remote sensing%support vector machine%vegetation index%inversion%Yellow River estuary
快速获取土壤盐分的含量、特征及空间分布信息是盐渍土治理、利用的客观需求。该文针对黄河三角洲盐渍土，以垦利县为例，基于Landsat 8 OLI多光谱影像，在传统植被指数的基础上引入短波红外波段进行扩展，提出了改进植被指数；然后基于改进前后对应的植被指数，分别采用多元逐步回归（multivariable linear regression，MLR）、反向传播神经网络（back propagation neural networks，BPNN）和支持向量机（support vector machine，SVM）方法构建土壤盐分含量的遥感反演模型，并进行模型验证、对比和优选；最后基于最佳模型进行研究区土壤盐分含量的空间分布反演和分析。结果显示：相对传统植被指数，扩展后植被指数可增强与土壤盐分的相关性，大幅降低指数间的多重共线性；采用上述3种方法建模，改进后模型的精度比改进前都有提高，验证集决定系数R2提高0.04～0.10，均方根误差RMSE降低0.13～0.73，相对分析误差RPD提高0.25～0.34，改进后模型RPD均大于2.0，普遍达到性能良好；对比3种建模方法，SVM建模精度最高，BPNN模型次之，MLR分析精度最低，最佳模型为基于改进植被指数的土壤盐分含量支持向量机反演模型，建模集R2和RMSE为0.75、3.48，验证集R2、RMSE和RPD为0.78、3.02和2.56，模型较为准确、可靠；基于该模型反演的研究区土壤盐分含量整体较高，盐渍化程度空间分布表现为自西南部农业生产区至东北沿海区域逐渐加重，与实地调查一致。研究表明基于Landsat 8 OLI多光谱影像，引入第7波段对植被指数进行改进，从而构建土壤盐分含量的支持向量机模型，可获得较好的土壤盐分空间分布反演结果。
快速穫取土壤鹽分的含量、特徵及空間分佈信息是鹽漬土治理、利用的客觀需求。該文針對黃河三角洲鹽漬土，以墾利縣為例，基于Landsat 8 OLI多光譜影像，在傳統植被指數的基礎上引入短波紅外波段進行擴展，提齣瞭改進植被指數；然後基于改進前後對應的植被指數，分彆採用多元逐步迴歸（multivariable linear regression，MLR）、反嚮傳播神經網絡（back propagation neural networks，BPNN）和支持嚮量機（support vector machine，SVM）方法構建土壤鹽分含量的遙感反縯模型，併進行模型驗證、對比和優選；最後基于最佳模型進行研究區土壤鹽分含量的空間分佈反縯和分析。結果顯示：相對傳統植被指數，擴展後植被指數可增彊與土壤鹽分的相關性，大幅降低指數間的多重共線性；採用上述3種方法建模，改進後模型的精度比改進前都有提高，驗證集決定繫數R2提高0.04～0.10，均方根誤差RMSE降低0.13～0.73，相對分析誤差RPD提高0.25～0.34，改進後模型RPD均大于2.0，普遍達到性能良好；對比3種建模方法，SVM建模精度最高，BPNN模型次之，MLR分析精度最低，最佳模型為基于改進植被指數的土壤鹽分含量支持嚮量機反縯模型，建模集R2和RMSE為0.75、3.48，驗證集R2、RMSE和RPD為0.78、3.02和2.56，模型較為準確、可靠；基于該模型反縯的研究區土壤鹽分含量整體較高，鹽漬化程度空間分佈錶現為自西南部農業生產區至東北沿海區域逐漸加重，與實地調查一緻。研究錶明基于Landsat 8 OLI多光譜影像，引入第7波段對植被指數進行改進，從而構建土壤鹽分含量的支持嚮量機模型，可穫得較好的土壤鹽分空間分佈反縯結果。
쾌속획취토양염분적함량、특정급공간분포신식시염지토치리、이용적객관수구。해문침대황하삼각주염지토，이은리현위례，기우Landsat 8 OLI다광보영상，재전통식피지수적기출상인입단파홍외파단진행확전，제출료개진식피지수；연후기우개진전후대응적식피지수，분별채용다원축보회귀（multivariable linear regression，MLR）、반향전파신경망락（back propagation neural networks，BPNN）화지지향량궤（support vector machine，SVM）방법구건토양염분함량적요감반연모형，병진행모형험증、대비화우선；최후기우최가모형진행연구구토양염분함량적공간분포반연화분석。결과현시：상대전통식피지수，확전후식피지수가증강여토양염분적상관성，대폭강저지수간적다중공선성；채용상술3충방법건모，개진후모형적정도비개진전도유제고，험증집결정계수R2제고0.04～0.10，균방근오차RMSE강저0.13～0.73，상대분석오차RPD제고0.25～0.34，개진후모형RPD균대우2.0，보편체도성능량호；대비3충건모방법，SVM건모정도최고，BPNN모형차지，MLR분석정도최저，최가모형위기우개진식피지수적토양염분함량지지향량궤반연모형，건모집R2화RMSE위0.75、3.48，험증집R2、RMSE화RPD위0.78、3.02화2.56，모형교위준학、가고；기우해모형반연적연구구토양염분함량정체교고，염지화정도공간분포표현위자서남부농업생산구지동북연해구역축점가중，여실지조사일치。연구표명기우Landsat 8 OLI다광보영상，인입제7파단대식피지수진행개진，종이구건토양염분함량적지지향량궤모형，가획득교호적토양염분공간분포반연결과。
Fast acquisition of the soil salt content, characteristics, and spatial distributing are the objective needs of saline soil management and utilization. This paper focused on the saline soil on the Yellow River Delta, and took Kenli County as an example. Based on the multi-spectral remote sensing image of Landsat 8 OLI, the traditional vegetation index (VI) was extended by adding the short-wave infrared band, and the modified vegetation index (MVI) was put forward. Then, based on the corresponding VI and MVI, using multivariate stepwise regression (MLR), a back propagation neural network (BPNN), and the support vector machine (SVM) method respectively, the remote sensing inversion models of soil salinity were built, validated, and compared. Finally, the spatial distribution of soil salinity was analyzed using the best model in the study area. The results indicated that the correlation between the vegetation indices and soil salinity was heightened and the multicollinearity between vegetation indices was greatly reduced by extending the traditional vegetation index. Extended normalized difference vegetation index(ENDVI) and extended ratio vegetation index (ERVI) which were added band 7 were selected as the modified vegetation index(MVI). Using MLR, a BPNN and the SVM method, the precision of the models based on the MVI was improved compared to the VI with the calibration coefficient of determination (R2) raised between 0.05 and 0.11, and the calibration root mean squares error (RMSE) reduced between 0.09 and 0.55, the validationR2 raised between 0.04 and 0.10, the validation RMSE reduced between 0.13 and 0.73, and the validation relative prediction deviation (RPD) raised between 0.25 and 0.34. The models based on MVI obtained generally good performance with the validation RPD greater than 2.00. The main reasons improved the model precision were that the band 7 on Landsat 8 OLI had more information and the MVI including band 7 could more protrude the difference in vegetation coverage and production status. Comparing the three modeling methods, the SVM achieved the highest accuracy, the second was the BPNN, and the MLR analysis resulted in the lowest accuracy. With the calibrationR2 and RMSE of 0.75 and 3.48, the validationR2, RMSE and RPD of 0.78, 3.02 and 2.56, the SVM model of soil salinity based on MVI was the best and obtained very high accuracy and reliability for remote sensing inversion of soil salinity content. The spatial distribution of soil salinity content in the study area was analyzed based on the best model. The statistical information of the inversed soil salinity was very close to the measured value of soil samples, the soil salinity content in the study area was very high generally, the area that belonged to severe saline soil and solonchak accounted for 77.91%, and the spatial distribution of soil salinization showed that the soil salinity content was gradually increased from the southwest agriculture region to the northeast coastal region, which was consistent with the field survey and geostatistical analysis. Therefore, the experiment indicated that the vegetation index was modified by introducing the band 7 based on Landsat 8 OLI, and the SVM model of soil salinity was built, which could obtain better inversion result of soil salinity spatial distribution.
