运用PLS算法由HJ-1A/1B遥感影像估测区域小麦实际单产

为进一步提高遥感估产精度,显示国产影像在农业估产中的应用效果。该研究以2010-2013年HJ-1A/1B影像为遥感数据,分析了卫星遥感变量与小麦实际单产的定量关系,运用偏最小二乘回归算法构建及验证了以实际单产为目标的多变量遥感估产模型,并制作了小麦实际单产空间等级分布图。研究表明:实际单产与所选用的大多数遥感变量间关系密切,且多数遥感变量两两间具有严重的多重相关关系;实际单产偏最小二乘回归模型的最佳主成分为5,且植被衰减指数、绿色归一化植被指数、调整土壤亮度的植被指数、比值植被指数和归一化植被指数为实际单产遥感估测的敏感变量;建模集和验证集实际单产估测模型的决定系数分别为0.74和0.70,均方根误差分别为754.05和748.20 kg/hm2,相对误差分别为11.5%和8.88%,且估测精度比线性回归算法分别提高20%以上和40%以上,比主成分分析算法分别提高18%以上和30%以上,说明偏最小二乘回归算法模型估测区域实际单产的效果要明显好于线性回归和主成分分析算法,该模型应用结果与小麦实际单产区域分布情况相符合,为提高区域小麦实际单产的遥感估测精度提供了一种途径。

Remote sensing estimation of wheat practical yield on regional scale using partial least squares regression algorithm based on HJ-1A/1B images

Abstract: Estimation of crop yield by remote sensing is a key research and application field in agriculture, and such research can provide timely and reliable yield information for regional food production. In order to further improve the accuracy of estimating wheat yield by remote sensing, and demonstrate the application of satellite imaging products in agricultural production, we used HJ-1A/1B images on April 26th 2010, April 28th 2011 and 2012, May 2nd 2013 at wheat anthesis stage as remote sensing data. 335 samples of wheat yield were collected from agriculture production field and divided into modeling dataset and validation dataset on a ratio of 3:2. Based on the minimum value of predictive residual error sum of square (PRESS), the number required for principal component model was determined. The yield estimation model was assessed through determination coefficient (R2), root mean square error (RMSE) and relative error (RE). This research was undertaken to make a systematic analysis on the quantitative relationship of satellite remote sensing variables to actual wheat yield. Depending on the partial least squares regression (PLS), the multivariable remote sensing estimation models and the space level distribution maps of actual wheat yield were constructed and verified by the modeling dataset and validation dataset, and the estimation effect of the PLS model was compared to linear regression (LR) and principal components analysis (PCA) algorithm models, respectively. The results of this research indicated that the majority of remote sensing variables were significantly (P< 0.05) related to practical yield, and there were significant (P < 0.05) multiple relationships among the majority of remote sensing variables. For the actual yield estimation model based on PLS, the number of the best principal components was 5. Plant senescence reflectance index (PSRI), green normalized difference vegetation index (GNDVI), optimal soil adjusted vegetation index (OSAVI), ratio vegetation index (RVI) and normalized difference vegetation index (NDVI) were identified as the sensitive remote sensing variables for estimating wheat yield. Through testing the actual yield estimation model based on PLS algorithm with modeling dataset and validation dataset, the R2 were 0.74 and 0.71, respectively, and the RMSE were 754.05 kg/hm2 and 748.2 kg/hm2, respectively, the RE were 11.50% and 8.88%, respectively. The PLS model with selected sensitive variables performed better to estimate wheat yield. PLS algorithm models to estimate wheat yield obtained the higher accuracy by above 20% and above 40% than the LR algorithm models, by above 18% and above 30% than the PCA algorithm models for modeling dataset and validation dataset, respectively. Based on the above PLS model and HJ-1A/1B image on May 2nd, 2013, the wheat practical yield spatial distribution level was mapped in central Jiangsu region. The results of applying the PLS models were correspondent with the actual distribution of wheat yield. It was concluded that PLS algorithm can provide an effective way to improve the accuracy of estimating wheat yield on regional scale based on aerospace remote sensing, and can contribute to large-scale application of the research results.