基于角果期高光谱的冬油菜产量预测模型研究

以连续3 a田间氮肥水平试验为基础,研究基于高光谱估产的可行性,明确最佳光谱监测方式和有效波段,降低光谱分析维数,提高产量估测时效性。2013—2016年分别于湖北省武穴市和沙洋县进行大田试验,通过测试角果期冠层光谱反射率、产量构成因子(单株角果数、每角粒数和千粒质量)和成熟期产量,利用偏最小二乘回归(PLS)分别对油菜原初光谱(RSR)和一阶微分光谱(FDR)与其产量及构成因子间构建定量分析模型并筛选有效波段。结果表明,基于全波段的FDR-PLS模型预测精度显著优于R-PLS,其最佳监测指标是冬油菜产量和角果数,验证集决定系数(R2)分别为0.90和0.91,均方根误差(RMSE)分别为379 kg/hm2和66个/株,相对分析误差(RPD)分别为3.11和3.12。基于各波段变量重要性投影(VIP)值,确定冬油菜产量有效波段分别为628、753、882、935、1061、1 224 nm;角果数有效波段分别为628、758、935、1 063、1 457、1 600 nm。此后,再次构建基于上述有效波段的冬油菜产量和角果数监测模型,决定系数分别为0.91和0.87,均方根误差分别为504 kg/hm2和82个/株,相对分析误差分别为2.34和2.52,估算精度较为理想。

Prediction Models of Winter Oilseed Rape Yield Based on Hyperspectral Data at Pod-filling Stage

Hyperspectral remote sensing can provide a non-destructive and effective approach for assessing the yield and yield components of oilseed rape timely. A quantitative technique was developed to estimate oilseed rape yield accurately depending on ground-based canopy reflectance spectra. Field experiments were conducted over three growing seasons at different sites (Wuxue and Shayang) in Hubei Province, China. The key parameters, including canopy hyperspectral reflectance during pod-filling period, seed yield and yield components (pod numbers per plant, seed numbers per pod and 1000 seed weight) were monitored. A partial least square (PLS) regression analysis was employed to perform the relationship between raw spectral reflectance (RSR), the first derivative reflectance (FDR) and seed yield and yield components. According to the calibration dataset, the best results were obtained with the FDR-PLS model for the prediction of yield and pod number, which yielded the highest coefficient of determination (R2cal) of 0.96 and 0.98, and the lowest root mean square error (RMSEcal) of 158kg/hm2 and 17 pods/plant, respectively. The tests using the independent validation dataset also showed that the FDR-PLS model could well forecast yield and pod number of winter oilseed rape, with values of R2val of 0.90 and 0.91, RMSEval of 379kg/hm2 and 66 pods/plant, and RPD of 3.11 and 3.12, respectively. The variable importance in projection (VIP) scores resulted from the PLS regression analysis were used to determine the effective wavelengths and reduce the dimensionality of the spectral reflectance data. The newly-developed FDR-PLS model using the effective wavelengths (628nm, 753nm, 882nm, 935nm, 1061nm and 1224nm) performed well in yield prediction with R2val of 0.91, RMSE val of 504kg/hm2 and RPDval of 2.34;Similar results were also obtained for pod number prediction with R2val of 0.87, RMSEval of 82 pods/plant and RPDval of 2.52 using the effective wavelengths (628nm, 758nm, 935nm, 1063nm, 1457nm and 1600nm). Consequently, the yield of winter oilseed rape could be reliably estimated with the in situ developed FDR-PLS method.