【摘要】 植株氮含量(PNC)是评价作物长势和氮营养状况的重要指标，因此，准确高效地获取PNC信息，对动态监测马铃薯长势及精准施控氮肥具有重要意义。首先于马铃薯现蕾期、块茎形成期、块茎增长期、淀粉积累期和成熟期获取无人机高光谱影像，并基于预处理的影像提取5个生育期冠层的原始光谱和一阶微分光谱；其次将提取的冠层光谱与马铃薯PNC进行相关性分析，筛选出PNC的敏感波长；然后分别利用灰度共生矩阵和1～3阶颜色矩，提取冠层原始光谱特征波长处高光谱图像的纹理和颜色2种图像特征，并将提取的特征与马铃薯PNC进行相关性分析，筛选出相关性较高的前5个图像特征；最后分别基于光谱特征、图像特征和图谱融合特征利用弹性网络回归(ENR)、贝叶斯线性回归(BLR)和极限学习机(ELM) 3种方法建立马铃薯PNC估算模型。结果表明：(1)马铃薯5个生育期的冠层光谱特征波长存在差异，但多数位于可见光区域。(2)冠层原始光谱特征波长图像的纹理和颜色特征与PNC的相关性较高，且现蕾期到淀粉积累期的相关性明显高于成熟期。(3)基于单一光谱特征和单一图像特征构建的马铃薯PNC估算模型在现蕾期到淀粉积累期效果较好，成熟期效果较差。(4)现蕾期到淀粉积累期，基于图谱融合特征的马铃薯PNC估算效果明显优于单一光谱特征和单一图像特征。(5)马铃薯各生育期，基于同种变量利用ENR构建的PNC估算模型效果较好，BLR次之，ELM较差。其中，以图谱融合特征为模型变量，利用ENR构建的PNC估算模型精度和稳定性最好，5个生育期的建模R~2分别为0.91、 0.75、 0.82、 0.77和0.69, RMSE分别为0.24%、 0.31%、 0.26%、 0.22%和0.29%, NRMSE分别为6.59%、 9.79%、 9.58%、 7.87%和11.03%。该研究可为马铃薯的氮营养监测提供一种快捷高效的技术手段。更多还原

【Abstract】 Plant nitrogen content(PNC) is an essential indicator of crop growth and nitrogen nutrition status. Therefore, accurate and efficient access to PNC information is vital for dynamically monitoring potato growth and proper N fertilizer application. In this study, the UAV hyperspectral images were obtained at the budding stage, tuber formation stage, tuber growth stage, starch accumulation stage, and maturity stage of the potato. After preprocessing, the original canopy spectrum and first-order differential spectrum of five growth stages were extracted; Secondly, the correlation analysis was carried out between the extracted canopy spectrum and potato PNC, and the sensitive wavelength of PNC was screened out; Then, the texture and color of two image features of the hyperspectral image at the wavelength of the original spectral features of the canopy were extracted using the gray co-generation matrix and the 1st to 3rd-order color moments, respectively, and the extracted features were correlated with the potato PNC to filter out the top five image features with higher correlation; Finally, based on spectral features, image features, and map fusion features, potato PNC estimation models were established by using elastic network regression(ENR), Bayesian linear regression(BLR), and limit learning machine(ELM). The results showed that:(1) there are differences in the characteristic wavelengths of canopy spectra in the five growth stages of potatoes. Still, most of them were located in the visible region.(2) The correlation between the texture and color characteristics of the original spectral characteristic wavelength image of the canopy and PNC was high. The correlation from the budding stage to the starch store stage was significantly higher than that in the mature stage.(3) The estimation models of potato PNC based on a single spectral feature and a single image feature have a good effect from the budding stage to the starch accumulation stage but a poor effect at the maturity stage.(4) From the budding stage to the starch accumulation stage, the estimation effect of potato PNC based on the map fusion feature was significantly better than the single spectral feature and the single image feature.(5) In each growth period of potato, the PNC estimation models constructed by ENR based on the same variable were better, BLR was the second, and ELM was poor. Among them, the accuracy and stability of the PNC estimation models constructed by ENR with fusion characteristics as model variables were the best. The modeling R~2 of five growth periods were 0.91, 0.75, 0.82, 0.77 and 0.69 respectively; RMSE were 0.24%, 0.31%, 0.26%, 0.22% and 0.29% respectively, and NRMSE were 6.59%, 9.79%, 9.58%, 7.87% and 11.03% respectively. This study can provide a fast and efficient technical tool for monitoring the nitrogen nutrition of potatoes.更多还原