氮磷互作对水稻冠层光谱的影响及其PNN识别

【目的】氮、磷均为作物必需的大量营养元素，其丰缺诊断直接关系到合理科学施肥，进而影响产量、效益以及环境。本文旨在研究准确、快捷、无损地区分水稻缺氮和缺磷信息的光谱识别方法，从而指导田间施肥决策，精确作物管理、节约种植成本并控制农田面源污染。【方法】基于水稻6个氮素及两个磷素营养水平交互下的盆栽试验，分别在分蘖、拔节和抽穗期测定水稻冠层的可见近红外反射光谱（350-1 330 nm）及植株全氮（TN）和全磷（TP）含量等数据，分析氮磷互作对水稻植株体内TN和TP含量以及冠层反射光谱的影响，并运用概率神经网络(PNN)分别对不同生育时期的冠层光谱进行氮水平、磷水平、氮磷交互水平和缺素水平4个尺度下的分类识别。为避免光谱测量时仪器误差和光照、风力、温度、水分等环境条件所造成光谱数据批次间的差异，PNN分类识别前对光谱数据进行标准化处理，并将其中2/3作为训练集，另外1/3作为测试集。【结果】植株全氮含量受氮肥、磷肥和氮磷交互作用的影响显著；植株全磷含量则主要受磷肥和氮肥水平的双重影响，但不存在氮磷交互作用。水稻冠层光谱对氮肥的响应规律不受磷肥水平的影响，缺氮使可见光区反射率升高，近红外区反射率下降。缺磷使近红外区反射率下降，但可见光区的响应则受氮肥水平的影响，施氮处理呈上升趋势，氮胁迫处理则呈现分蘖期下降、拔节期上升、抽穗期下降的趋势。利用冠层光谱PNN模型可以对各个生育时期氮水平、磷水平、氮磷交互水平和缺素水平等不同施肥尺度进行识别，拔节期分类精度最高，抽穗期分类精度相对最低。4种分类尺度下PNN模型对磷素水平的分类精度最高，分蘖期和拔节期分别为83%和94%；其次是缺素水平，分别为78%和88%；对氮素水平以及氮磷交互水平等有较多个分类输出的识别精度较低，为61%-75%。值得一提的是，PNN模型对水稻施肥关键生育时期分蘖期和拔节期水稻植株缺氮缺磷、缺氮不缺磷、缺磷不缺氮、不缺氮不缺磷等4种缺素水平的分类中，所有只缺氮处理没有被预测为只缺磷处理，所有只缺磷处理也没有被误判为只缺氮处理，表明冠层光谱PNN模型能有效区分开氮磷胁迫。【结论】水稻的冠层光谱受到氮、磷水平的共同影响，利用水稻冠层光谱建立的PNN模型不仅能分别辨识各氮素、磷素施肥水平，并且能有效地区分开水稻缺磷和缺氮处理，避免混淆，对有目的性的指导施肥具有重要的意义和价值，可避免不恰当的施肥策略造成的环境、产量和经济损失。

Effects of Interaction of N and P on Rice Canopy Spectral Reflectance and Its PNN Identification

【Objective】Nitrogen (N) and phosphorus (P) are macronutrients for crops and the diagnosis of N and P in crops is the premise of scientific fertilization. Accurate, fast, and nondestructive detection of the deficiency of N and P in rice has a great meaning on precision fertilization, cost-saving and agricultural non-point source pollution control.【Method】A two-factor pot experiment of N (6 levels) and P (2 levels) was carried out, canopy spectral reflectance and plant TN and TP content were measured simultaneously at tillering, jointing and heading stages. The interactive effects of N and P on rice growth (N and P content) and canopy reflectance at 350-1 330 nm was investigated and PNN model was used to classify the N and P levels based on the canopy reflectance. In order to avoid the error of different batches caused by instrument, light, wind, temperature, water and other environmental conditions, the reflectance spectra data were standardized. A total of 2/3 of the data were used to train the PNN model and the other 1/3 data were used to test the PNN model. 【Result】 Rice N content was significantly influenced by the N rate, P rate and the interaction of N and P. But rice P content was only affected by P rate and N rate, the interaction of N and P did not exist. The response of canopy reflectance spectra to N rate was not influenced by P rates, and N deficiency increased the reflectance at visible band and decreased those in the near infrared region. Under P-deficiency, the reflectance at near-infrared bands decreased at all N levels, but the reflectance at visible bands increased in N application treatments while declined at tillering stage, increased at jointing stage, then decreased at heading stage when the N was seriously deficient. The identification accuracy of PNN model was highest at jointing stage and lowest at heading stage. The identification accuracy at tillering and jointing stages was 83% and 94% for P levels, and 78% and 88% for N and P deficiency levels, respectively. However, the identification accuracy for N levels and interaction of N-P levels was only 61%-75%. It was noted that all the N-deficient treatments were not identified as P-deficiency and vice versa at tillering and jointing stages, which showed that PNN model could distinguish N-deficiency from P-deficiency.【Conclusion】Rice canopy reflectance was influenced by N and P fertilizer levels. The PNN model of the rice canopy reflectance can not only identify nitrogen and phosphorus fertilizer levels based on single factor, but also can distinguish N-deficiency from P-deficiency, and has important meaning and value in rice fertilization decision and may avoid yield and economic loss and environmental pollution caused by improper fertilization strategy.