基于高光谱的番茄氮磷钾营养水平快速诊断

Rapid diagnosis of tomato N-P-K nutrition level based on hyperspectral technology

为了精确、快速和稳定的对番茄氮、磷、钾3种元素的营养水平进行诊断,该文提出利用反射光谱技术诊断方法,选用遗传算法优选波段;采用主成分分析方法提取敏感波长下的纹理特征;通过逐步回归、主成分回归、偏最小二乘法回归分别建立基于光谱和图像特征的番茄叶片氮、磷、钾素模型。针对单一技术不能全面反映叶片营养信息的问题,采用人工神经网络对光谱和图像技术进行特征层的信息融合,建立了多信息融合的诊断模型,求得氮、磷、钾的相关系数R分别为0.9651、0.9216、0.9353;均方根误差RMSE分别为0.19、0.33、0.29。结果表明采用光谱与图像的融合技术模型比单一光谱模型提高的精度分别为6.25%、3.97%、7.92%,比单一图像模型提高的精度为3.80%、5.43%、3.26%,有更好的诊断作用,能够实现对番茄作物氮、磷、钾素营养水平的高精度快速检测。

Because of the short growth cycle, large yield and high fertilizer requirements of facility crop, which were the characteristics causing high cost and high complement in cultivation medium, nitrogen (N), phosphorus (P) and Potassium (K) deficiencies frequent occurrence in the growth of facility crops. Accurate monitoring and diagnosis of nutrient content in facility crops during the growth process was very important-. In order to diagnose tomato nitrogen, phosphorus and potassium nutrition level more accurately, rapidly and stably, in the aspect of the spectral diagnostics technology, the changes of reflectance on characteristic wavelengths were taking into consideration to assess the nutritional status of crops.-. Sensitive bands were selected by using genetic algorithms. Then, the quantitative models of tomato nitrogen, phosphorus, potassium were established via stepwise regression, principal component regression and partial least squares method respectively based on reflectance spectra. The results verified that the stepwise regression models outperformed the principal component and partial smallest squares regression models of nitrogen and phosphorus, while principal component regression get the best models of phosphorus. The correlation coefficient R of the best models were nitrogen (0.9026) >phosphorus (0.8819) > potassium (0.8561) . The root mean square error (RMSE) were nitrogen (0.3191) <phosphorus (0.4978)<potassium (0.5128). Imaging technology can analyze the change of texture and other characteristics that were caused by plant nutrient deficiency. Texture features were extracted from images under - sensitive wavelength by using principal component analysis. The nutrients models of the tomato leaf nitrogen, phosphorus and potassium based on image features were established by stepwise regression, principal component regression and partial least squares method respectively. The results verified that the principal component regression models outperformed others models of nitrogen and phosphorus, while partial least squares method get the best models of phosphorus. The correlation coefficient R of the best models were nitrogen (0.9271) > potassium (0.8991) >phosphorus (0.8673). The root mean square error (RMSE) were nitrogen (0.3413)<phosphorus (0.3994)<potassium (0.5628). For overcoming the inadequacies of models build with single feature sauce, diagnosed models of multi-information fusion was established for tomato nutrients stress via artificial neural network modeling. Feature layer fusion was combining with the internal components and external morphological caused by crop nutrients stress. The correlation coefficient R of nitrogen, phosphorus and potassium were 0.9651, 0.9216 and 0.9353. The root mean square error (RMSE) were 0.19, 0.33 and 0.29. Results fully showed that the spectra reflection technology and image technology after feature layer integration models were better than spectra reflection or single image technology. Artificial neural network models of nitrogen, phosphorus and potassium improve the correlation coefficient R accurately were 6.25%, 3.97%, 7.92% than the single spectral model and 3.80%, 5.43%, 3.26% than the single image model. Furthermore, the detection root mean square error was reducing. The results showed that the multi-information fusion models achieve a substantial increase in model accuracy and have better diagnostic accuracy in achieving high accuracy compared with a single feature model, thus the rapid and high sensitivity detection of nutritional stress of the tomato leaves could be realized, which provides basis to methods about crop nutrients for the development of fast and accurate diagnostic instrument with important academic value and application prospect.