基于图谱信息融合决策支持的小麦地上鲜生物量快速检测方法研究

The wheat is one of the food which is needed mostly in our country and world, so the grain production and the quality safety of agricultural products is needed to increase urgently. The above-ground wheat biomass in the whole stage was an important index to evaluate the yield of wheat and how it grew. The single detection method was applied widely in detection of wheat biomass for the advantage of quickly and non-destructively, such as spectroscopy and machine vision technique. But there were some downsides to relying solely on single technology like low accuracy of detection model against precision agriculture decision. In this project,the wheat in the whole stage was considered as measuring object, canopy image in the early growth, side image in the middle or later period and canopy spectral were acquired. The change patterns of spectral and image were studied in the whole period. By using image processing and spectroscopy analysis method, a serial characteristic image parameters and spectral parameters were extracted, which could reflect the wheat biomass and its growing. The image and spectra information were used to construct a biomass prediction model by data fusion method of partial least squares method. The validation was measured to test the accuracy of detection model. Finally, a fast and non-destructive detection and evaluation method for the wheat above-ground fresh biomass was built, which could provide method for the precision agriculture decision.

小麦是我国乃至世界上需求最多的粮食作物之一，迫切需要提高粮食产量和农产品质量。而整个生育期内小麦的地上鲜生物量是评价小麦产量和长势的重要指标。现有的光谱技术和机器视觉技术等单一的检测方法以其快速无损的优点在作物生物量检测方面得到了广泛的应用。但是，这些方法存在一些缺陷如建立模型的准确率较低，不利于精准农业决策。本项目以整个生育期内的小麦为测量对象，获取生育早期的冠层图像、中后期的侧面图像和整个生育期内的冠层光谱，分析小麦生物量在整个生育期内的光谱和图像变化规律，综合应用图像处理和光谱分析，提取出反应小麦地上鲜生物量和长势的图像特征指标和光谱特征指标，应用偏最小二乘法等方法建立融合光谱、图像信息的小麦地上鲜生物量预测模型。通过对验证集小麦地上鲜生物量的预测来验证模型的准确性。最终构建一套小麦地上鲜生物量快速无损检测方法，为精准农业决策提供方法。

地上鲜生物量（AGB）是反映植物生长状况的重要指标之一，可用于作物产量预测。目前在田间尺度小区域内小麦生物量预测仍然存在中后期预测精度较低的问题。本项目研究了小麦图像特征和光谱特征以及图谱融合特征来预测田间小区域内冬小麦地上鲜生物量在不同阶段内。从冬小麦冠层光谱信息中提取特征波长和6种植被指数作为光谱特征参数，并通过图像处理技术提取八种基于灰度共生矩阵的纹理特征和计算NDTI作为图像特征参数，建立线性回归模型比较不同生育期（苗期，后期和全期）的单一图谱特征参数与生物量之间的关系，分析图谱特征中与AGB相关性较强的特征参量，分别作为最终的图谱特征参量（植被指数、特征波长、纹理特征最优子集、NDTI最优子集）,然后使用基于图谱融合特征的随机森林回归（RF）、偏最小二乘回归(PLSR) 与使用单一图谱特征参量的简单线性回归模型（LR）对冬小麦地上鲜生物量（AGB）的预测能力进行对比。这些融合图谱特征参量分别为植被指数与纹理特征最优子集、特征波长与纹理特征最优子集、特征波长与NDTI最优子集、NDTI最优子集和纹理特征最优子集、植被指数与特征波长、植被指数与NDTI最优子集共6种组合。结果表明，在不同生长阶段中，相对于其余5种图谱融合特征，其中基于TEX+SPA的融合图谱特征参量的AGB估测模型精度最好R2=0.943,RMSE=0.014、R2=0.788,RMSE=0.059、R2=0.87,RMSE=0.054,相对于单一图谱估测AGB精度分别提升了5.9%、4.4%、4.9%。由此可见，相比单一的光谱信息和图像信息模型，本研究提出的融合图谱特征充分的利用图像与光谱信息，能准确预测作物AGB。