基于无人机采集的视觉与光谱图像预测棉花产量

为了高效管理农田,该文提出了一种应用低空遥感视觉与光谱图像预测棉花产量的方法。盛花期前的棉花图像由无人机遥感平台在距地面50m的飞行高度下采集,采集的局部图像通过拼接处理得到棉花地的全景RGB图像与CIR(color-infrared,彩色红外)图像。基于全景图像提取并计算了色度、植株覆盖率与归一化植被指数(normalized difference vegetationindex,NDVI)3个特征参数,用于构建棉花产量的预测模型。包括产量与特征参数的原始数据集随机分为训练集(90%)与测试集(10%)。训练集数据首先基于产量概率分布特征去除了10%的离群值,然后通过均值滤波器滤波,处理后的数据用于构建预测模型。通过SAS软件对比分析了单变量、双变量以及三变量构建的线性回归模型,预测模型由P值、决定系数R2、每0.4 hm2面积下估计值与真实值之间的平均绝对误差百分比(mean absolute percentage error,MAPE)这3个参数进行评估。试验结果表明,单变量、双变量以及三变量构建的共7个线性回归模型,其P值均小于0.05,则7个线性回归模型均具有统计学意义(5%显著性水平)。其中,由三变量构建的多元线性回归模型具有最大的决定系数R2=0.9 773,因此适应性最优。基于测试集验证模型精度,试验结果表明,采用多元线性回归模型进行产量估计,估计值与实际值之间的平均绝对误差百分比为4.0%。因此,无人机搭载图像传感器采集提取视觉与光谱特征能够有效用于作物产量的预测。

Cotton yield prediction using remote visual and spectral images captured by UAV system

Abstract: Crop yield prediction is helpful for farmers to make reasonable agronomic decisions such as crop insurance, planning of harvest, storage requirement, cash flow budgeting and the determination of input factors. This paper presented an application for cotton yield prediction based on the remote visual and spectral images captured by a UAV (unmanned aerial vehicle) remote sensing system. Field experiments were carried out at a cotton field approximately 5 hm2. Images of the cotton field were captured by the UAV remote sensing system at a 50 m flight height before fully blooming. One RGB panoramic image and one CIR (color-infrared) panoramic image were obtained by stitching the collected partial images with 75% overlap. Yield data with GPS (global position system) information were registered respectively to the RGB and CIR images according to the GCPs (ground control points). Then 3 features including color feature, coverage rate and vegetation index were extracted from the images and used to develop predication models. Color feature was the hue channel in HSV (hue, saturation, value) color model. Coverage rate was calculated by using Excess Green value based on RGB (red, green, blue) color model. Vegetation index was NDVI (normalized difference vegetation index) value usually used to monitor the crop growth. The datasets including yield data and 3 features were randomly divided into training data (90% of raw data) for modeling and test data (10% of raw data) for evaluating the model accuracy. The training data were preprocessed though removing 10% partial outliers based on probability distribution of yield data and being smoothed by mean filter respectively. Then multiple linear regression models were built to predict cotton yield, and parameters of the models were estimated based on the least square method and calculated by SAS software. 3 parameters were used to evaluate the models including P value in F-test statistic, coefficient of determination R2, and MAPE (mean absolute percentage error) between predicted values and the ground truth data in acre area level. The results showed that 7 linear regression models built by 1 variable (color feature, coverage rate or vegetation index), 2 variables (color feature and coverage rate, color feature and vegetation index or coverage rate and vegetation index) and 3 variables (color feature, coverage rate and vegetation index) were all statistically significant at significance level 5%. However, the adequacy of the multiple linear regression model with 3 variables was the best due to the highest R2=0.977 3 compared to other 6 regression models built by 1 variable and 2 variables. So the multiple linear regression model with 3 variables was used to predict yield. And the average MAPEs on the multiple linear regression models built by 3 variables was 4.0% in acre area level. The results from this study indicate that combining visual and spectral features from image sensors can be used effectively to predict yield.