基于高光谱和卷积神经网络的大田马铃薯早疫病严重程度分级方法

Severity classification of potato early blight in field based on hyperspectral and convolutional neural network

早疫病是影响马铃薯产量的主要病害之一,大田病害监测对控制早疫病发展有重要意义。使用配备高光谱成像仪的无人机(UAV)在田间尺度上获取患不同严重程度早疫病的马铃薯高光谱影像,分别提取并计算健康、轻度感染、中度感染和重度感染马铃薯的冠层光谱数据,通过光谱变换得到包括原始光谱在内的4种光谱,再进行特征波段选取,利用卷积神经网络(CNN)基于全波段和特征波段对马铃薯早疫病不同发病程度进行识别。结果表明,一阶微分光谱随机蛙跳(RF)降维后的特征波段+CNN模型的效果最好,总体识别准确率为91.18%,比一阶微分光谱随机蛙跳(RF)降维后的特征波段+反向传播网络(BP)总体准确率提高了1.96个百分点,平均精准率和平均召回率分别提高了3.00个百分点和2.00个百分点,平均F1得分提高了0.02;对不同感染等级的识别精度分别达到了95.0%、88.0%、83.0%和97.0%。

Early blight is one of the major diseases affecting potato yield,and field disease detection is important for controlling disease development.The unmanned aerial vehicle(UAV)equipped with a hyperspectral imager was used to acquire hyperspectral images of potatoes with different severity of early blight on the field scale,and the canopy spectral data of healthy,mildly infected,moderately infected,and severely infected potatoes were extracted and calculated,respectively.Four kinds of spectra including the original spectra were obtained by spectral transformation,and then the feature bands were selected.Convolutional neural network(CNN)was used to perform the identification of different degrees of potato early blight based on the full band and feature bands.The results showed that the feature bands after first-order differential spectra random leapfrog(RF)dimension reduction+CNN model had the best effect,and the overall recognition accuracy rate was 91.18%.Compared with the feature bands after first-order differential spectra RF dimension reduction+back propagation network(BP),the overall accuracy rate was increased by 1.96 percentage points,the average precision rate and the average recall rate were increased by 3.00 percentage points and 2.00 percentage points respectively,and the average F1 score was increased by 0.02.The identification accuracy of potato early blight with different infection levels reached 95.0%,88.0%,83.0%and 97.0%,respectively.