基于多叶位快速叶绿素荧光和1D-DRDC-Net的棉苗盐胁迫诊断方法

Cotton Seedling Salt Stress Diagnosis Method Based on Multi-leaf Rapid Chlorophyll Fluorescence and 1D-DRDC-Net

摘要:

盐胁迫会导致棉花纤维品质及产量下降，尤其在苗期时其遭受盐胁迫影响最大。为了实现棉苗盐胁迫的快速诊断，本文利用快速叶绿素荧光技术获取了不同盐胁迫程度下棉苗冠层叶片的OJIP曲线，并结合深度残差网络(Deep residual network, ResNet)和空洞卷积（Dilated convolution）结构构建了基于“叶位-通道”荧光数据融合的1D-DRDC-Net（1D-deep residual dilated convolutional neural network）棉苗盐胁迫深度学习诊断模型。结果表明，盐胁迫导致棉苗体内含水率下降，丙二醛(Malondialdehyde, MDA)含量、超氧化物歧化酶(Superoxide dismutase, SOD)活性、过氧化物酶(Peroxidase, POD)活性升高；在垂直方向上盐胁迫对棉苗的影响趋势表现为植株上部分叶片各参数变化明显，其中对胁迫最敏感的叶位为L1，而成熟叶片受到的影响相对较小。相比于其它模型，1D-DRDC-Net对棉苗不同胁迫时间下3个盐浓度梯度（0、100、200 mmol/L）的诊断精度为76.67%， F1值为76.48%，比支持向量机（Support vector machine, SVM）、反向传播神经网络（Back propagation neural network, BPNN）准确率均提高5个百分点，比随机森林（Random forest, RF）提高14.45个百分点，比双向长短期记忆网络（Bidirectional long short-term memory，Bi-LSTM）提高3.34个百分点。基于“叶位-通道”的荧光信息融合策略准确率优于仅使用单一敏感叶位荧光信息8.89个百分点，其鲁棒性和泛化能力均优于只采用普通卷积核和取消“跳跃连接”的模型。最终，建立的1D-DRDC-Net模型在棉苗受到胁迫7、14、21 d后，对植株是否受到盐胁迫的诊断准确率分别达到83.33%、88.33%和95.00%，研究结果可为棉花栽培管理提供理论依据。

Abstract:

Salt stress can lead to a decrease in cotton fiber quality and yield, especially during the seedling stage when it is most affected by salt stress. In order to achieve rapid diagnosis of salt stress in cotton seedlings, rapid chlorophyll fluorescence technology was used to obtain OJIP curves of cotton seedling canopy leaves under different degrees of salt stress, and deep residual network (ResNet) and dilated convolution structures were combined to construct a 1D-deep residual dilated convolutional neural network (1D-DRDC-Net) cotton seedling salt stress deep learning diagnosis model-based on “leaf-position channel” fluorescence data fusion. The results showed that salt stress significantly led to a decrease in water content in cotton seedlings, an increase in malondialdehyde (MDA) content, superoxide dismutase (SOD) activity, and peroxidase (POD) activity. The trend of salt stress on cotton seedlings in the vertical direction showed significant changes in various parameters of the upper leaves of the plant, with L1 being the most sensitive leaf position to stress, while mature leaves were relatively less affected. Compared with other models, the diagnostic accuracy of 1D-DRDC-Net for three salt concentration gradients (0 mmol/L, 100 mmol/L and 200 mmol/L) under different stress times in cotton seedlings was 76.67%, with an F1-score of 76.48%, which was 5 percentage points higher than the accuracy of support vector machine (SVM) and back propagation neural network (BPNN), 14.45 percentage points higher than that of random forest (RF), and 3.34 percentage points higher than that of bidirectional long short-term memory (Bi-LSTM). The fluorescence information fusion strategy-based on “leaf-position channel” was more effective than using only a single sensitive leaf position fluorescence information by 8.89 percentage points. Its robustness and generalization ability were stronger than that of models that only use ordinary convolution kernels and cancel “skip connections”. Finally, the established 1D-DRDC-Net model achieved diagnostic accuracies of 83.33%, 88.33%, and 95.00% on the 7th, 14th, and 21st day after cotton seedlings were subjected to salt stress, respectively. The research results can provide theoretical basis for cotton cultivation management.