Research On Precision Identification And Application Of Lettuce Phenotype Based On Deep Learning

In recent years,smart agriculture,through scientific monitoring,precise management,and efficient use of agricultural resources,has become a central development trend in promoting a transition to sustainable and efficient agricultural production.Phenotyping,as an external representation of crop growth and genetic characteristics,is of great significance for the realization of smart agriculture due to its accurate identification and assessment.The precise identification of crop phenotypes can provide an efficient,accurate,and objective method of monitoring and management for agricultural production,which is of practical significance in ensuring crop quality and yield,precise breeding,and improving economic benefits.Additionally,a series of applications centered on phenotypic indicators have provided crucial technical support for optimizing crop cultivation management and have also advanced the efficient and high-quality development of smart agriculture.This paper focuses on lettuce as the research subject,utilizing common water and nitrogen stress conditions during its growth as experimental parameters.We collected RGB images,hyperspectral data,and physiological information throughout the complete growth cycle of lettuce.Multiple deep learning-based computational models have been developed to address key issues in the growth process of lettuce,and a comprehensive deep learning dataset has been constructed.The cost-effective and rapid precise identification and assessment of lettuce phenotypic indicators have been achieved,which allows for a more intuitive and comprehensive grasp of the growth status and trends of lettuce.Based on this,precise predictions for future growth images and phenotypic indicators of lettuce have been made,as well as the accurate graded diagnosis of water and nitrogen stress conditions in lettuce and early detection of pest damage.At the same time,non-invasive precise detection of key physiological information in lettuce has been completed through multimodal fusion of phenotypic indicators and spectral data.Ultimately,a low-cost,intelligent monitoring and management system for the entire lifecycle of lettuce has been realized,providing technical support for the high-quality and efficient production of lettuce.The main research content and results are as follows:（1）Identification and Evaluation of Lettuce Phenotypic IndicatorsThe AUNet,an image segmentation network model based on deep learning,was constructed,integrating a spatial attention mechanism to achieve precise segmentation of lettuce images.Utilizing segmented leaf and planting pot images,45 phenotypic indicators of lettuce were rapidly and accurately identified,including representative phenotypic indicators of lettuce growth state defined in this study.Various statistical methods were employed to filter and analyze phenotypic indicators as well as dynamically track the changes in phenotypic traits during the growth process.Quantitative analysis of the accumulation and evolution of phenotypic indicators at different growth stages enabled an intuitive and comprehensive analysis of lettuce growth state and trends,revealing the adaptability of lettuce to different water and nitrogen stress conditions and its response to environmental changes.（2）Prediction and Analysis of Lettuce Growth and DevelopmentUsing the lettuce planting pot as a reference,the size of the lettuce images was adjusted accurately to construct a dataset suitable for a spatiotemporal prediction network model.This data set was used to train a deep learning-based model for predicting and analyzing lettuce growth and development.The model merged the image spatial feature extraction capability of CNNs with the sequential data processing ability of LSTMs,utilizing an encoder-decoder architecture to precisely predict future growth trend images of lettuce.Moreover,by combining MSE,SSIM,and perceptual loss computation methods,we optimized and constructed a Multi-loss Conv LSTM Encoder-Decoder model,further enhancing the quality of predicted images and the accuracy of phenotypic indicators.Notably,the average MAPE of predicted lettuce image geometric indicators was less than 0.55%,while that for color and texture indicators was less than 1.7%.（3）Identification,Grading Diagnosis of Lettuce Water and Nitrogen Stress States,and Early Pest DetectionVarious types of classification recognition and object detection models were established using lettuce phenotypic indicators and RGB image data,achieving precise identification and grading diagnosis of lettuce water-nitrogen stress states and early detection of pest infestation.Specifically,a DNN network model trained with the Feature Phenotype Dataset offered efficient,rapid,and accurate identification and grading diagnosis of lettuce water-nitrogen stress states,providing a feasibility basis for stress state monitoring and diagnosis using non-image data.The Res Net50Evo-SE model established using image data combined with the Squeeze-and-Excitation channel attention mechanism reached an accuracy rate of0.9897 for the recognition and grading of water-nitrogen stress states,representing an increase of 1.25%and 0.49%compared to the original Res Net50Evo and Vi T models,respectively,while also consuming fewer computational resources.The Yolo-CBAM model,established using image data in conjunction with the Convolutional Block Attention Module,accurately detected and localized color and morphological features of lettuce,while also effectively recognizing states of leaf miner pest infestation,facilitating early pest detection.（4）Non-Invasive Detection of Lettuce Physiological InformationUsing a model-level data fusion approach,a Plant-MTL network model based on multimodal learning was constructed.This model effectively integrated phenotypic indicators and spectral data of lettuce,enabling accurate prediction of chlorophyll and fresh weight content.The multimodal learning method combining data showed a certain improvement in the accuracy of predicting lettuce physiological information compared to using spectral data or phenotypic indicators alone,demonstrating a synergistic effect between the two types of data.The Plant-MTL model,established based on model-level data fusion,outperformed models built using traditional machine learning methods such as LS-SVM,LASSO,and PLS,as well as the Plant-MLP model built using feature-level data fusion.The Plant-MTL network model is the optimal model for predicting lettuce physiological information,with R_train² and R_test² for chlorophyll content prediction reaching 0.9559and 0.9502,respectively,and R_train² and R_test² for fresh weight prediction reaching0.9820 and 0.9838,respectively,achieving effective accuracy gains over other models.