Deep Learning Based Spatiotemporal Prediction Model Of Wheat Growth And Development Under Controlled Environment

Wheat was the world’s largest sown and most widely distributed food crop.To ensure food security and environmental sustainability,it was urgently required to develop higheryielding varieties to meet future demand.Selecting and breeding superior wheat varieties required extensive growing,screening,and measuring of germplasm material.However,conventional breeding of wheat usually requires 8–10 years to complete a breeding cycle.Wheat grows slowly under natural conditions,which was the key issue limiting the cycle of wheat breeding trials.Efficient screening of superior alleles in segregating generations determined the success or failure of breeding.High-throughput screening tools were urgently needed for the selection of superior germplasm and the gene localization of related traits.Wheat growth and development spatiotemporal prediction models were investigated in this paper to predict wheat growth and development.Predicting and visualizing wheat growth and development and analyzing its phenotype(leaf number,projected area,plant height,and plant-type)will help breeders and others to get an early understanding of wheat growth and development.Statistical analysis of phenotypic variation,the correlation between different traits,and localization of genetic loci regulating the target phenotypic variation will be further carried out.Predicting wheat growth and development early and measuring phenotypic traits holds the potential to speed up experimental plant cycles and accelerate plant breeding by reducing the time of wheat growth,imaging,and measurement.Predicting wheat growth and development will be promising to solve the issues of long cycles,low efficiency,and great uncertainty in the wheat breeding industry.The main research contents and conclusions of the paper are as follows.(1)Construction of wheat growth and development dataset.Four wheat varieties(Fielder,Bainong 58(AK58),Jimai 22(JM22)and Konong 199(KN199))were used as research materials.The growth and development image sequences of fixed wheat were collected at fixed points and at regular intervals.Images were also pre-processed to construct a wheat growth and development prediction dataset.The datasets included multi-view wheat growth and development dataset and continuous-view wheat growth and development dataset.The datasets will provide a data basis for the subsequent research on wheat growth and development and growth prediction models.(2)Spatiotemporal prediction model of growth and development for wheat based on STLSTM.For the problem of the complex morphological structure of wheat plants and severe overlapping shading between leaves,a method of modeling spatio-temporal characteristics of wheat growth and development based on Densenet201 and bi-directional long and short term memory networks was designed.This investigates the feasibility of modeling the spatio-temporal characteristics of wheat growth and development.Further,the modeling of spatial and temporal characteristics of wheat growth and development is carried out using tightly coupled spatio-temporal correlations of wheat growth and development.A spatiotemporal long and short term memory network(ST-LSTM)based modeling method for spatio-temporal characteristics of wheat growth and development is proposed.Then the encoder and decoder structures were fused to propose a wheat growth and development prediction model to achieve wheat growth and development image sequence prediction.The proposed prediction model solved the problem that current growth prediction studies can only predict the dynamic changes of a single phenotype and cannot visualize plant growth and development completely.The structural similarity(SSIM)of the first prediction time node was 94.45%,and the mean square error(MSE)was 6.02.The mean SSIM value for all prediction time nodes was 80.71%,the mean MSE value was 58.96,and the mean peak signal to noise ratio(PSNR)value was 28.79.The results showed that the predicted growth and development had high consistency and similarity with the actual growth and development.(3)Prediction model of growth and development for wheat based on ST-LSTM and MIM.To improve the prediction timeliness of the wheat growth and development prediction model,the wheat growth and development prediction model based on ST-LSTM and MIM was proposed based on the wheat growth and development prediction model based on STLSTM.Memory in Memory(MIM)was used to improve the ability to model the nonsmooth trend of growth and development.The predicted wheat growth and development was more similar to the actual growth and development.The SSIM increased to 95.67% at the first prediction time step.And the predicted results still maintained better similarity with the actual growth and development(>85% structural similarity)as the number of prediction time steps increased.(4)Determination model of the growing period for wheat based on DenseNet201 and BLSTM.To further evaluate the accuracy of wheat growth and development prediction,a determination model of growing period for wheat based on DenseNet201 and BLSTM was proposed.Different pre-trained convolutional neural networks and long short-term memory networks were combined to synthetically model the spatiotemporal characteristics of wheat growth and development and classify the growing period of each wheat image in the sequence.The highest determination accuracy of the growing period was 97.80,which was achieved by the model combining Densenet201 and bidirectional long short-term memory.Of these,the determination precisions of the tillering,re-greening,jointing,booting,and heading period were 100%,97.80%,97.80%,85.71%,and 95.65%,respectively.The predicted growth and development of wheat was similar to the actual growth and development.The results of growth stage determination were consistent.The errors in the number of leaves and the length and width of the minimum outer rectangle were small.The trends of projected area changes of predicted wheat growth and development were similar to the real ones.The results showed the predictive validity and accuracy of the growth and development prediction model for wheat.(5)Generalizability validation of the growth and development spatiotemporal prediction models for wheat.The wheat growth and development spatiotemporal prediction models were extended and applied to panicoid crops such as maize and barley as well as the model plant Arabidopsis thaliana to verify the generality and effectiveness of the wheat growth and development spatiotemporal prediction model.The SSIM between the predicted growth development and the real growth development obtained by the growth development prediction model based on ST-LSTM and MIM was significantly better.The SSIM decreased slowly with the number of prediction time steps increased.The prediction timeliness problem was alleviated to some extent.And there was also an improvement in the blur of the predicted images.And the MSE between the predicted growth and development results of the ST-LSTM-based growth and development prediction model and the real growth and development was smaller.In addition,this paper analyzed and determined the effective prediction durations of 5,5,and 14 days for wheat,Arabidopsis,and panicoid crops,respectively,and the best prediction interval of 1 hour for wheat.Those results will provide a reference for further research on the spatiotemporal prediction of growth and development in terms of dataset acquisition and model construction.In summary,the wheat growth and development spatiotemporal prediction model proposed in this paper could better predict the growth and development of plants(wheat,Arabidopsis thaliana,maize,millet,and sorghum)and resolve phenotypes(leaf number,projected area,plant height,and plant-type)in advance.Breeding experts further carried out the statistical analysis of phenotypic variation,the correlation analysis between different traits,and the locus localization of genetic loci regulating the target phenotypic variation and other efficient screening for excellent alleles.This will shorten the breeding trial cycle and make breeding trials faster and more efficient completion.At present,the growth and development prediction model for wheat was used for wheat grown in pots under controlled conditions in an artificial climate chamber,and the effective prediction of growth and development of wheat in pots provides a reference for further breeding in a field environment to shorten the breeding trial cycle and high throughput screening.In the future work,water,fertilizer,and environmental information will be integrated in the growth and development prediction study of population wheat in a complex field environment.