Research On High-resolution Prediction Algorithm For Maize Growth And Development Based On STRPM

In the field of intelligent breeding,predicting the phenotypic characteristics of plant growth and development can help researchers understand the response mechanisms and adaptability of plant growth and development,which can help researchers design more effective breeding programs,accelerate the breeding of new varieties,improve crop yield and quality,and make important contributions to food security.Phenotypic prediction and growth simulation of plants are difficult,especially to make predictions at high resolution,because high-resolution images contain more visual details.To address these problems,this paper proposes a maize growth and development prediction algorithm based on Spatiotemporal Residual Predictive Model(STRPM)on the maize phenotype time-series dataset Panicoid Phenomap-1.The algorithm can better predict the growth state of maize at future time points and can visualize the image of maize growth and development.In this paper,the STRPM algorithm is improved in two main aspects as follows:(1)First,to simulate the slow growth process of plants.In this paper,3D convolution is used to construct encoders to extract temporal and spatial information from the maize timeseries dataset separately,which can better extract the features of temporal and spatial dimensions in the continuous crop growth process and can improve the prediction performance of the model at longer time steps.(2)Second,in order to train a more stable model and predict a more realistic image.In this paper,we use WGAN-GP(Wasserstein GAN with Gradient Penalty)to improve the adversarial training method of STRPM.Moreover,this paper adds VGG-16 perceptual loss function to the generator.VGG-16 perceptual loss function takes into account not only the pixel-level differences of images,but also the high-level semantic information,which can make the model better capture the semantic information between images and thus improve the quality and accuracy of predicted images.Experimentally,it is proved that the improved STRPM-based maize growth and development prediction model at high resolution in this paper has a mean MSE of 64.65,a mean SSIM of 0.9626,a mean PSNR of 24.64,and a mean LPIPS of 10.03 for predicting the next eight images at a time step of 8,which is better than other advanced models.Moreover,the resolution of the predicted images in this paper reached 512×512,which is nearly 4 times higher than the previous related works at home and abroad,and the image reproduction is also higher.In this paper,we have initially achieved the prediction of plant growth and development,and this research can help accelerate intelligent breeding.By predicting the growth and development of different kinds of plants,we can find the most suitable plant species for specific environments,improve the adaptability and resilience of plants,reduce the number and cost of experiments,and thus accelerate the promotion of new varieties.