Research On Wheat Yield Prediction Method Based On Multimodal And Time-series Networks

Large-scale yield estimation in the field or plot during wheat filling period,contributing to high-throughput plant phenotyping and precision agriculture.To address the problem of poor yield estimation of wheat at large scale and multiple species,this study used a combination of multispectral and RGB drones to collect images from a high-tech agricultural park in Hefei,Anhui Province,China,to generate vegetation indices and canopy structure information.Five machine learning methods were used:partial least squares,random forest,support vector regression machine,BP neural network,and long and short-term memory network.The time-series data of remote sensing information generated using UAV multi-sensors were fully exploited during the winter wheat filling period to determine their potential for estimating wheat seed yield and to clarify the effect of multimodal data fusion on the improvement of yield estimation accuracy.From the perspective of distinguishing different heat-tolerant genotypes of wheat,the yield estimation of wheat grain filling period data was performed using a long and short-term memory network based on the preferred machine learning model.The main conclusions are as follows:(1)The spectral reflectance characteristics of vegetation indices showed a decreasing trend as the time of the filling period increased.For the time-series data of the wheat filling period,the long short-term memory network had the highest estimation effect,and the R~2 was improved by 0.21 compared with the BP neural network,which had the worst estimation effect.(2)Three types of wheat were distinguished into heat-tolerant wheat,moderately heat-tolerant wheat,and heat-sensitive wheat,and the long and short-term memory network with the best yield estimation effect was selected for regression prediction,and the yield estimation effect was better than that without distinguishing genotypes,respectively.Among them,sensitive wheat had the best prediction effect,and the model prediction accuracy:R~2 was 0.91 and RMSE%was 3.25%.(3)Among them,the yield prediction accuracy R~2 of fusing vegetation index with canopy structure information improved by about 0.07 overall than using vegetation index alone and was more adaptable to spatial variation.This study shows that differentiating wheat by different heat-tolerant genotypes,using a low-cost UAV for data fusion to extract canopy parameters,and applying a long-and short-term memory network for yield estimation of different heat-tolerant genotypes of wheat separately can provide relatively accurate and robust crop yield estimation models that can help make informed crop management decisions for harvesting and emergency forecasting of large areas of wheat.