Study On Cotton Yield Estimation Based On Multi-time Series UAV Image

Timely and accurate estimation of cotton yield not only plays an important role in cotton production management,but also has important strategic significance for cotton trade and agricultural policy planning.Therefore,it is necessary to find a fast,accurate and efficient cotton yield estimation method.At present,remote sensing technology is mainly used to estimate the yield of large-area cotton field,and single growth period and single temporal satellite remote sensing images are widely used,and a lot of research results have been achieved.However,satellite remote sensing usually has the problems of low resolution and poor timeliness.At the same time,the single temporal remote sensing image is difficult to fully consider the growth of crops in different key growth periods,and the yield estimation results are one-sided and lack of reliability.Based on this,in this study,UAV remote sensing platform was used to collect the canopy images of three important growth stages of cotton,namely seedling stage,bud stage and flowering stage.Combined with machine learning algorithm and deep learning technology,three growth parameters of plant height,leaf area index and aboveground biomass in each growth stage were obtained,and single temporal yield estimation model and deep learning yield estimation model based on multi temporal images were established,The practicability and accuracy of the two models were comprehensively analyzed,and the best yield estimation model was selected.The main research contents and achievements are as follows:（1）Combined with convolution neural network,cotton growth parameters were monitored.In this paper,the cotton canopy image was used as input,and three kinds of convolutional neural network（CNN）frameworks including Alex Net,VGGNet and Goog Le Net were used to build the growth parameter monitoring model,so as to realize the accurate monitoring of plant height,leaf area index and aboveground biomass.The results showed that the Alex Net model was the best growth parameter monitoring model at the flower bud stage,and the R² of plant height,leaf area index and aboveground biomass were 0.854,0.964 and 0.951,respectively.At the same time,according to the growth characteristics of cotton growth period and canopy image characteristics,the monitoring effect of the three models in the bud and flowering period was comprehensively analyzed.The results showed that the monitoring effect of the three models in the flowering period was better than that in the bud period.（2）Based on traditional machine learning,single growth period cotton yield estimation is realized.Taking three growth parameters of cotton seedling stage,bud stage and flowering stage as yield estimation factors,four machine learning algorithms including k-nearest neighbor,random forest,support vector machine and BP（back propagation）neural network were selected to construct yield estimation models based on single growth stage.The results showed that the overall performance of the random forest model was the best,and the yield estimation accuracy of the bud stage was the highest,with the determination coefficient of 0.781.The determination coefficients of seedling stage and flowering stage were 0.751 and 0.701,respectively.（3）Cotton yield estimation based on time series.Time series data sets were generated from the images of three growth stages of cotton.The characteristics of recurrent neural network in time series feature extraction were given full play.Considering the influence of network depth on feature extraction effect,LSTM and Bi LSTM models were constructed to realize yield prediction based on time series data.Based on the advantages of Bi LSTM model in spatial feature extraction,CNN-Bi LSTM model with serial structure was constructed to realize cotton yield prediction based on spatial and temporal scales.The results show that the CNN₁₄-Bi LSTM model with convolution layer depth of 14 was the best,which is superior to LSTM and Bi LSTM models.Its R² was 0.885,RMSE was147.167 g,and MAPE was6.711%.Compared with the single growth period cotton yield estimation model,the accuracy of CNN₁₄-Bi LSTM model was also higher than that of Random Forest model,which further confirms that the best yield estimation model is the multi time series cotton yield estimation model based on deep learning.In this paper,taking cotton as the research object,using UAV platform to collect images of seedling stage,bud stage and flowering stage,and combines with deep learning technology to realize cotton yield estimation based on time series images.The research results can provide reference for similar crop yield estimation.