Study On Stratified Yield Estimation Of Densely Planted Cotton Field By UAV Low Altitude Imaging

In China,cotton is a primary textile raw material and an important economic crop.Moreover,timely and accurate estimation of cotton production is important for guiding cotton production and guaranteeing cotton trade.Currently,the main yield estimation techniques are based on manual sampling and statistical methods.These methods are complex,labor-intensive and have long lead times.In recent years,advances in image processing technology and the widespread use of UAV image collection platforms have provided technical support for faster and more accurate crop yield estimation.However,due to the high planting density and small plant spacing in close planting cotton fields,the lower layer of cotton is heavily obscured,which makes the imaging-based yield estimation method difficult.Based on this issue,this study was con-ducted by taking close planting cotton field in Xinjiang during defoliation period as the research object,combining deep learning technology with cotton yield estimation as the research objective,and focusing on three aspects of UAV low-altitude image acquisition,boll stratification discrimination and cotton yield esti-mation.The main work of the study is shown as following:(1)To improve the resolution and acquisition efficiency of images.A five-point sampling method based on UAV low-altitude imagery was proposed in this study.The method based the random sampling principle and five representative low altitude images of the cotton block area were collected to represent the infor-mation of the whole cotton field.Compared with the method of whole-area image acquisition of cotton field,this method has higher acquisition efficiency and it is more suitable for image acquisition in Xinjiang close planting field.(2)To achieve accurate discrimination of cotton boll areas in close planting cotton field images.In this study,the cotton field images were firstly segmented into whole boll regions by using the idea of hierarchical discrimination.Secondly,based on the overall segmented image,the upper boll region segmentation and the lower region segmentation were performed.For the overall boll region segmentation,the compiler and de-coder blocks based on the Segnet network framework were redesigned.An extended cotton boll segmenta-tion framework(CD-Seg Net)with four neural network models(model 1,model 2,model 3 and model 4)was obtained and trained on the same data set.The four trained neural network models were evaluated using the mean intersection over union(MIOU),Fl-score,class pixel accuracy(CPA)and Recall.The results revealed that among the four models,model 4 had the best MIOU and CPA of 77.13%and 90.82%,respectively.Finally,when compared with the traditional machine learning segmentation algorithms(SVM and RF),the CD-Seg Net segmentation algorithm outperformed the traditional algorithms in all evaluation metrics.(3)For the upper cotton boll segmentation,CMYK-UNet segmentation method was proposed in this study.The CMYK-K color space component of the cotton field image was extracted to enhance the upper boll image features,and the CMYK-K component map was used as the input image.A Unet segmentation model was then trained and tested.The results demonstrated that the CMYK-UNet segmentation method outperformed the traditional method(OTSU)in all evaluation indexes,and the accuracy of upper boll seg-mentation was 93.55%.(4)Since cotton yield is derived from cotton boll,the phenotypic traits of cotton boll are closely related to the yield estimation results.In this study,the cotton boll was divided into two layers by using the idea of stratification.A linear regression model for yield estimation was constructed for each layer of cotton boll image rate.Based on the constructed yield estimation model,yield estimation was carried out in four test cotton fields,and the estimated yield values were compared with the harvested yield of cotton pickers.The results showed that the average relative error of yield estimation was 5.28%and 6.22%,the coefficient of determination(R~2)was 0.99 and 0.95,and the root mean square error(RMSE)was 353.6 kg/ha and 451.6kg/ha for the stratified yield estimation model and the non-stratified yield estimation model,respectively.Finally,the comparison of the results of the two estimation models showed that the hierarchical estimation had better robustness and accuracy.