Research On Leaf Area Index Extraction And Yield Estimation Of Grapes Based On UAV Data

Grape is one of the earliest domesticated fruit types.Like other crops,real-time monitoring of various field-scale growth indicators is of great importance in decision-making for field management and obtaining high-quality,high-yield grape products.Although a large number of research have been carried out on the monitoring of vineyards with remote sensing technologies,the methods cannot be copied directly to use in another place due to differences in grape varieties,management styles,climate,and other factors.At present,there are a series of problems were remained in related research,such as inconsistent inversion methods,the need for a large number of manual sample collections,and incomplete factors considered in feature extraction.Therefore,based on previous research,using multi-temporal UAV RGB and multispectral data,taking the pergola-trained vineyards in Turpan region of China as the research object,this paper conducted in-depth research on leaf area index(LAI)extraction and yield estimation by using a variety of shallow and deep learning methods.The main works of this study are as follows:(1)The extinction coefficient suitable for extracting LAI of pergola-trained vineyards was determined.According to the LAI extracted by the direct measurement,the extinction coefficient which used in digital cover photography algorithm was adjusted,and a large amount of sample data was collected.Due to the lack of related instrument,the UAV multispectral images do not have location information,and it is hard to generate orthoimages accurately.Therefore,using shooting time differences,we extracted the coordinates and other parameters from corresponding RGB images and assigned them to multispectral images.Then generated orthophotos successfully,and provided reliable data sources for next research.(2)An ensemble learning model based on five shallow machine learning methods was constructed,and the LAI was extracted more accurately from UAV data.Using different sensors and features,various models were trained and evaluated,and come to following conclusions:(a)the prediction accuracy of ensemble learning models is higher than that of most base learners;(b)The textural features are better than the spectral ones,and the combination of them can improve model performances;(c)Models based on multispectral data performs better than the models based on RGB data;(d)Feature selection can improve the model performances.(3)An improved Res Net model was constructed,and a data augmentation method was proposed,which can effectively improve the accuracy and efficiency of LAI extraction.The image data was prepared by cropping the corresponding location of field sampling in orthophotos,and a CNN model was established by reconstructing the Res Net model.To improve the generalization ability of the model,in addition to use general image data augmentation methods such as rotation and flipping,a Mosaic data augmentation method suitable for regression problems was proposed.The following results were obtained:(a)The Mosaic method can improve the model prediction accuracies;(b)The large-sized input images do not necessarily improve the model performances,but increases the training time;(c)For small-sized input images(16×16,32×32,64×64),the Mosaic data augmentation method can improve the efficiency and accuracy of LAI extraction;(d)Compared with shallow machine learning methods,CNN model can predict LAI more accurately.Especially for higher or lower LAI values,deep learning has stronger predictive ability than shallow machine learning methods,but the requirements for hardware is relatively high.(4)A Conv LSTM-VTR grape yield estimation model based on convolution,LSTM,and ensemble learning was proposed.The spatiotemporal variation of LAI,textural/spectral features,and yield was analyzed.LAI and various features showed stronger correlation with yield in earlier growth stages,and decreased at the middle to later stages.By combining convolution,long short-term memory network(LSTM),shallow machine learning,and ensemble learning methods,the Conv LSTM-VTR yield estimation model was constructed,and the highest R~2 in prediction reached 0.589,which is better than that of the Conv LSTM model alone.However,this study still has some limitations,such as the inability to perform absolute radiometric correction on UAV data due to a lack of instruments,and the inability to collect more ground data due to the limited time and labor sources.However,this research is the first systematic study of typical pergola-trained vineyards in Xinjiang and provided a scientific and technical basis for future research and related precision management practices.