| LAI(Leaf Area Index)is half of the total leaf area of vegetation under unit projected area.It is an important parameter in agricultural science and remote sensing science.How to accurately and quickly measure leaf area index has always been an important research subject.This paper focuses on the LAI measurement method based on digitial hemispheric photography.To our knowledge,it is the first time that deep learning technology is applied to the measurement to solve the difficult task of identifying and segmenting vegetation images,and to provide a feasible solution for the accurate and reliable LAI measurement instruments.This paper mainly researches the problem of LAI measurement based on deep learning and digitial hemispheric photography from the following aspects:(1)Discuss several key issues in digitial hemispheric photography: data acquisition and processing,segmentation decision classification,image segmentation,and the widely used Lambert-Beer law.It is clear that the accuracy of image segmentation is an important factor to the final LAI value.Facts affirmed that the adaptive Otsu method and the HSV threshold segmentation method can also obtain better segmentation results under certain conditions.Based on empirical logic and linear discriminant analysis,two feasible traditional segmentation decision methods are given.(2)The image segmentation decision problem in digitial hemispheric photography is analyzed,and then a deep learning algorithm is used to classify the vegetation pictures to optimize the LAI measurement method.Inception model was used for image classification training,and the classification decision results of this method were compared with traditional methods.This method greatly improved the accuracy of LAI measurement,and the correlation coefficient between the measured value and the value by LAI2200 increased from 0.47 to 0.91.(3)Try to use multiple deep learning algorithms to solve the problem of vegetation image segmentation to increase picture utilization.Then through the Semantic Segmentation Suite,two feature extraction models and seven image segmentation models are connected.The performance of the model is compared and analyzed from different angles: the same data with different model and the same model with different data.It is found that the segmentation accuracy of most dual-end image semantic segmentation models exceeds 0.95,but it takes up more then 50 s to inference on the embedded device Jetson TX2,which cannot meet the actual measurement needs.(4)Under the conditions of time and performance,Pix2 pix model is selected as the device transplantation model.The Pix2 pix model is derived from the c_GAN model.Its adversarial training ideas and lightweight structure ensure the performance and low computing requirements of the model.Meanwhile,it points out the problems of noise and blurring on the binary problem.The Dense CRF is added to the pix2 pix model’s generating structure through the average field approximation algorithm.The improved model can eliminate noise,refine edges,and obtain clearer image segmentation results.(5)Develop the ‘LAI_analysis’ software by ourself and transplant it to Nvidia Jetson TX2.Meanwhile,TensorRT technology was used to optimize the inference calculation speed,so that the speed of single inference and calculation reached 8s.Finally,a systematic experiment was performed on the experimental prototype,which shows that the prototype basically achieves the preset goals of the subject. |
