Research On Retrieval Algorithm For Canopy Parameters Of Typical Vegetation Based On Deep Learning And Hemispheric Images

Vegetation canopy parameters can reflect vegetation canopy structure,function and growth status,and are important input factors for many ecological research models,which are important for climate change and ecosystem research.Among the many methods for vegetation canopy parameter retrieval,the indirect measurement method based on hemispheric images is widely used,but their intermediate processes such as image segmentation and estimation of Clumping index can introduce errors.To address the above problems,this research conducts a canopy parameter retrieval study including leaf area index and fractional vegetation cover based on deep learning with typical vegetation as the research object and hemispheric images of vegetation canopy as the data to improve the retrieval accuracy.The details of the research contents are as follows:(1)A total of 17 simulation models of the same typical vegetation types as those in the real scenario experiments were built based on 3D modeling software,and multiple sampling points were set to obtain long time series hemispheric images of vegetation canopy in a day,and measured parameters of the vegetation canopy.The experimental dataset was constructed by combining the images and canopy parameters obtained in the real scenario experiments.(2)A lightweight hemispherical image segmentation network DHPGU-Net,which is suitable for a variety of typical vegetation,is constructed to improve the accuracy of vegetation segmentation in hemispherical images.The network combines codec structure and lightweight network structure.The experimental results show that DHPGU-Net has an Io U of over 96.95%,a Precision of over 97.80%,and a Recall of over 98.93% in multiple datasets,with high accuracy of model segmentation.In the test data of different scenes,compared with the retrieval results based on the threshold segmentation method,in the regression analysis of the simulation environment measurement value and the LAI-2200 reference value,the retrieval results based on the DHPGU-Net segmentation method improve the fractional vegetation cover correlation coefficient by 0.03-0.14,the leaf area index correlation coefficient by 0.03-0.22,the RMSE decreases by 0.28-0.81,and the MAE decreases by 0.28-0.62,which shows that this method can effectively improve the retrieval accuracy.(3)An end-to-end deep regression model for canopy parameter retrieval in hemispheric images was constructed,and the effects of using multiple networks as feature extraction structures,model regression layers,loss functions,and optimizers on the regression results were investigated.In the canopy parameter retrieval experiments,the regression analysis results with the measured values of the simulated environment and the reference value of LAI-2200 showed that the correlation coefficients of leaf area index and fractional vegetation cover were over 0.91 and 0.94,and compared with the segmentation-based retrieval method,the correlation coefficients of leaf area index increased by 0.12-0.32,RMSE decreased by 0.33-0.98,and MAE decreased by 0.28-0.83,which further improved the retrieval accuracy of vegetation canopy parameters in hemispheric images.(4)A lightweight deep regression network of vegetation canopy parameter was constructed,and its effectiveness and accuracy in retrieval of vegetation canopy parameters were verified in experimental data.A prototype vegetation canopy parameter retrieval system based on Py Side2 and Qt Quick was developed and deployed to run in edge devices.The system integrates the improved segmentation-based and the proposed regression-based canopy parameter retrieval methods in this research for vegetation canopy parameter measurement.