Combining Deep Learning And Object-Oriented Approachs To Estimate Photosynthetic And Non-Photosynthetic Vegetation Cover From Unmanned Aerial Vehicle Images

Soil erosion is a pressing global environmental issue,and accurate monitoring of regional photosynthetic vegetation（PV）and non-photosynthetic vegetation（NPV）cover and spatial distribution is crucial for improving the accuracy of soil erosion models.Traditional methods for extracting PV and NPV,such as ground surveys and remote sensing inversions,have limitations.However,the emergence of low-altitude remote sensing technology using unmanned aerial vehicles（UAVs）has opened up new opportunities for vegetation information extraction.Currently,most UAV-based vegetation studies have focused on photosynthetic vegetation,while research on non-photosynthetic vegetation,which plays a significant ecological role and contributes to soil and water conservation,is scarce.Additionally,studies on PV and NPV information extraction from UAV images using object-oriented and deep learning approaches are limited.Although three deep learning semantic segmentation models,namely Deep Lab V3+,PSPNet,and U-Net,have shown promise in vegetation extraction from UAV images,they heavily rely on large amounts of labeled data,which presents practical limitations.Furthermore,UAV images often contain shadows,making it challenging to accurately classify features in shadowed areas.Studies that consider PV and NPV classification in shadows are relatively rare.Therefore,further research is needed to explore effective methods for extracting PV and NPV information from UAV images,particularly in shadowed areas,to enhance the accuracy of vegetation mapping and monitoring for soil erosion assessment.In this study,a DJI Phantom 4 Pro UAV was employed to conduct an annual aerial survey in 2019 of six representative vegetation sample areas located on the Loess Plateau,including Yulin overgrown sandy sparse woodland,Hengshan grassland,Wuqi lemon woodland,Wuqi artificial oil pine woodland,Huangling secondary woodland,and Yangling agricultural land sample area.Building upon established methods for UAV-based vegetation information extraction,this study proposes a framework that combines deep learning semantic segmentation models with object-oriented classification methods for fast extraction of PV and NPV.The feasibility of incorporating shadow feature classification into the framework is also explored.The segmentation results of three different models are compared using overall classification accuracy,Kappa coefficient,and other indicators in the confusion matrix.The accuracy and generality of vegetation information extraction for different vegetation types are evaluated,and the results are compared with ground survey data and satellite-based remote sensing inversion for the same period.The main conclusions of this study are as follows:（1）A framework for fast extraction of UAV PV and NPV is proposed.The framework includes three main steps.Firstly,an object-oriented optimal algorithm based on scale seeking and feature preference is used to generate pre-classification results.Secondly,the optimal classification results from the object-oriented approach,combined with manual correction,are used to create segmented datasets for PV and NPV using a size of 512 pixels and a step size of 256 pixels.Finally,Res Net 50 is selected as the backbone network to build three deep learning semantic segmentation models,namely Deep Lab V3+,PSPNet,and U-Net.（2）The experiments on scale seeking and feature preference show that the optimal parameter combinations for multi-scale segmentation of UAV images in the six typical vegetation sample areas are a shape factor of 0.1,a tightness factor of 0.5,and a segmentation scale of 60-84.And compared with spectral features,texture features,and geometric features,NGRDI,EXG and VDVI features are of higher importance in UAV PV and NPV extraction.（3）A comparison of the classification results of three algorithms,RF,SVM,and KNN,showed that the classification results of the object-oriented random forest algorithm performed best in classifying features in PV and NPV of six typical vegetation type samples,with an overall classification accuracy range of 88.3%-95.2%and a kappa coefficient range of 0.80-0.93.The classification results of RF algorithm with manual correction were selected to produce a label database,which can provide a good data base for the establishment of the subsequent deep learning semantic segmentation model.（4）The comparison results of the semantic segmentation models Deep Lab V3+,PSPNet and U-Net for each class showed that Deep Lab V3+had better performance in many indexes for the sample area of sandy sparse forest in Yulin,with the overall classification accuracy of91.9%and the Kappa coefficient of 0.87,which was higher than those of the PSPnet and U-Net models.PSPnet and U-Net models.For the five vegetation type sample areas of Hengshan Grassland,Wuqi Lime Strip Woodland,Wuqi Artificial Oil Pine Woodland,Huangling Secondary Woodland and Yangling Agricultural Land,PSPNet performs better in many metrics among the three deep learning semantic segmentation models,with overall classification accuracy of 86-92%and Kappa coefficient of 0.76-0.86.（5）Deep Lab V3+was selected for the model generalization evaluation study of Yulin overgrown sandy woodland sample area;PSPNet was selected for the model generalization evaluation study of five vegetation type areas,namely,Hengshan grassland,Wuqi lemon strip woodland,Wuqi artificial oil pine woodland,Huangling secondary woodland and Yangling agricultural land.The results show that the optimal network model established by using the region A comparison experiment can well extract the PV and NPV of the sample areas of different locations in the same period of the corresponding land type,the sample areas of different periods in the same location,and the sample areas of different locations in different periods,and the selected optimal model has good generality.（6）Based on the constructed optimal model,PV and NPV were automatically extracted from long-term monitoring areas of six vegetation type sample areas,including Yulin overgrown sandy sparse woodland,Hengshan grassland,Wuqi lemon strip woodland,Wuqi artificial oil pine woodland,Huangling secondary woodland,and Yangling agricultural land.The dynamic distribution of PV and NPV in these six vegetation types at different times throughout the year revealed that,except for Yangling farmland,the PV in the other five vegetation types showed a unimodal distribution,with the highest values occurring during the middle of the growing season from July to September,and the lowest values occurring at the beginning of the growing season from February to March.The highest values of NPV occurred at the end of the growing season,from late October to February-March of the following year.The findings of further analysis revealed that the estimates of f _PV and f _NPV for the six vegetation types using the UAV images were in close agreement with surface observations,with R² values ranging from 0.8 to 0.9 for f _PV and 0.5 to 0.9 for f _NPV.Additionally,the correlation between the estimates obtained from the constructed optimal model and the image trilateration model based on Sentinel-2A images was also high.These results indicate that the optimal model developed in this study can effectively extract PV and NPV information from UAV remote sensing imagery and provide reliable spatial distribution and change information for long-term monitoring areas.This approach helps in understanding the dynamic changes in vegetation cover in different vegetation types on the Loess Plateau.In summary,this study proposes and demonstrates the use of object-oriented and deep learning semantic segmentation methods for unmanned monitoring of photosynthetic and non-photosynthetic vegetation,allowing for quick extraction of PV and NPV cover and spatial distribution information in typical vegetation type areas on the Loess Plateau.Researchers can utilize these methods to gain insights into the composition of photosynthetic and non-photosynthetic vegetation in different vegetation types on the Loess Plateau,with the ultimate goal of accurately monitoring and evaluating soil erosion in this region.This method presents a new source of data for precise monitoring and evaluation of soil erosion on the Loess Plateau.