基于深度学习语义分割模型的草地植被盖度估算对比研究

Comparative Study of Grassland Vegetation Coverage Estimation Based on Deep Learning Semantic Segmentation Models

草地植被盖度是评估草地生态系统健康和管理效果的重要指标。草地植被盖度常用经验目测或传统图像分类方法获得,存在植被盖度估算主观性较强、精度不够、模型泛化能力不足等问题。本研究利用深度学习语义分割模型对草地植被图像进行分割,并基于分割结果估算草地植被盖度,在像素尺度比较3种深度学习语义分割模型(Unet++、DeepLabv3+、Segformer)和Canopeo模型以及经典机器学习模型随机森林(Random Forest)在草地植被分割任务上的性能,结果表明:①Unet++模型分割性能最优,其平均交并比(MIoU)达0.79,F1分数(F1-score)达0.87,明显优于其他模型;相比之下Random Forest模型的表现较差,其MIoU为0.47,F1-score为0.55。②在图像尺度草地植被盖度估算中,Unet++、DeepLabv3+和Segformer模型估算的草地植被盖度均与实测草地植被盖度较为一致,估算精度明显高于Canopeo模型和Random Forest模型,深度学习语义分割模型中Unet++模型估算的草地植被盖度精度最高,决定系数(R2)达0.98,均方根误差(RMSE)低于3.8%,说明深度学习语义分割模型能够较为准确地估算草地植被盖度。③由于Unet++模型具有比其他模型更优的草地植被分割性能,因此将Unet++模型作为最终的草地植被盖度估算模型,并应用于荒漠草原、典型草原和草甸草原3个实验样地,模型可快速准确地获取样地的草地植被盖度。研究显示,Unet++等深度学习语义分割模型在草地植被盖度估算中表现出较高的准确性和适用性,能为草地植被盖度估算提供高效可靠的工具。

Grassland vegetation cover is a key indicator for assessing the health of grassland ecosystems and the effectiveness of their management.Traditionally,grassland vegetation cover has been measured through visual estimation or conventional image classification methods.These methods have drawbacks such as high subjectivity,low accuracy,and poor model generalization.In this study,we apply deep learning semantic segmentation methods to segment grassland vegetation images and estimate vegetation cover based on the segmentation results.We compare the performance of three semantic segmentation models(Unet++,DeepLabv3+,Segformer),the Canopeo model,and a classical machine learning model,Random Forest,for pixel level grassland vegetation segmentation.The results show that:(1)The Unet++model outperforms the others,achieving a mean intersection over union(MIoU)of 0.79 and an F1-score of 0.87,while the Random Forest model performs poorly with an MIoU of 0.47 and an F1-score of 0.55.(2)For estimating vegetation cover at the image level,the Unet++,DeepLabv3+,and Segformer models show high consistency with measured vegetation coverage,with significantly better accuracy than the Canopeo and the Random Forest models.Among the deep learning semantic segmentation models,Unet++achieves the highest estimation accuracy,with a coefficient of determination(R2)of 0.98 and a root mean square error(RMSE)of less than 3.8%.This demonstrates that deep learning models can accurately estimate grassland vegetation cover.(3)Given its superior performance,the Unet++model was selected as the final model for estimating grassland vegetation cover.It was applied to three experimental plots in desert steppe,typical steppe,and meadow steppe regions,efficiently and accurately estimating the grassland vegetation cover in these areas.This study demonstrates that deep learning semantic segmentation models like Unet++offer relatively high accuracy and applicability in estimating grassland vegetation coverage,providing an efficient and reliable tool for vegetation coverage estimation.