期刊
METHODS IN ECOLOGY AND EVOLUTION
卷 13, 期 11, 页码 2363-2371出版社
WILEY
DOI: 10.1111/2041-210X.13984
关键词
canopy density; canopy hemispherical photography; computer vision; convolutional neural network; forest monitoring; machine learning; UNet; vegetation density
类别
资金
- Bundesamt fur Naturschutz [3518830200]
- Bundesministerium fur Bildung und Forschung [01LN1301A]
- Commonwealth Scientific and Industrial Research Organisation
- Deutsche Forschungsgemeinschaft [491292795]
- Leibniz Institute for Zoo and Wildlife Research (IZW) in the Forschungsverbund Berlin eV Berlin, Berlin, DE
- Point Defiance Zoo and Aquarium
- United States Agency for International Development [72044020CA00001]
- University of Western Australia
This paper introduces an R package imageseg for image segmentation using convolutional neural networks, which can be used to evaluate forest structural metrics with high accuracy and generalization ability across different forest types and biomes.
Convolutional neural networks (CNNs) and deep learning are powerful and robust tools for ecological applications, and are particularly suited for image data. Image segmentation (the classification of all pixels in images) is one such application and can, for example, be used to assess forest structural metrics. While CNN-based image segmentation methods for such applications have been suggested, widespread adoption in ecological research has been slow, likely due to technical difficulties in implementation of CNNs and lack of toolboxes for ecologists. Here, we present R package imageseg which implements a CNN-based workflow for general purpose image segmentation using the U-Net and U-Net++ architectures in R. The workflow covers data (pre)processing, model training and predictions. We illustrate the utility of the package with image recognition models for two forest structural metrics: tree canopy density and understorey vegetation density. We trained the models using large and diverse training datasets from a variety of forest types and biomes, consisting of 2877 canopy images (both canopy cover and hemispherical canopy closure photographs) and 1285 understorey vegetation images. Overall segmentation accuracy of the models was high with a Dice score of 0.91 for the canopy model and 0.89 for the understorey vegetation model (assessed with 821 and 367 images respectively). The image segmentation models performed significantly better than commonly used thresholding methods, and generalized well to data from study areas not included in training. This indicates robustness to variation in input images and good generalization strength across forest types and biomes. The package and its workflow allow simple yet powerful assessments of forest structural metrics using pretrained models. Furthermore, the package facilitates custom image segmentation with single or multiple classes and based on colour or grayscale images, for example, for applications in cell biology or for medical images. Our package is free, open source and available from CRAN. It will enable easier and faster implementation of deep learning-based image segmentation within R for ecological applications and beyond.
作者
我是这篇论文的作者
点击您的名字以认领此论文并将其添加到您的个人资料中。
推荐
暂无数据