The quantification and identification of ground cover plays a key role in erosion modelling, weed measurement, plant disease identification and other environmental applications. Currently, a variety of methods are used to mechanically classify digital images for ground cover. Only a few of these methods can distinguish green vegetation, straw/dormant vegetation, and exposed soil using only the Red-Green-Blue (RGB) spectrum. This research presents an approach to classifying ground cover using standard JPEG images and readily available Matlab (2018b) functions. The approach uses block segmentation, as opposed to pixel-wise or object-based segmentation, and compares multiple machine learning approaches with varying pixel block size and classification acceptance thresholds. The most successful classification approach found through this study was the decision tree algorithm with a 70-pixel block size and 60% classification acceptance threshold. Images were reduced to three feature descriptors: colour, texture, and oriented gradients to represent the respective RGB spectrum for an image. Both the training set and test set images used in this research came from field and greenhouse studies done between 2016 and 2019. The produced classifications were compared to manual coverage classifications using Samplepoint, a grid-based method, with R-squared values of 0.86 for green vegetation, 0.87 for straw/dormant vegetation, and 0.96 for exposed soil, respectively. This method showed strong performance for images containing exposed soil and either green vegetation or straw/dormant vegetation. The method was less effective for images with large quantities of both green vegetation and straw/dormant vegetation likely due to their similar shape.

中文翻译：

---

使用图像特征描述符和决策树学习算法测量土壤覆盖率

地面覆盖的量化和识别在侵蚀建模、杂草测量、植物病害识别和其他环境应用中起着关键作用。目前，有多种方法用于对地面覆盖的数字图像进行机械分类。这些方法中只有少数可以仅使用红-绿-蓝 (RGB) 光谱来区分绿色植被、稻草/休眠植被和裸露的土壤。本研究提出了一种使用标准 JPEG 图像和现成的 Matlab (2018b) 函数对地面覆盖进行分类的方法。该方法使用块分割，而不是逐像素或基于对象的分割，并将多种机器学习方法与不同的像素块大小和分类接受阈值进行比较。通过这项研究发现的最成功的分类方法是具有 70 像素块大小和 60% 分类接受阈值的决策树算法。图像被简化为三个特征描述符：颜色、纹理和定向梯度，以表示图像的相应 RGB 光谱。本研究中使用的训练集和测试集图像均来自 2016 年至 2019 年间进行的实地和温室研究。使用基于网格的方法 Samplepoint 将生成的分类与手动覆盖分类进行比较，其中 R 平方值为 0.86绿色植被，秸秆/休眠植被分别为 0.87，裸露土壤为 0.96。该方法对于包含裸露土壤和绿色植被或稻草/休眠植被的图像显示出强大的性能。