Datasets for SALMA: A Machine Learning Tool for Precise Leaf Morphology Measurements of Small Leaves

Structure Dataset for the software/paper "SALMA: A Machine Learning Tool for Precise Leaf Morphology Measurements of Small Leaves". Please see the code repository for executables and code: https://github.com/lnilya/salma This archive comprises three datasets based on the SALMA paper. Each dataset is contained in its own folder: 1. Main Dataset: This dataset is the primary dataset used in the paper, containing 3332 individual leaf images from 64 plant species in New Zealand, scanned at 600 dpi. Please see 2. Low-Contrast Dataset: The dataset consists of digital photographs under various adverse lighting conditions (e.g., blurriness, varying contrast, blue or red colouration). It contains 130 leaves from four alpine species. 3. Diverse Morphology Dataset: Consists of 212 leaves from six species containing damage from dessication and herbivory, compound leaves, narrow-leaved grasses and leaf fragments. This dataset includes some species extracted from Both et al. 2020. If not otherwise indicated, all images have been obtained through field work. Folder Structure Leaf Sheets: Contains the raw scans of leaves and their human-delineated masks. Subfolders correspond to species Files are numbered consecutively *.jpg files are scans *.png (same file name as jpgs) are human-delineated masks. Leaf Images: Single leaves extracted from the leaf sheets for easier evaluation. The results presented in our paper use these images. Each file contains a crop of a single leaf from a sheet. Sheet identity is ignored, and leaves are numbered consecutively. If the rectangular crop of the leaf contains a part of another leaf, it is overlayed with a white mask. *.jpg files are scans *.png (same file name as jpgs) are human-delineated masks Leaf Segmentations by Method: Contains the result of various algorithms on segmenting the individual leaves. The results presented in our paper are computed from these segmentations by comparing the contained *.png masks to the human-delineated masks in the Leaf Images folder. The subfolders are: Otsu: Applies Otsu's threshold on the intensity channel. Optimal: Picks an intensity threshold that most resembles the human-delineated mask output. This is not an automated algorithm, but an upper bound on any intensity-threshold-based algorithm (see our paper for discussion). FAMeLeS: Output of the ImageJ plugin published by Montès et al. 2024 ColsGrad/SALMA: Output of SALMA algorithm using colour and gradient information. This folder also contains a *.pickle file with the trained model for that species (and resolution). The number of leaves used for training varies between 1 and 4; the best-performing model and its results are contained in this folder. ColsGrad/SALMA_1..4: Output of SALMA algorithm using colour and gradient information and using 1,2,3 or 4 leaves for training, respectively. (This can be used to compare how SALMA's performance reacts to a varying number of leaves used for training the model.) Leaf Sheets - Corrected: Only applies to the Low-contrast dataset. This folder contains the leaf sheets after white balance correction using the ImageJ plugin from Mascalchi 2016. In the low-contrast dataset, these corrected images are used for subsequent steps (i.e., segmentation). Acknowledgements We acknowledge Rachael Lockhart, Olivia Bird, Indira Leon Garcia and Janelle Veenendaal for contributing data. We are grateful to Moses Njau, Vaustine Obura, and Job Munene for segmentation of the images.结构 本数据集配套于软件/论文《SALMA：用于小型叶片形态精确测量的机器学习工具》。 可执行程序与代码请参阅代码仓库：https://github.com/lnilya/salma 本压缩包包含基于SALMA论文的三类数据集，每类数据集均置于独立文件夹中： 1. 主数据集：本数据集为论文中使用的核心数据集，包含来自新西兰64个植物物种的3332张单叶图像，扫描分辨率为600 dpi。详见 2. 低对比度数据集：该数据集包含多种恶劣光照条件下拍摄的数码照片（如模糊、对比度异常、偏蓝或偏红色彩），涵盖4个高山物种的130张叶片图像。 3. 多样形态数据集：包含6个物种的212张叶片图像，这些叶片存在干燥损伤、植食性啃食痕迹、复叶、窄叶禾草以及叶片碎片等特征。本数据集部分物种取自Both等人2020年的研究。 若无特殊说明，所有图像均通过野外采集获取。 ## 文件夹结构 ### 叶片原始扫描件（Leaf Sheets） 存储叶片原始扫描图与人工标注的掩码（mask）。 - 子文件夹以植物物种命名 - 文件按顺序连续编号 - 后缀为.jpg的文件为扫描图像 - 后缀为.png的文件（与对应jpg文件同名）为人工标注的掩码。 ### 单叶图像（Leaf Images） 从叶片原始扫描件中裁剪出的单叶图像，便于后续评估。论文中展示的实验结果均基于该类图像生成。 - 每个文件均为单张扫描件中的单叶裁剪区域，忽略原始扫描件的编号信息，叶片按顺序重新编号。 - 若该矩形裁剪区域包含其他叶片的部分区域，则会叠加白色掩码进行标记。 - 后缀为.jpg的文件为扫描图像，后缀为.png（与对应jpg文件同名）的文件为人工标注的掩码。 ### 算法分割结果（Leaf Segmentations by Method） 存储各类算法对单叶进行分割的结果。论文中展示的实验结果通过将该文件夹内的.png格式掩码与单叶图像文件夹中的人工标注掩码进行比对计算得到。其子文件夹包括： - Otsu阈值法（Otsu）：在亮度通道上应用Otsu阈值分割算法。 - 最优阈值法（Optimal）：选取与人工标注掩码输出最为匹配的亮度阈值。该方法并非自动算法，而是所有基于亮度阈值的分割算法所能达到的性能上限（详细讨论参见本论文）。 - FAMeLeS：Montès等人2024年发布的ImageJ插件的分割输出结果。 - ColsGrad/SALMA：基于色彩与梯度信息的SALMA算法分割输出结果。该文件夹同时包含对应物种（及分辨率）的训练模型文件（格式为.pickle）。用于训练的叶片数量介于1至4张之间，本文件夹存储了性能最优的模型及其分割结果。 - ColsGrad/SALMA_1..4：分别使用1、2、3、4张叶片进行训练的基于色彩与梯度信息的SALMA算法分割输出结果（可用于对比不同训练样本数量对SALMA算法性能的影响）。 ### 校正后叶片原始扫描件（Leaf Sheets - Corrected） 仅适用于低对比度数据集。该文件夹存储了使用Mascalchi 2016年发布的ImageJ插件进行白平衡校正后的叶片原始扫描件。在低对比度数据集的后续处理（如分割）中，均使用该校正后的图像。 ## 致谢 感谢Rachael Lockhart、Olivia Bird、Indira Leon Garcia与Janelle Veenendaal提供数据集。同时感谢Moses Njau、Vaustine Obura及Job Munene完成图像标注工作。