The InvL Dataset of Invasive Species from drones in a Land Environment

<strong>Summary</strong> An original dataset for semantic segmentation, InvL, is introduced, which to the best of the authors' knowledge, is the first dataset of its kind with pixel-level annotations pertaining to invasive species (Himalayan balsam) in a land environment from drones with height 10-30 m (GSD 3-13 mm) using optical imaging and having diverse outdoor blur, noise and contrast. <strong>Objectives</strong> Automated detection and quantification of invasive species from drones would enable more efficient mapping using simple low-cost technologies. This would help to <em>remove invasive and later protect species. </em>However, it is difficult to distinguish invasive species from common ones from higher altitudes using drone images. Training a machine learning algorithm to accurately detect a given invasive species from images taken in the field requires a massive amount of human-generated training data. Objective is to encourage people to open share images that can be used to develop technology. We are interested expand database and add author for every 1 GB added to the new version InvL or every 200 hours spent to improve annotation quality or annotation type. Objective is to expand establish a reference dataset for automatic extraction using gold standard training sets to gain Intersection over Union (IoU) over 0.8 from drone height over 30 meters with least GSD using low cost drones. <strong>Data description</strong> This data set contains images of Himalayan balsam (Impatiens glandulifera) taken mostly with the DJI Mavic 2 PRO in Finland from heights 10, 15, 20 and 30 meters, low speed 1- 2 m/s, no filters, 90 degrees angle, still images, GSD varying from 3-13mm, image size ~15 MB, between July (usually no flowers) and August (flowers). <strong><em>Table 1. Drone images of invasive species</em></strong> <strong><em>Label</em></strong> <strong><em>Name of data file/set</em></strong> <strong>Dataset size (GB)</strong> <strong>File type (extension)</strong> <strong>Data repository and identifier</strong> <strong>IoU estimate </strong><strong>- model</strong> <strong>Comment</strong> <em>Himalayan_balsam</em> <em>InvL/Himalayan_balsam</em> <em>../ann – contains annotations</em> <em>…/ann/H10m_Lahti_Peitsikatu_H10m_10072020</em> <em>H10 = 10 m UAV heigth</em> <em>Lahti = city</em> <em>10072020 (last) = date of fligth</em> <em>…/original – orginal UAV images, folder structure same as for “ann”</em> 2 InvL.zip (folder structure with annotations and original images) <em>Himalayan balsam </em> <em>IoU = 0.53, FCN, 15 m with augmentation IoU=0.37, FCN, 30 m</em> <em>Himalayan balsam, green</em> <em>Elevation varies in area +- 5m, so actual UAV flight height varies</em> <em>Each image around 15 MB</em> <strong>Limitations</strong> Annotations do not indicate width or margins. There is no indication of confidence of annotations. Blur, noise or contrast values are not calculated. Mostly species are easily visible and images with bright sunlight have been removed in selection of images. Annotator needs good and tested instruction to annotate images. Annotator needs to do work for tens of hours with developed and tested instructions to guarantee good quality. After image annotator each image should be checked by an “expert” and reannotated if necessary (silver standard, if this is done twice it is gold standard). If expert is used each image should be annotated once (silver standard) and twice (gold standard). Annotation time for each 15 MB image should be around one hour. Image GSD should be around 5 mm and less for smaller species. In different field areas species color and size varies and mixed vegetation may cause problems. In images there are flowers and non-flowers. <strong>Abbreviations</strong> FCN: a fully convolutional neural network.