Phuc Le
Intern at Fauna & Flora International - Vietnam Programme
June - September 2021
This report summarizes my experimentation process in evaluating the usefulness and accessibility of MegaDetector and Wildlife Insights for Fauna & Flora International - Vietnam Programme in their camera-trapping projects. MegaDetector is an off-the-shelf machine learning model that can detect the presence of animals in camera-trapping images, and Wildlife Insights is a developing platform aiming for an end-to-end technological solution for camera-trapping projects around the world. The results show that it is possible to incorporate these technologies into the current workflow to speed up the process while informing the conservation decision. Concerning the usefulness, Wildlife Insights, in its early stage, can be used as an accessible window into a dataset and sharing medium among collaborators. Meanwhile, MegaDetector gives a "warp speed" button which can be pressed when it is short of time and resources and a quick answer is needed. For accessibility, both can be used without cumbersome installation. One can use Wildlife Insights directly on the web. And although MegaDetector is a Machine Learning model, there is an alternative option, e.g., Google Colaboratory, to run it on the web at a relatively low cost. The application of technological solutions into the current human-based workflow is promised to scale up camera trapping projects and push them towards global conservation goals.
Camera-trapping is an effective way to quantitatively monitor the presence of animals in an ecological system in a specific time period. A camera trap project produces thousands of data points that can be leveraged to evaluate the health of an ecological system. However, due to the large availability of data, there exist difficulties of data pre-processing, selecting images that contain animals, and data management, especially after the completion of data analysis. These difficulties have been presented in the current workflow of Fauna & Flora International - Vietnam Programme, hereafter referred to as FFI Vietnam, which I will try to solve in this report.
After the cameras are retrieved from the field, camera trap images are extracted from the SD cards into a staff computer and uploaded to OneDrive into a clearly-defined project folder. At the same time, an Excel spreadsheet which has the information of camera-trapping deployments such as date-time setup, date-time retrieval, geographic location, etc,. is finalized. After that, an FFI Vietnam staff member will reorganize the OneDrive project directory following a folder structure convention; this task is usually done by using camtrapR (an R library designed for camera-trapping). When the images are properly stored on OneDrive, a group of conservationists will start cataloging the images on digiKam (Figure 1). digiKam is an all-purpose image management software that allows a flexible hierarchical tagging system, which is suitable for tagging species to different taxonomic levels. As they have to go through the entire dataset, it often takes more than two weeks to catalogue all images for a group of three conservationists. After that, they start analyzing the data. The analysis is kept by the conservationists in charge but is not included in the shared storage.
Figure 1. An example of cataloging images with digiKam.
Although the protocol is clear and systematic, there is still room for improvement. There are two ready-made tools, namely MegaDetector and Wildlife Insights, that can help reduce the logistical burden for conservationists. This will allow them to use their time more effectively, by moving more quickly from insights to conservation actions. In this report, I demonstrate what I have done during my internship at Fauna & Flora International - Vietnam Programme (June to September 2021) to explore these off-the-shelf technologies that may help solve some practical issues inherent in this protocol. The assessment and proposed workflow are based on and kept suited for camera-trapping projects at FFI Vietnam, but other organizations can adapt it accordingly.
Traditionally, after receiving the data, a group of conservationists needs to collectively go through the whole dataset to find the images with animals and identify them. In this procedure, about 75% of the time they have to look at blank images (this statistic is derived from FFI Vietnam's camera trap project in Kong Plong during 2019-2020). Blank images can be caused by animals that move too quickly past the camera trap and are missed, or more frequently by 'false triggers', i.e. caused the movement of non-animal objects such as falling leaves or swinging grasses. Black images do not contribute any information to the ecological understanding of the location and are discarded during the data analysis phase. Although human-intensive species identification is crucial to guarantee the credibility of post-processed data and its readiness for analysis, its time-consuming nature often requires a group of full-time staff to annotate for weeks, which can cause unwanted delays for the project. Thus, for camera-trapping to be effective and scalable, this step must be condensed whilst still maintaining the accuracy of the data.
Because cataloging animal images is merely about repetition and pattern recognition, it is very amenable to automation using Machine Learning algorithms. Recently, with the rise of Machine Learning's broad applications in interdisciplinary fields, there have been more Machine Learning models that are designed with this purpose in mind. Among them, the MegaDetector by Microsoft stands out as the most powerful, accurate, continually developed, and supported model available.
With the high volume of images, there are also complications regarding data management. Besides ensuring data security, raw data sharing, and proper backups, two other important aspects of data management that are often neglected in the context of global wildlife conservation are its interactive capability and openness, i.e. anyone who is not involved in the project.
Commonly, all images are centralized in an online cloud storage such as OneDrive, which serves as an online backup, in addition to a physical backup. However, due to the lack of an appropriate interactive user interface, the data on the drive is simply static images. Furthermore, images and analysis data are not centralized but distributed across the organization's drive, which makes it harder to navigate around a specific project. Thus, it is not easy for external parties to have a quick glimpse of the images and statistical inferences and interact with the dataset if they do not sync the cloud onto their local computer. As a result, new staff or external collaborators are hindered from the static data to instantly acquire a comprehensive understanding of finished projects. In this fashion, to some extent, it hinders collaboration.