Drones, automatic counting tools and artificial neural networks in wildlife population censusing
Dominik Marchowski1*
1 Ornithological Station, Museum and Institute of Zoology, Polish Academy of Sciences. Nadwiślańska 108, 80-680 Gdańsk, Poland.
* Corresponding author.
E-mail address: dominikm@miiz.waw.pl (D. Marchowski) ul. Nadwiślańska 108, 80-680 Gdańsk, Poland, https://orcid.org/0000-0001-7508-9466
Key words: Unmanned aerial vehicle, Waterbirds, Machine learning, Bird monitoring, Microbiological tools in ecological research, Birds’ reactions to drones
Abstract
  1. The use of a drone to count the flock sizes of 33 species of waterbirds during the breeding and non-breeding periods was investigated.
  2. In 96% of 343 cases, drone counting was successful. 18.8% of non-breeding birds and 3.6% of breeding birds exhibited adverse reactions: in the former, the birds were flushed, whereas the latter attempted to attack the drone.
  3. The automatic counting birds was best done with the microbiology software - ImageJ / Fiji: the average bird counting rate was 100 birds in 82 seconds.
  4. Machine learning using neural network algorithms proved to be an effective and fast way of counting birds – 100 birds in 23 seconds. However, as the preparation of images and machine learning time are time-consuming, this method is recommended only for large data sets and large bird assemblages.
  5. The responsible study of wildlife using a drone should only be carried out by persons experienced in the biology and behaviour of the animals concerned.
Introduction
Biological diversity conservation has become, next to dealing with the consequences of climate change, one of the most important challenges for humanity (Díaz et al. 2006). The total loss of some species and the rapid decline of others has taken on a so far unknown dynamic (Hooper et al. 2012). Some species will probably not be described at all before they become extinct (Costello et al. 2013). Therefore, we should make every effort to use new technological solutions in ecological research in such a way that the knowledge gained with their use can be effectively used in nature conservation (Arts et al. 2015). In a rapidly changing world, we need ecological research methods that are fast, effective and minimally invasive. In consequence, we will be able to dynamically react to negative changes in the environment (Díaz-Delago et al. 2017).
Large vertebrates, especially birds, have been considered indicators of the state of the environment (Amat & Green 2010). Many long-term bird monitoring programmes have been established in many places around the world (e.g. Farina et al. 2011, Reif 2013, Niemi et al. 2016). New initiatives are constantly emerging, and as a result of this ever denser network of research programmes, we are acquiring an increasingly precise model of ecological processes on Earth (Gregory & Strien 2010). To meet this challenge, we need new, more effective methods and tools.
With regard to modern techniques of gathering ecological data, we are starting to face analytical issues (Shin & Choi 2015) comparable to those in other fields, such as microbiology or biochemistry. Using the principle of similarities of natural structures, such as the similarities of a river network to a blood vessel network (e.g. LaBarbera & Rosso 1989; Neagu & Bejan 1999), I found that waterbird colonies could be construed as bacterial colonies. Since software is commonly used for counting microorganisms, it has been tested many times and its precision confirmed (Barbedo 2012). In the case of aerial photos, it seems justified to use analytical methods previously reserved for areas such as microbiology, such as the ImageJ open software platform used for automatic object counting (Schindelin et al. 2012) or Passing Bablok regression, used to compare methods in clinical laboratory work (Bilić-Zulle 2011).
The use of Unmanned Aerial Vehicles (hereafter drones) in ecological research has already been described in research on breeding (e.g. Chabot et al. 2015; Ratcliffe et al. 2015) and non-breeding birds (e.g. Hodgson et al. 2018; Jarrett et al. 2020), as well as marine (e.g. Koski et al. 2015; Adame et al. 2017) and terrestrial mammals (Vermeulen et al. 2013; Hu et al. 2020). This has turned out to be an effective method for studying larger vertebrates, mainly birds and mammals, but also reptiles (Elsey & Trosclair 2016), as well as for other ecological studies (Puttock et al. 2015; Michez et al. 2016). Nonetheless, research using drones is still at an early stage, and further studies are needed to establish both methodological standards and the actual effectiveness of working with this tool (Barnas et al. 2020). Apart from efficiency and time saving, an important issue is the invasiveness of this method and the safety of the studied object. Initial research in this area has already been carried out on a limited group of species (e.g. Vas et al. 2015; Jarrett et al. 2020).
Waterbirds are a special group of animals, as they often make use of hard-to-reach habitats, such as islands in water bodies or wetlands. Being bioindicators of environmental quality, they are also a frequent object of monitoring studies (Amat & Green 2010; Amano et al. 2017). Therefore, the use of a drone for research on this group of animals should be doubly beneficial: 1) easy and quick access to hard-to-reach areas, and 2) limited disturbance of birds – there is no need to enter a breeding colony or disturb a flock.
The objective of this research was to assess the extent to which the above assumptions were correct. I focused on the effectiveness of a population censusing method using a drone, its invasiveness, the automated analysis of the data obtained by the drone, and the application of Artificial Intelligence (AI) for interpreting the results. The field study was carried out on colonial breeding waterbirds and gregarious waterbirds forming flocks during the non-breeding period. I selected various species that occupy different habitats: these can be divided into four categories – open water, arable fields and meadows, wetlands and islets.
My research questions were: 1) Will it be possible/safe to count nests / incubating birds / individuals using a drone over a colony or a flock? 2) Will the appearance of a drone over the breeding colony or flock cause the birds to react and, if so, what kind of reaction takes place? 3) Is automatic counting using dedicated software and Machine Learning applicable to bird censusing?
Methods