Line of action:

Status:

Execution date:

The project, according to the organization, arises from the need for an effective biodiversity census and monitoring system to have accurate and updated information on the size, distribution and evolution of animal populations, and to improve the capacity for their conservation. The University of Huelva points out that photo-trapping cameras can obtain this information, but although they are capable of taking huge amounts of images, there is no automatic method for extracting knowledge from these images. According to them, the identification of the species photographed and the digitization of the data is normally carried out by technicians and is so laborious that most of the knowledge is not exploited. The entity proposes this project with the purpose of implementing a wildlife census and monitoring system by automatic identification of the species in the photo-trapping images, which, in this case focuses on mammalsbut that can be extrapolated and used for other species and on a national scale.

The main objective of the project has been to implement an automatic biodiversity census and monitoring system through the identification of species in photo-trapping images using deep learning(DL) algorithms.

The specific objectives are:

• Build and manage a database of photo-trapping images and relevant image information.
• Acquire photo-trapping images for training DL algorithms to identify species of interest in the images.
• Create a DL-based artificial intelligence system for species recognition in photo-trapping images.
• Collect photo-trapping images for census and monitoring of species of interest.
• Build and manage a system for verifying species in images with a low level of confidence in their classification.
• Analyze the information contained in the database and generate technical reports.
• Disseminate and publicize the results of the project.

• Design of the database structure that allows the organization of the photo-trapping images and their orderly storage according to their capture date.
• Implementation of the database and the generated files.
• Creation of a system for querying and updating the database.
• Maintenance of the database with the revisions and actions performed.
• Compilation of photo trapping images available from other projects. In this activity, different agreements have been established with several researchers and individuals who have used photo-trapping images as a field technique in their projects, in order to obtain permission to cede their images.
• Collection of photo trapping images available on the web that are tagged for non-commercial use. Based on the previous activity, non-commercial images collected from different sources have been registered and classified.
• Design and training of a convolutional neural network CNN (Convolutional Neuronal Network) for species recognition in photo-trapping images.
• Creation of a program for the recognition of species in new images by means of the CNN and to update the database.
• Improvement of the CNN by retraining it with the new images incorporated in the database. The application of artificial intelligence (AI) has been carried out. Thus, the accuracy of CNN for taxon and group recognition has been improved.
• Selection of 12 census species. Finally, all species of wild and domestic mammals present in the Doñana Natural Area have been censused, improving the initial approved proposal.
• Study of the location and selection of sites for photo-trapping cameras.
• Installation of photo-trapping cameras, maintenance and downloading of captured images.
• Classification of images using the artificial intelligence system developed.
• Creation of a citizen science project on the org platform, in order to request the collaboration of volunteers in the verification of the species present in the images classified by artificial intelligence (CNN) under the name “Iberian Camera Trap Project“.
• Management of the verification project “Iberian Camera Trap Project”. This action consisted of answering questions and doubts generated by volunteers in the platform’s forums, as well as technical incidents that arose during the course of the project.
• Analysis of demographic parameters for 5 taxa of interest in Doñana National Park. Among them, 3 correspond to wild mammals (deer, wild boars and foxes) and 2 to domestic mammals (horses and cows).
• Dissemination and communication of the project to raise awareness of the operation and applications of the census system implemented throughout the initiative.

The objective of the AI-CENSUS project has been to implement an automatic biodiversity census and monitoring system. In order to improve the management and conservation of the species, it is essential to have accurate and quality information on the size, distribution and evolution of their populations.

This information can be obtained through remote monitoring of the species by means of photo-trapping cameras, but, although the cameras are capable of taking huge amounts of images, there is no automatic method for extracting knowledge from them. Normally, the identification of the photographed species and the digitization of the data is carried out manually and is such an arduous task that most of the knowledge cannot be exploited.

Within the framework of the AI-CENSUS project, a wildlife census and monitoring system has been implemented through the automatic identification of species in photo-trapping images using artificial intelligence (AI) algorithms. For this purpose, a convolutional neural network (CNN) was trained using deep learning algorithms (
Deep Learning
) with correctly classified photo-trapping images.

The images were taken in Doñana National Park through the installation of 38 photo-trapping cameras that took 3,673,723 images between October 2020 and December 2022. Some of these photos were uploaded to the citizen science platform
Zooniverse
where more than 15,000 volunteers classified 1,500,000 images. The images classified by the volunteers and reviewed by the researchers were used to train the CNN. In this way, the rest of the images were classified using AI.

The network proved to be highly effective for the classification of medium and large mammal species present in the National Park. The IA achieved sensitivity values between 0.72 – 0.98 when using all classified images and between 0.95 – 0.99 when using only the images in which the system had the highest confidence.

On the other hand, the IA accuracy reached values between 0.47 – 0.98 when all images were used and between 0.89 – 0.99 when only the highest confidence images were used. In addition, the network was able to classify 777,600 images per day, compared to the 1,528 images classified per day by volunteers. Beyond the classification of the images, using hierarchical dynamic occupancy models , it was possible to obtain demographic data on the monitored populations, such as the probability of occupancy, colonization or probability of extinction.

In this regard, 7 taxa or groups of species have been identified through the IA: Leporidae, Cervidae, Cow, Horse, Boar, Fox, Human and “Other species”. In addition, by ranking on CNN 12 different species have been recorded: foxes (Vulpes vulpes), badger (Meles meles), honeydew (Herpestes ichneumon), genet (Genetta genetta), Iberian lynx (Lynx pardinus), wild boar (Sus scrofa), red deer (Cervus elaphus), fallow deer (Dama dama), cow or bull(Bos taurus), Iberian hare(Lepus granatensis), rabbit(Oryctolagus cuniculus) and horse or mare(Equus ferus caballus).

In this context, the results of the project have allowed us to endorse the usefulness of artificial intelligence to provide accurate data on the demography of the populations and to improve their conservation and management. Likewise, through the combination of species monitoring by means of photo-trapping cameras, species identification by means of artificial intelligence and the use of advanced statistical models, constant and updated information on the size, distribution and evolution of animal populations can be obtained.