Artificial Intelligence (AI)
The use of AI in combination with visual data for the (semi)-automatic classification of camera trap pictures is a technology gaining rapidly in interest. AI-Classification uses neuronal networks, i.e. mathematical algorithms which assigns an input-object to an output. For example, if an input-object is a picture, the output would be the classification for that picture. The optimisation of this algorithm for a specific task is called calibration and requires already labelled data. Labelled data corresponds to the input-objects ultimately used as the definitive network to make a prediction. Therefore in the context of the camera trap monitoring, to have a large collection of labelled reference pictures. During the calibration, the algorithms are refreshed iteratively to minimise the loss function. The loss function calculates the difference between the predicted output of the network (the most likely species according to the network) and the training label (the actual species on the picture). The process is repeated iteratively, for either a predefined duration, or until the desired performance (e.g. a limit for the accuracy of the species classification) is achieved based on the validation date.
KORA uses camera traps to systematically monitor large carnivore populations in Switzerland. Camera traps are particularly used in the monitoring of lynx and wildcat, which can be identified individually due to their coat patterns. However, these focus species feature in only a small percentage of the total pictures taken. Various other mammal species are often also captured by the camera traps, but their images are rarely utilised. The rapid technological development of camera traps and their wide application generates huge amounts of images in a relatively short time. This creates big challenges: How can useful information for the conservation or management of all these species be gained from this huge amount of raw data in a rapid manner and with justifiable costs?
A consortium has been formed to tackle this challenge. It consists of a unique group of academic institutions and national competence centres, who are all confronted with the same problem. They bring together an extraordinary large and diverse data set. The consortium also invited corporations that process data from camera traps in Switzerland to join them. The consortium consists of ecologists, biologists and wildlife managers with years of experience in processing camera trap data, as well as experts in machine learning and developers of IT solutions. The proposed solution lies in adapting the newest developments in methods for AI in combination with the unique, domain specific data set of the consortium for the model calibration.
The analysis of data from camera trap monitoring consists of four challenging and time-consuming tasks:
1) Removal of images that show human activities – a legal requirement for the protection of privacy 2) Recognition of empty images from e.g. false triggering 3) Identification of the pictured species 4) Identification of individuals based on their markings, e.g. lynx and wildcats.
Currently, these tasks are performed largely manually, by well-trained human specialists.
This approach will release important personnel resources which currently need to be invested into the above manual tasks and the management of raw data. Even more importantly, the AI will reduce processing times and make camera trap approaches relevant for the real-time monitoring and protection of wild animals