NAIRA a tool to automatic mammals genera identification in Camera Trapping Pictures

Camera traps are an efficient tool for detecting terrestrial mammals and birds. Vast amounts of presence data and information about distribution and size of mammals and birds have been generated in a very short time using this approach. The first step into a monitoring study, using camera trapping is to analyse the photographs. However, processing times, when it is done manually, can take longer because of the large amount of data collected. Pre-selecting relevant pictures using an automatic system and automatically identify the animals is an alternative to reduce the analysis time. The challenges to automatic identify the mammal´s genus from camera trap photos are: few examples of some genera (unbalanced classes problem), variation in light levels, constant changes in the scene, animal partially occluded, blurred photographs and other variations resulting from the natural dynamics of the ecosystem. Until now there has not been a computational tool to help in the specific task of labelling the animal´s genus. Thus, it is necessary to design new tools that automate the processing of these photographs. Here we introduce a new version of the software NAIRA. This software uses Machine Learning algorithms as an alternative to automatically labelling mammal genera. The photographs are classified into pictures with animals and without animals using a fuzzy classifier, after that the images are segmented to extract the area with animal and a second classification distinguishes between birds and mammals using an Artificial Neural Network (ANN). Finally NAIRA identifies the genera in the detected mammals using Support Vector Machine (SVM) and Bag of Words (BoW). This version of software includes a proposal to identify the level of incertitude in the machine decision. The results over a database (Andean, Caribbean and Pacific regions in Colombia) with 70780 photos and strong unbalance between classes are promising; it was possible to automatically identify photos of animals and to differentiate among birds and 20 mammal genera (average accuracy 95%). The incertitude analyse is usefully to prevent a misclassification when the algorithm does not have a high level of certitude. In this case, only the 4,45% of pictures (with high classification incertitude level) has to be analysed by an human expert. The attendees will know in detail the functions of the software to recognize the advantages of the processing of this type of images with NAIRA.