The DeepFaune initiative: a collaborative effort towards the automatic identification of French fauna in camera-trap images

Abstract

Camera-traps have revolutionized the way ecologists monitor biodiversity and population abundances. Their full potential is however only realized when the hundreds of thousands of images collected can be rapidly classified with minimal human intervention. Machine learning approaches, and in particular deep learning methods, have allowed extraordinary progress towards this end. Trained classification models remain rare however, and for instance are only emerging for the European fauna. This can be explained by the technical expertise they require but also by the limited availability of large datasets of annotated pictures, which are key to obtaining successful recognition models.In this context, we set-up the DeepFaune initiative (https://deepfaune.cnrs.fr), a large-scale collaboration between dozens of partners involved in research, conservation and management of wildlife in France. The aim of DeepFaune is to aggregate individual datasets of annotated pictures to train species classification models based on convolutional neural networks, an established deeplearning approach.Here we report on our first milestone, a two-step pipeline built upon the MegaDetector algorithm for detection (discarding empty pictures and cropping the animal) and a classification model for 18 species or higher-level taxa as well as people and vehicles. The classification model achieved 92% validation accuracy and showed > 90% sensitivity and specificity for many classes. Most importantly, these performances were generally conserved when tested on an independent out-of-sample dataset. In addition, we developed a cross-platform graphical-user-interface that allows running the pipeline on images stored locally on a personal computer.In conclusion, the DeepFaune initiative provides a freely available (for non-commercial purposes) toolbox with high performance to classify the French fauna in camera-trap images.

Automated identification of invasive rabbitfishes in underwater images from the Mediterranean Sea

“…A future priority will be to couple this identification model with a general fish detection model able to segment all fish individuals from a video frame or a picture (e.g., Knausgård et al, 2022) for the model to subsequently process those extracted images. Even if current detection models are only able to identify a single species (Lopez‐Marcano et al, 2021), recent advances in other taxa show encouraging results such as DeepFaune (Rigoudy et al, 2023) for French terrestrial fauna or for insects (Teixeira et al, 2023). However, these detection models require images with all individuals annotated, which is time‐consuming as underwater images could contain >100 fish.…”

Automated visitor and wildlife monitoring with camera traps and machine learning

“…Furthermore, object detection models are often limited to broad categories (person, animal and vehicle), while wildlife studies usually address questions on the species level. In this case, we suggest applying a combined approach of initially using MegaDetector to filter out empty pictures, and for example, pictures of human activities, followed by either manual classification and counting of remaining wildlife images on the species level or using additional models for automated species or activity identification (Redmon et al., 2016; Rigoudy et al., 2022, Fig. 7).…”

A versatile semiautomated image analysis workflow for time-lapsed camera trap image classification

“…1) is adapted from Böhner et al (2022), including pre-processing of images, model training, manual quality checks, and final data formatting. Specifically, we build on the results of Rigoudy et al (2022) andFennell et al (2022) who combined Megadetector with manual classification by combining custom trained models with Megadetector. The workflow consists of the following 3 steps in addition to pre-processing of images and final data formatting (Fig.…”

“…Specifically, it consists of (1) identifying good-quality images, (2) separating empty images from images with animals, humans, or vehicles (3) cropping out animals from images and classifying them by species, and (4) manual inspection of a selection of images. Fennell et al (2022) and Rigoudy et al (2022) used similar approaches combining Megadetector with custom trained species identification, but contrary to their case, in our study limiting false negatives was crucial. We investigate optimal thresholds and procedures for the trade-off between false negatives and manual reviewing time.…”

“…While pre-trained models are often limited to the fauna of specific regions (e.g. Deepfaune; Rigoudy et al, 2022 or Conservation AI), the training of custom models may require millions of images and computer processing power inaccessible to many researchers. A more streamlined, accessible, and versatile solution that utilizes the strengths of computer vision models is needed.…”