Automatic monitoring of wildlife is becoming a critical tool in the field of ecology. In particular, Radio‐Frequency IDentification (RFID) is now a widespread technology to assess the phenology, breeding and survival of many species. While RFID produces massive datasets, no established fast and accurate methods are yet available for this type of data processing. Deep learning approaches have been used to overcome similar problems in other scientific fields and hence might hold the potential to overcome these analytical challenges and unlock the full potential of RFID studies.
We present a deep learning workflow, coined “RFIDeep”, to derive ecological features, such as breeding status and outcome, from RFID mark‐recapture data. To demonstrate the performance of RFIDeep with complex datasets, we used a long‐term automatic monitoring of a long‐lived seabird that breeds in densely packed colonies, hence with many daily entries and exits.
To determine individual breeding status and phenology and for each breeding season, we first developed a one‐dimensional convolution neural network (1D‐CNN) architecture. Second, to account for variance in breeding phenology and technical limitations of field data acquisition, we built a new data augmentation step mimicking a shift in breeding dates and missing RFID detections, a common issue with RFIDs. Third, to identify the segments of the breeding activity used during classification, we also included a visualisation tool, which allows users to understand what is usually considered a “black box” step of deep learning. With these three steps, we achieved a high accuracy for all breeding parameters: breeding status accuracy = 96.3%; phenological accuracy = 86.9%; and breeding success accuracy = 97.3%.
RFIDeep has unfolded the potential of artificial intelligence for tracking changes in animal populations, multiplying the benefit of automated mark‐recapture monitoring of undisturbed wildlife populations. RFIDeep is an open source code to facilitate the use, adaptation, or enhancement of RFID data in a wide variety of species. In addition to a tremendous time saving for analysing these large datasets, our study shows the capacities of CNN models to autonomously detect ecologically meaningful patterns in data through visualisation techniques, which are seldom used in ecology.
