DeLoCo: Decoupled location context-guided framework for wildlife species classification using camera trap images

The automated classification of wildlife species using camera trap images is of paramount importance for wildlife surveys and biodiversity conservation. Deep learning methods, which are particularly adept at handling large datasets, has demonstrated considerable promise in this field. While the camera trap location context may offer supplementary information for species classification, existing methods frequently fail to adequately incorporate this contextual information. To increase classification accuracy and facilitate ecological information processing, we explore the correlation between location context and species classification tasks, proposing the Decoupled Location Context-guided framework (DeLoCo). DeLoCo incorporates wildlife image backgrounds from camera trap locations to assist in species classification by decoupling co-supervised species and location classification tasks. Moreover, the weighted loss strategy based on correlation strength is proposed to prioritize image samples from locations with fewer classes and minimize the impact of samples from locations with diverse classes. Experiments on two typical camera trap datasets (IWildCam and TerraInc) validates that our approach outperforms eight benchmark methods. This demonstrates the great advantages of utilizing our innovative DeLoCo method for the efficient ecological data processing in camera trap images with location information. The code is available here: https://github.com/rid1cul0us/wildlife . • Explores the correlation between wildlife image backgrounds and species. • Proposes a framework using camera trap backgrounds to aid in species classification. • Evaluates DeLoCo and eight benchmarks on two typical camera trap datasets.