Elephant entry in human settlements causes a major threat to human–elephant conflict (HEC), but the rapid detection thereof while crossing boundary remains a difficult role, due to the lack of necessary boundary sensing infrastructure. A combination of wireless sensor networks and recent advances in deep learning has paved an effective way for elephant detection. This paper presents and evaluates a Boundary Sense Deep Learning architecture (BSDL) for automated detection of elephants, integrating, elephant image feature learning and elephant classification. A novelty in this approach enhances the deep learning Boundary Sense (BS) architecture to include a background subtraction module which highlights the visual region important for elephant recognition. The objective of the proposed BSDL architecture is validated by achieving better accuracy when compared with the raw image, time efficiency, simplifying and boosting up of Deep Convolution Neural Network (DCNN) by introducing preliminaries; elephant recognition with Multi view dataset dramatically improves the state of art and fine tuning using back-propagation. The dataset includes 1,28,069 images of various postures of elephants collected under different circumstances from different regions (Hosur, Bannerghatta, Shimoga (Sakrebailu) and Dubare). The trained model is designed for identifying only one target species, the elephants. The conclusions drawn from our work prove that the achievement percentage is 97% accuracy in the BSDL architecture, when compared with the existing deep learning algorithms using raw image. © 2018, © 2018 Informa UK Limited, trading as Taylor & Francis Group.