Edge-AI System for Real-Time Predator Identification and Automated Farm

Alerting Using EfficientNet-B0 on Raspberry Pi 5

Abstract—Predation by wild carnivores remains a significant This paper makes the following contributions:
cause of livestock loss in rural farming regions, and continuous
A comparative evaluation of a custom CNN, MobileNetV2,
manual surveillance is impractical at scale. This paper presents a
and EfficientNet-B0 for four-class wildlife classification (Wolf,
vision-based Edge-AI system for real-time predator identification
Hyena, Boar, Other) under identical training conditions;
and automated farm alerting. Three convolutional neural network
architectures — a custom CNN baseline, MobileNetV2, and A complete real-time deployment pipeline on a Raspberry Pi 5
EfficientNet-B0 — were trained via transfer learning on a balanced with an IMX708 camera, using TensorFlow Lite for on-device
four-class dataset (Wolf, Hyena, Boar, Other) and compared under inference;
identical conditions. EfficientNet-B0 achieved the highest test Integration with a mobile alerting application (image
accuracy (95.0%), outperforming MobileNetV2 (91.2%) and the transmission and SMS notification) for remote monitoring; and
custom CNN baseline (62.5%), while remaining lightweight Field validation of the end-to-end system across multiple real-
enough for on-device inference. The selected model was converted world predator detection scenarios.
to TensorFlow Lite and deployed on a Raspberry Pi 5 with an
IMX708 camera module, forming a fully offline detection pipeline. II. RELATED WORK
Upon detection, the system captures an image and transmits a real-
time alert, including SMS notification, to a companion mobile Convolutional Neural Networks (CNNs) have become the standard
application built with MIT App Inventor. Field validation across approach for image-based wildlife recognition since the emergence
multiple detection scenarios confirmed reliable real-time operation, of deep learning [3], [4]. Lightweight architectures designed for
demonstrating that lightweight transfer-learning architectures can mobile and embedded deployment, such as the MobileNet family
deliver practical, low-cost livestock protection without reliance on [5], [6] and the compound-scaled EfficientNet family [7], have
cloud infrastructure. made it practical to run competitive classifiers on single-board
computers rather than requiring cloud-based GPU inference.
Index Terms—Edge AI, Convolutional Neural Networks,
Prior work has explored Edge-AI and IoT frameworks for
EfficientNet-B0, MobileNetV2, Transfer Learning, Raspberry Pi,
protecting crops and livestock from wildlife intrusion, typically
Wildlife Monitoring, Predator Detection, Internet of Things.
combining a lightweight classifier with a notification mechanism
I. INTRODUCTION [8]. Similarly, TinyML-based approaches have been proposed for
acoustic wildlife monitoring on microcontroller-class hardware.
Human–wildlife conflict remains a persistent challenge for farming Compared with these efforts, the present work provides a direct,
communities located near forests and natural reserves, where controlled comparison of three architectures of increasing
predators such as wolves, hyenas, and wild boars pose a direct complexity on the same predator-classification task and hardware
threat to livestock. Traditional mitigation approaches, including target, and integrates the resulting model into a complete alerting
fencing, guard animals, or manual night patrols, are labor-intensive pipeline validated under real operating conditions rather than
and do not scale to large or remote properties. Recent advances in offline evaluation alone.
deep learning and low-cost embedded hardware have made it
feasible to bring automated visual recognition directly to the field, III. PROPOSED SYSTEM
without dependence on cloud connectivity, recurring data costs, or
A. System Architecture
specialized surveillance infrastructure [1], [2].
The proposed system follows a five-stage pipeline: (1) image
Edge Artificial Intelligence (Edge AI) — running inference acquisition via the IMX708 camera module; (2) on-device
directly on resource-constrained hardware rather than in the cloud preprocessing (resizing to 224×224 and normalization); (3)
— is particularly well suited to this problem: it preserves low inference using the deployed TensorFlow Lite model; (4) event
latency, functions in areas with unreliable internet access, and handling, whereby a predator classification above the confidence
avoids the privacy and bandwidth costs of continuous video threshold triggers image capture and local storage; and (5) alert
streaming. However, deploying deep learning models on transmission over the local network to a mobile application, with
embedded platforms requires an explicit trade-off between an optional SMS fallback. (Fig. 1 illustrates the complete pipeline,
classification accuracy and computational cost, which motivates a from camera acquisition through mobile alert delivery.)
systematic comparison of candidate architectures before
deployment. B. Dataset and Preprocessing
A balanced dataset of 4,000 labeled images was assembled across Model Val. Acc. Test Acc. Params
four classes — Wolf, Hyena, Boar, and Other — with 1,000
MobileNetV2 88.67% 91.20% 2.27M
images per class. The "Other" class included humans, domestic
EfficientNet-B0 95.0% 95.0% 5.33M
animals, and background scenes to reduce false positives. The
dataset was split 70/15/15 into training, validation, and test subsets TABLE I. MODEL COMPARISON ON THE TEST SET
(2,800 / 600 / 600 images). All images were resized to 224×224
pixels and normalized according to each backbone's expected input EfficientNet-B0 outperformed MobileNetV2 by 3.8 percentage
range. Training data was augmented with random flipping, points and the custom CNN baseline by 32.5 points, confirming the
rotation, zoom, brightness, contrast, and translation to improve benefit of transfer learning and compound scaling for this task.
generalization. Despite having roughly 2.3× more parameters than MobileNetV2,
EfficientNet-B0 remained practical for on-device inference on the
C. Model Architectures Raspberry Pi 5 after TensorFlow Lite conversion.
Three classifiers were trained and evaluated under matched Field validation was conducted across five real-time scenarios:
conditions: wolf detection, hyena detection, boar detection, image transmission
1) Custom CNN: A convolutional network trained from scratch, to the mobile application, and end-to-end alert notification. All five
used as a baseline to establish the difficulty of the task without scenarios completed successfully, with the deployed system
transfer learning. correctly classifying live camera frames, capturing predator
images, and delivering alerts to the mobile application without
2) MobileNetV2: Initialized with ImageNet-pretrained weights and manual intervention. (Fig. 2 shows example real-time detection
fine-tuned on the target dataset, selected for its inverted residual output for each predator class.)
blocks and low computational footprint (Sandler et al. [6]).
These results indicate that a carefully selected lightweight
3) EfficientNet-B0: Initialized with ImageNet-pretrained weights, architecture, combined with transfer learning, can approach the
with the last 30 layers unfrozen for fine-tuning. EfficientNet-B0's accuracy typically associated with heavier server-side models
compound scaling of depth, width, and resolution [7] was expected while remaining deployable on low-cost embedded hardware
to yield stronger feature extraction at a comparable parameter suitable for unattended, offline field use.
budget.

All models were trained with the Adam optimizer (learning rate V. CONCLUSION AND FUTURE WORK
1×10⁻⁴), categorical cross-entropy loss, batch size 16, dropout rate This paper presented a complete Edge-AI system for real-time
0.4, class weighting, early stopping (patience 7), and learning-rate predator identification and farm alerting, combining a comparative
reduction on plateau. study of three CNN architectures with a fully deployed, field-
D. Hardware and Deployment validated hardware and mobile-alerting pipeline. EfficientNet-B0
was identified as the most effective backbone, achieving 95.0%
The selected model was converted to TensorFlow Lite format and
test accuracy while remaining suitable for deployment on a
deployed on a Raspberry Pi 5 (8 GB RAM, quad-core ARM
Raspberry Pi 5. The complete system, from image acquisition to
Cortex-A76 @ 2.4 GHz) paired with a Sony IMX708 autofocus
mobile notification, was validated under real operating conditions,
camera module (120° field of view, night-vision capable),
demonstrating a practical, low-cost pathway to intelligent livestock
interfaced via MIPI CSI-2. This configuration was chosen for its
protection without reliance on cloud infrastructure.
balance of computational headroom, low cost, low power
consumption, and native Python/TensorFlow compatibility, Future work will focus on expanding the dataset to additional
making it suitable for unattended field deployment. predator species and a broader range of environmental conditions,
improving robustness under low-light and nighttime scenarios, and
E. Mobile Alerting
incorporating a complementary acoustic classification channel to
A companion Android application was developed using MIT App enable multimodal (audio-visual) predator detection and reduce the
Inventor. Once configured with the Raspberry Pi's IP address, the false-negative rate under partial visual occlusion.
application listens for alert events; upon predator detection, it
displays the captured image, the predicted class, and system status, REFERENCES
and can additionally forward an SMS notification to a configured
phone number for cases where immediate attention is required. [1] I. Goodfellow, Y. Bengio, and A. Courville, Deep Learning.
Cambridge, MA, USA: MIT Press, 2016.
IV. EXPERIMENTAL RESULTS [2] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, "Gradient-
based learning applied to document recognition," Proc. IEEE,
Table I summarizes classification performance for the three
vol. 86, no. 11, pp. 2278–2324, 1998.
architectures on the held-out test set.
[3] Y. LeCun, Y. Bengio, and G. Hinton, "Deep learning," Nature,
Model Val. Acc. Test Acc. Params vol. 521, no. 7553, pp. 436–444, 2015.
Custom CNN 65.2% 62.5% 1.71M [4] F. Chollet, Deep Learning with Python, 2nd ed. Shelter Island,
NY, USA: Manning, 2021.
[5] A. G. Howard et al., "MobileNets: Efficient convolutional
neural networks for mobile vision applications,"
arXiv:1704.04861, 2017.
[6] M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and L.-C.
Chen, "MobileNetV2: Inverted residuals and linear
bottlenecks," in Proc. IEEE CVPR, 2018, pp. 4510–4520.
[7] M. Tan and Q. Le, "EfficientNet: Rethinking model scaling for
convolutional neural networks," in Proc. ICML, 2019, pp.
6105–6114.
[8] K. V. Reddy et al., "Edge AI in sustainable farming: Deep
learning-driven IoT framework to safeguard crops from wildlife
threats," IEEE Access, vol. 12, pp. 77707–77723, 2024.
[9] N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R.
Salakhutdinov, "Dropout: A simple way to prevent neural
networks from overfitting," J. Mach. Learn. Res., vol. 15, pp.
1929–1958, 2014.
[10] S. Ioffe and C. Szegedy, "Batch normalization: Accelerating
deep network training by reducing internal covariate shift," in
Proc. ICML, 2015, pp. 448–456.