Balancing complexity and performance of machine learning models for avian pests sound detection in agricultural environments

Publication Type

Journal Article

Name of Author

• Kazeneza Micheline, African Centre of Excellence in Internet of Things, University of Rwanda,Kigali, Rwanda
• ANNA SERGEEVNA BOSMAN, Department of Computer Science, University of Pretoria, Pretoria, South Africa
• DAMIEN HANYURWIMFURA, African Centre of Excellence in Internet of Things, University of Rwanda,Kigali, Rwanda
• ANTHONY VODACEK, 3Chester F. Carlson Center for Imaging Science, Rochester Institute of Technology, Rochester, NY 14623, USA 4School of ECE, Carnegie Mellon University Africa, Kigali, Rwanda

Publication Date (Issue Year)

2024

Journal Name

IEEE Xplore

Abstract

Agricultural pest control traditionally relies on inefficient visual inspections. Acoustic monitoring offers a promising alternative by analyzing pest-specific sounds. While effective, implementing acoustic monitoring in agricultural settings faces practical constraints, particularly the limited computational resources available in remote farming environments. This necessitates optimized machine learning (ML) solutions for low-power edge devices. This study evaluates ML models for bird pest detection on resourceconstrained platforms. We evaluated convolutional neural networks (CNNs), recurrent neural networks (RNNs), and traditional ML models by comparing standalone and knowledge-distilled versions of EfficientNetB0 and gated recurrent unity (GRU) against EfficientNetB4, Long short-term memory (LSTM), MobileNetV2, LightGBM, and support vector machine (SVM). Analysis revealed significant performance variations across computational requirements. LightGBM achieved 98% accuracy with minimal resources (8,500 parameters, 7KB, 0.6ms inference), demonstrating good efficiency. SVM (97% accuracy) and distilled GRU (86% accuracy) also showed favorable performance-to-resource ratios. Knowledge distillation substantially enhanced the accuracy of EfficientNetB0 (from 73% to 98%) and modestly improved GRU (from 84% to 86%). We examined platform compatibility across computing tiers, discovering that while high-performance edge devices (Jetson Nano, Raspberry Pi 4) support all studied models effectively, microcontrollers require specialized approaches. Advanced microcontrollers (such as ESP32-S3 and STM32H7) can accommodate optimized implementations, while highly constrained platforms (such as Arduino Nano) require TinyML techniques. This research contributes (i) an on-farm audio dataset, (ii) comprehensive crossmodel evaluation metrics, and (iii) deployment optimization strategies for acoustic pest detection systems in resource-constrained agricultural environments.

Keywords

Acoustic monitoring, bird pest detection, deep learning, knowledge distillation, machine learning.

Grantee Name(s)

EMMANUEL NDASHIMYE

Project Title

Agricultural Data from Acoustic Monitoring

Type of Grant

Research Award

Thematic Area

ICTs Including Big Data and Artificial Intelligence

Recommended Citation

Micheline, K., BOSMAN, A. S., HANYURWIMFURA, D., & VODACEK, A. (2024). Balancing complexity and performance of machine learning models for avian pests sound detection in agricultural environments. IEEE Xplore https://doi.org/10.1109/ACCESS.2024.0429000