A framework for an early warning system for the management of the spread of locust invasion based on artificial intelligence technologies.
No Thumbnail Available
Date
2024
Authors
Halubanza, Brian
Journal Title
Journal ISSN
Volume Title
Publisher
The University of Zambia
Abstract
As the global population continues to grow, ensuring food security remains a paramount challenge, especially in light of threats like devastating locust invasions. The agricultural sector in Zambia, particularly in the Sikaunzwe area of Kazungula district, Southern Province, faces unique challenges including inaccurate locust species identification, a lack of field staff, and the inaccessibility of infested areas. Despite advancements in AI and sensor applications for pest management, existing approaches often fail to robustly adapt to varied agronomic conditions or to integrate real-time environmental data effectively. Furthermore, these methods generally lack sufficient engagement of local communities, crucial for the sustained success of locust management strategies. This research addresses these problems by introducing a comprehensive framework that enhances early warning and management of locust invasions. The methodologies employed include Focus Group Discussions (FGDs), semi-structured questionnaires, and field experiments using a Deep Learning model embedded in Internet of Things (IoT) devices and Cloud Computing. The research is guided by the Unified Theory of Acceptance and Use of Technology (UTAUT) and the Design Science Research Methodology, facilitating systematic development and evaluation of an AI-based early warning system. The development of mobile applications and SMS services has significantly enhanced the reach and effectiveness of locust management strategies. The research culminated in the creation and implementation of an advanced Convolutional Neural Network (CNN) model, specifically the MobileNet version 2 quantized model, tailored for automatic identification of locust species. This model achieved an average precision rate of 91% for Locusta migratoria and 85% for Nomadacris septemfasciata using a custom dataset of 1700 images from the study area. Beyond AI-driven identification, the research integrated low-cost IoT devices capable of capturing real-time locust images and uploading them to an online database only if they met an 80% accuracy threshold, while also collecting vital environmental data like temperature and humidity. This integration of AI, IoT, and real-time data collection represents a transformative approach to integrated locust pest management, setting a scalable model for future adoption in similar agricultural contexts. The framework not only addresses immediate locust management challenges but also enhances the broader path of technological progress in agriculture.
Description
Thesis of Doctor of Philosophy (PhD) in Computer Science.