DEVELOPMENT OF DEEP LEARNING MODELS TO IMPROVE PLANT BIOSECURITY FOR SUSTAINABLE AGRICULTURE

Abstract

Agriculture is the golden thread that fastens all the sustainable development goals globally. Its profound relationship to the global economy, biodiversity, and human his- tory is unquestionable. The increasing environmental concerns have transitioned agricul- ture from conventional to sustainable practices. This transformation prioritizes ecological balance, long-term agriculture productivity, and natural resource conservation. However, plant stress and indiscriminate use of chemicals significantly threaten agricultural pro- ductivity and quality, undermining the pillars of agricultural sustainability. In this con- text, plant biosecurity becomes a crucial element of sustainable agriculture, focusing on monitoring, preventing, and managing pests, diseases, and invasive species that endan- ger crop health. Achieving plant biosecurity begins with identifying plant stress, which requires continuous monitoring of the agricultural landscape. However, traditional tech- niques and manual inspection are time-consuming and require domain expertise, mak- ing automated monitoring solutions crucial for effectively identifying biotic stress and strengthening crop protection and food security. Digitalization, particularly deep learning, has emerged as a powerful tool for data analysis in many areas, including agriculture. Researchers from various disciplines lever- age deep learning for stress monitoring and propose innovative solutions to address plant resilience, sustainability, and biosecurity issues. However, they face challenges in de- ploying proposed solutions in real-world settings. To address this, a systematic literature review was conducted to identify key research gaps. The identified challenges include the lack of available datasets, an over-reliance on supervised learning, high costs associated with data labelling, neglect of computational efficiency metrics, limited generalizability of models, and regional disparities in research output. iv This work also comprehensively evaluates the strengths, weaknesses, opportunities, and threats of employing deep learning in the field of monitoring plant biotic stress. By examining internal and external factors influencing technology development and imple- mentation, the analysis highlights advantages that can drive progress while addressing challenges that may hinder adoption. Ultimately, this evaluation offers a balanced per- spective on the potential impacts of deep learning applications on the future of plant biosecurity, considering both opportunities and risks. Considering the challenges identified through literature review and motivated by the Digital Agriculture Mission, the authors propose a new framework using semi-supervised and ensemble learning. This framework utilizes unlabelled data, reduces annotation costs and efforts, and enhances classification and detection models for monitoring plant disease. The proposed framework was rigorously validated with benchmark datasets, a crucial pro- cess as it provides reassurance of the framework’s effectiveness and potential for practical application. The testing process, which demonstrated significant performance improve- ments in classifying plant diseases and outperforming existing methods, ensures that the proposed framework is reliable and effective. Additionally, this study explores the potential of integrating sustainable computing with deep learning to maintain the ecological facet and balance the three pillars of sus- tainable agriculture practices: social, economic and environmental. Consequently, the Comprehensive Sustainable Smart Agriculture Framework is introduced to address the often-neglected environmental aspect of agriculture sustainability. This framework incor- porates two crucial facets of sustainable computing: software and deployment optimiza- tion, aimed at improving model efficiency to reduce energy consumption and computa- tional demands. To validate the Comprehensive Comprehensive Sustainable Smart Agri- culture Framework, we propose and test a novel model, Sustainable Smart Agriculture Model , specifically designed for plant disease classification in Indian crops. The Sus- tainable Smart Agriculture Model surpasses existing state-of-the-art models, showcasing outstanding performance while requiring fewer resources. v This research further advances plant biosecurity by exploring the feasibility of popu- lar deep learning object detection models for accurately locating weeds in Indian cotton farms. This approach addresses a major challenge encountered by cotton farmers in India, who often struggle with the effective management of weeds. By providing accurate and timely identification of weed species, the proposed model empowers farmers to imple- ment targeted interventions. This thesis presents deep learning models to enhance plant biosecurity for sustainable agriculture, thereby supporting the three pillars of sustainability—social, economic, and environmental—and fostering their synergistic interaction. This comprehensive contribu- tion emphasizes the critical role of integrating advanced technologies to attain long-term sustainable agricultural practices.