From Deep Learning Recognition of Pest Instar to Fuzzy-Guided Explainable Reinforcement Learning for Nutrient Management in Smart Farming : 스마트 농업에서 해충 유충 단계의 딥러닝 인식과 영양 관리를 위한 퍼지 기반 설명 가능한 강화학습 연구

다국어 초록 (Multilingual Abstract)

Ensuring food security for a growing global population requires smarter and more sustainable agricultural practices. This thesis presents an integrated approach that fuses intelligent perception and interpretable decision-making to address two critical pillars of sustainable farming: pest management and nutrient optimization. The first contribution focuses on the early detection of Fall Armyworm (Spodoptera frugiperda) larvae using a robotic vision system inspired by the human visual pathway. A dual-stage deep learning architecture was developed, combining a VGG19-based classifier for peripheral scanning with a Faster-RCNN detector (VGG16 backbone) for foveal analysis. Stereo RGB vision, supported by SIFT-based feature correspondence and ChArUco calibration, enabled precise 3D localization of pest instances. The system was validated in a physics-based robot simulation environment (CoppeliaSim), confirming its capability for real-time pest recognition and targeted neutralization. The second contribution addresses the interpretability of reinforcement learning (RL)-driven fertilizer application strategies. A framework was developed wherein both Fuzzy Inference Systems (FIS) and Neuro-Fuzzy Inference Systems (NFIS) approximate the policies of trained RL agents. Latent features extracted from autoencoders were used to enhance NFIS performance. Experimental results demonstrated that NFIS could match RL policy behavior with high fidelity (RMSE = 0.30), while also improving nitrogen use efficiency compared to the conventional expert strategy. FIS models offered fully transparent rule bases but at the cost of lower policy fidelity, revealing a trade-off between interpretability and expressiveness. Together, these contributions demonstrate how perception and decision-making systems, when fused through deep learning and fuzzy logic, can lead to practical, explainable, and sustainable solutions for integrated pest and nutrient management in smart farming.

국문 초록 (Abstract)

전 세계적으로 증가하는 인구의 식량 안보를 보장하기 위해서는 보다 지능적이고 지속 가능한 농업 관행이 요구된다. 본 연구는 지속 가능한 농업의 핵심 요소인 해충 관리와 영양 최적화를 동시에 해결하기 위해, 지능형 인식과 해석 가능한 의사 결정을 통합한 접근 방식을 제안한다. 첫 번째 기여는 인간의 시각 경로에서 영감을 받은 로봇 비전 시스템을 통해 열대거세미나방(Spodoptera frugiperda) 유충의 조기 탐지를 가능하게 한다. 주변 환경을 스캔하는 VGG19 기반 분류기와 초점 분석을 수행하는 Faster R-CNN 검출기(VGG16 백본)를 결합한 2단계 딥러닝 아키텍처를 설계하였다. SIFT 기반 특징 매칭 및 CharUco 보정을 적용한 스테레오 RGB 비전은 해충 개체의 정밀한 3차원 위치 추정을 가능하게 하였다. 제안된 시스템은 물리 기반 로봇 시뮬레이션 환경(CoppeliaSim)에서 검증되었으며, 실시간 해충 인식 및 표적 중화 기능의 유효성을 입증하였다.

목차 (Table of Contents)

• I. Introduction 1
• 1.1 Overview 1
• 1.2 Research Objectives 4
• 1.3 Structure of the Thesis 5
• II. Previous Works 7
• 2.1 Literature Review 7
• 2.1.1 Machine Learning for Integrated Pest Management (IPM) 7
• 2.1.2 Explainable Reinforcement Learning for Integrated Nu- trient Management (INM) 13
• 2.1.3 Identified Gaps 16
• 2.2 Core Scientific Contributions 17
• 2.2.1 Contribution 1: Precision Pest Instar Detection with Supervised Learning (IPM) 18
• 2.2.2 Contribution 2: Fuzzy-Guided Explainable Reinforce- ment Learning for Fertilizer Policy Optimization 19
• III.Proposed Methods 21
• 3.1 Precision Pest Instar Detection with Supervised Learning (IPM) 21
• 3.1.1 Overview 21
• 3.1.2 Data Collection Location and Image Sensors 23
• 3.1.3 Deep-Learning Architecture for Insect Pest Classifi- cation and Detection 27
• 3.2 Explainable Fertiliser Application with Reinforcement Learn- ing (RL) (INM) 34
• 3.2.1 Overview 34
• 3.2.2 Simulation Environment: gym-DSSAT 36
• 3.2.3 Fuzzy Inference System (FIS) 38
• 3.2.4 Neuro-Fuzzy Inference System (NFIS) 41
• 3.2.5 Autoencoder for Latent Feature Extraction 44
• 3.2.6 Evaluation Metrics 46
• IV. Experimental Results 49
• 4.1 Precision Pest Instar Detection with Supervised Learning 49
• 4.1.1 Experimental Setup 49
• 4.1.2 Insect Classifier and Detector Performance 50
• 4.1.3 Cosimulation Results 53
• 4.2 Explainable Fertiliser Application with Reinforcement Learn- ing (RL) (INM) 55
• 4.2.1 Experimental Setup 55
• 4.2.2 Baseline Policy Benchmarking 59
• 4.2.3 Results: FIS Approximation of Expert Policy 61
• 4.2.4 Results: Autoencoder Performance 65
• 4.2.5 Results: Surrogate Approximation of RL Policy 67
• 4.3 Interpreting NFIS: Stage-Wise Fertilisation Interpretation from NFIS Output 70
• V. Conclusion and Future Research Directions 73
• 5.1 Precision Pest Recognition and Localization 73
• 5.2 Explainable Fertilizer Decision-Making with RL Surrogates . 74
• 5.3 Limitations and Future Work 75
• References 78
• Abstract (In English) 85
• Abstract (In Korean) 87