Fusion Model of MLP, XGBoost and QAT- Optimized PointNet++ for Predicting Urban Tree Biomass Change

다국어 초록 (Multilingual Abstract)

This thesis presents a novel fusion-based approach for predicting urban tree biomass change using 3D point cloud data through the integration of Quantization-Aware Training (QAT) optimized PointNet++ models with independent regression techniques. The research addresses the critical challenge of developing accurate yet computationally efficient biomass estimation models suitable for deployment on resource-constrained edge devices, enabling real-time urban forest monitoring. The proposed framework comprises three main components: (1) QAT-optimized PointNet++ models for tree species classification (34 species) and trunk-crown part segmentation, (2) frozen encoder feature extraction from pre-trained models, and (3) a dual-head regression architecture combining MLP for feature fusion and XGBoost for biomass change prediction. The system processes 1,024-point 3D point clouds captured via airborne LiDAR, integrating geometric measurements, environmental sensor data, and deep learning-derived features. Experimental evaluation on 2,694 tree samples collected from five geographically diverse regions in South Korea (Daegu, Wonju, Daejeon, Sejong, and Jeju) demonstrates exceptional performance. The species classification model achieves 94.14% accuracy, while the part segmentation model attains 81.82% accuracy. The fusion regression model achieves R² = 0.9663 with RMSE = 0.4437 on the test dataset. Quantization-Aware Training successfully reduces model size by 4× (from 21 MB to 2 MB for segmentation, also 21.1 MB to 2 MB for classification) while maintaining competitive accuracy. INT8 quantized models achieve 3.2× inference speedup (27.4 ms per tree on CPU), making them suitable for edge deployment. Feature importance analysis reveals that species classification features (61.55 importance score) and crown density (50.02 score) are the most influential predictors, confirming the value of species-aware modeling. The research advances the intersection of deep learning, model compression, and environmental monitoring by demonstrating that multi-task learning combined with quantization techniques can achieve both high accuracy and practical deployment feasibility. This work provides a foundation for next-generation automated forest inventory systems supporting urban forestry management, carbon sequestration monitoring, and climate change mitigation strategies.

국문 초록 (Abstract)

이 논문은 3D 포인트 클라우드 데이터를 사용하여 도시 나무 바이오매스 변화를 예측하는 새로운 융합 기반 접근 방식을 제시합니다. 이 연구는 리소스 제약이 있는 엣지 장치에서 배포가 가능한 정확하면서도 계산 효율적인 바이오매스 추정 모델을 개발하는 중요한 문제를 다룹니다. 제안된 프레임워크는 세 가지 주요 구성 요소로 이루어져 있습니다: (1) 나무 종류 분류(34종) 및 줄기-수관 부분 분할을 위한 QAT 최적화된 PointNet++ 모델, (2) 사전 훈련된 모델에서 동결된 인코더 특징 추출, (3) MLP를 이용한 특징 융합과 XGBoost를 통한 바이오매스 변화 예측을 결합한 듀얼 헤드 회귀 아키텍처. 이 시스템은 공중 LiDAR로 캡처된 1,024 포인트 3D 포인트 클라우드를 처리하며, 기하학적 측정값, 환경 센서 데이터 및 딥 러닝에서 파생된 특징들을 통합합니다. 대한민국의 다섯 개 지역(대구, 원주, 대전, 세종, 제주)에서 수집된 2,694개의 나무 샘플에 대한 실험적 평가 결과 뛰어난 성능을 보여주었습니다. 종 분류 모델은 94.14%의 정확도를 달성했으며, 부분 분할 모델은 81.82%의 정확도를 보였습니다. 융합 회귀 모델은 테스트 데이터셋에서 R² = 0.9663, RMSE = 0.4437을 달성했습니다. QAT는 모델 크기를 4배 줄이는 데 성공했으며(분할 21MB에서 2MB로, 분류도 21.1MB에서 2MB로 감소), 정확도를 유지하면서 엣지 배포에 적합한 성능을 제공합니다. INT8 양자화된 모델은 3.2배의 추론 속도 향상(CPU에서 나무당 27.4ms)을 보였으며, 이는 엣지 배포에 적합합니다. 특징 중요도 분석 결과, 종 분류 특징(61.55 중요도 점수)과 수관 밀도(50.02 점수)가 가장 중요한 예측 변수임을 확인했습니다. 이 연구는 딥 러닝, 모델 압축 및 환경 모니터링의 교차점을 발전시키며, 다중 작업 학습과 양자화 기술을 결합하여 높은 정확도와 실용적인 배포 가능성을 동시에 달성할 수 있음을 입증합니다. 이 작업은 도시 숲 관리, 탄소 격리 모니터링 및 기후 변화 완화 전략을 지원하는 차세대 자동화된 산림 목록 시스템을 위한 기초를 제공합니다.

목차 (Table of Contents)

• CHAPTER 1. Introduction １
• CHAPTER 2. Related Works ７
• 2.1 Research background ８
• 2.1.1 The Progression of PointNet and PointNet++ Architectures ９
• 2.1.2 Recent Advances in Point Cloud Processing １０
• 2.1.3 Feature SElection and Preproessing in Point Cloud Moels １1
• 2.1.4 LiDAR-Specific Adaptations １２
• 2.1.5 Applying PointNet to Complex Geometries Found in Forests １３
• 2.2 Model Optimization and Quantization １４
• 2.2.1 Model Optimization for Inference Efficiency １４
• 2.2.2 Studies That Use Quantization for PointNet++ １５
• 2.2.3 Point Cloud Registration and Detection Techniques １６
• 2.3 Biomass Change Estimation １７
• 2.3.1 Applications in Environmental and Ecological Monitoring １７
• 2.3.2 Biomass Estimation for Ecological Monitoring １８
• 2.3.3 Integration of UAV and LiDAR Technologies １８
• 2.4 Model Fusion and Multi-Task Learning ２０
• 2.4.1 Multi-Task Learning Fundamentals ２０
• 2.4.2 Feature-Level Fusion Strategies ２１
• 2.4.3 Decision-Level Fusion and Ensemble Methods ２１
• 2.4.4 Fusion Architectures for Point Cloud Analysis ２２
• CHAPTER 3. Methodology ２４
• 3.1 PointNet++ Architecture Design ２４
• 3.1.1 Classification Model Architecture ２４
• 3.1.2 Part-Segmentation Model Architecture ２７
• 3.2 Quantization-Aware Training (QAT) ３０
• 3.2.1 QAT Methodology and INT8 Configuration ３０
• 3.2.2 QAT Training for PointNet++ part-segmentation Model ３３
• 3.2.3 QAT Training for PointNet++ Classification Model ３９
• 3.3 Fusion Model for Biomass Change Prediction ４３
• 3.3.1 Frozen Encoder Feature Extraction ４５
• 3.3.2 Multi-Modal Feature Fusion Strategy ４７
• CHAPTER 4. Experimental Results and Discussions ５７
• 4.1 Experimental Configuration and Dataset Analysis ５７
• 4.1.1 Dataset Overview and Collection ５７
• 4.1.2 Data Preprocessing and Augmentation ６６
• 4.1.3 Hardware and Software Environments ７０
• 4.1.4 Evaluation Metrics Definition ７２
• 4.2 Point Cloud Understanding Performance ７５
• 4.2.1 Part-Segmentation Results (FP32 Baseline) ７６
• 4.2.2 Part-Segmentation Results (QAT INT8) ７８
• 4.2.3 Species Classification Results (FP32 Baseline) ８１
• 4.2.4 Species Classification Results (QAT INT8) ８３
• 4.3 Quantization Efficiency and Deployment Analysis ８５
• 4.3.1 Model Compression and Size Reduction ８５
• 4.3.2 Inference Speed and Latency Improvements ８７
• 4.3.3 Accuracy vs Efficiency Trade-off Analysis ８８
• 4.4 Biomass Change Prediction and Fusion Model Analysis ９０
• 4.4.1 MLP Model Performance ９０
• 4.4.2 XGBoost Regression Head Performance ９１
• 4.4.3 Comparative Model Performance ９３
• CHAPTER 5. Discussion and Limitations ９７
• 5.1 Discussion ９７
• 5.2 Limitations ９８
• CHAPTER 6. Conclusion and Future work １００
• 6.1 Conclusion １００
• 6.2 Future work １０１
• References １０２
• 국문초록 １１２
• Acknowledgements １１４