🌲 Dead Tree Segmentation Project

🎯 Project Overview

This project addresses the critical environmental challenge of monitoring forest health through automated dead tree detection. Using 4-channel aerial imagery (RGB + Near-Infrared), we implement and compare multiple semantic segmentation models to achieve accurate pixel-level classification of dead vegetation.

Key Features

• Multi-channel Input: RGB + NIR (Near-Infrared) 4-channel image processing

• Multiple Architectures: Implementation of UNet, UNet++, SegFormer, DeepLabv3, and DeepLabv3+

• Dual Evaluation Metrics: Both Mean IoU (mIoU) and Dead Tree-specific IoU analysis

• Robust Data Pipeline: Comprehensive preprocessing with stratified cross-validation

• Performance Analysis: Detailed cross-validation results and timing benchmarks

🛠️ Tech Stack

• Languages: Python

• Frameworks: PyTorch, Segmentation Models PyTorch (SMP)

• Computer Vision: OpenCV, Albumentations

• Data Science: NumPy, Pandas, Matplotlib

• Deployment: Streamlit, Google Colab, Hugging Face Spaces

• Tools: Git, Jupyter Notebook

🖼️ Example Results

Input RGB	Ground Truth	UNet++ Prediction

More qualitative comparisons (all 5 models) are available in the [results/]

📊 Model Performance Overview

Mean IoU-focused Models（in training）

| Model | Val mIoU | Val Dice | Val Pixel Acc |

|-------|----------|----------|---------------|

| UNet | 0.714 ± 0.011 | 0.614 ± 0.019 | 0.984 ± 0.002 |

| UNet++ | 0.717 ± 0.018 | 0.618 ± 0.033 | 0.984 ± 0.002 |

| SegFormer | 0.702 ± 0.015 | 0.795 ± 0.023 | 0.981 ± 0.003 |

| DeepLabv3 | 0.663 ± 0.026 | 0.512 ± 0.055 | 0.980 ± 0.002 |

| DeepLabv3+ | 0.697 ± 0.015 | 0.581 ± 0.031 | 0.983 ± 0.001 |

Dead Tree IoU-focused Models

| Model | Dead Tree IoU | Val Dice | Inference Speed |

|-------|---------------|----------|-----------------|

| UNet (Dead Tree Focus) | 0.443 ± 0.033 | 0.618 ± 0.041 | - |

| UNet++ (Deep Supervision) | 0.415 ± 0.002 | 0.581 ± 0.004 | - |

| SegFormer | 0.434 ± 0.030 | 0.596 ± 0.027 | 68.31 FPS |

| DeepLabv3 | 0.349 ± 0.049 | 0.512 ± 0.055 | - |

| DeepLabv3+ | 0.413 ± 0.029 | 0.581 ± 0.031 | - |

🚀 Quick Start

Data Preprocessing

from data_preprocess.four_channel_data_preprocessing import DeadTreeDataPreprocessor

  

# Initialize preprocessor with stratified sampling

preprocessor = DeadTreeDataPreprocessor(

dataset_root="path/to/USA_segmentation",

random_seed=42,

stratification_bins=4,

test_split_ratio=0.2,

k_folds=5

)

  

# Load and validate dataset

success = preprocessor.load_and_preprocess_dataset()

  

# Create stratified cross-validation splits

splits = preprocessor.create_stratified_splits(train_val_ratio=0.8)

  

# Generate comprehensive statistics

report = preprocessor.generate_statistics_report()

print(report)

🔬 Technical Implementation

Model Architectures

1. UNet Family

• UNet: Classic encoder-decoder with skip connections

• UNet++: Enhanced with nested skip connections and deep supervision

• Specialized Focus: Versions optimized for either mIoU or Dead Tree IoU

2. SegFormer

• Architecture: Transformer-based hierarchical encoder with lightweight decoder

• Performance: Best balance of accuracy and inference speed (68.31 FPS)

• Multi-scale: Efficient multi-scale feature extraction

3. DeepLab Family

• DeepLabv3: Atrous convolution with ASPP (Atrous Spatial Pyramid Pooling)

• DeepLabv3+: Enhanced decoder with fine-grained feature recovery

Data Pipeline Features

• 4-channel Processing: RGB + NIR channel alignment and validation

• Stratified Sampling: Dead tree ratio-based stratification for balanced training

• Cross-validation: 5-fold CV with stratified splits

• Augmentation: Synchronized transformations for image-mask pairs

• Quality Control: Comprehensive validation and error handling

📈 Evaluation Metrics

Primary Metrics

• Mean IoU (mIoU): Overall segmentation quality

• Dead Tree IoU: Specific performance on dead tree detection

• Dice Coefficient: Overlap-based similarity measure

• Pixel Accuracy: Overall pixel-level classification accuracy

Timing Analysis (SegFormer)

• Optimal Batch Size: 16

• Maximum FPS: 68.31

• Average Inference Time: 14.64ms per sample

• Hardware: Tesla T4 GPU, 14.7GB Memory

📜License

MIT License © 2025

---

Last Updated: September 2025