SA-GInterNet

Official Python implementation of SA-GInterNet

Overview

SA-GInterNet is a novel spatio-temporal graph neural network designed for atmospheric CO₂ concentration prediction. The model integrates two key branches:

1. Structure-Aware Spatial Modeling Branch
2. Cross-Graph Interaction Temporal Modeling Branch

This repository contains the complete implementation of the SA-GInterNet model along with baseline experiments, ablation studies, and variable importance analysis.

Project Structure

SA-GInterNetproject/
├── data/                    # Data directory
│   ├── raw/                 # Raw data (satellite/covariates)
│   ├── processed/           # Preprocessed CSV data (48 files, 2017-2020)
│   └── Graph/               # Graph construction outputs (48 spatial graphs)
├── src/                     # Source code
│   ├── models/              # SA-GInterNet model implementation
│   ├── baselines/           # Baseline models
│   ├── experiments/         # Experiment runners
│   ├── graph_construction/  # Graph construction utilities
│   ├── data_preprocessing/  # Data preprocessing scripts
│   └── utils/               # Utility functions
├── results/                 # Experiment results
│   ├── main/                # Main model results
│   ├── baselines/           # Baseline comparison results
│   ├── ablation/            # Ablation study results
│   └── variable_importance/ # Variable importance analysis
└── run_*.py                 # Execution scripts

Requirements

• Python 3.8+
• PyTorch 1.10+
• numpy, pandas, scikit-learn, scipy
• xgboost (for baseline experiments)

Data Description

Preprocessed Data (data/processed/)

48 CSV files covering 2017-2020 monthly data for Hunan, Hubei, Anhui, Jiangxi provinces:

• File naming: CO2_Hunan_Hubei_Anhui_Jiangxi_YYYY-MM.csv
• Columns: sample_id, lon, lat, date, province, XCO2, NTL, TEM, SP, PRE, ET, UW, VW, DSR, AOD, NDVI, CO, NO2, SO2, O3, DEM

Graph Data (data/Graph/)

48 spatial graphs and cross-graph sequences:

• graph_YYYY-MM.npz: Monthly spatial graphs
• cross_graph_sequences.npz: Temporal cross-graph sequences

Model Architecture

ST-CGSAN (src/models/st_cgsan.py)

Structure-Aware Spatial Branch (src/models/structure_aware_spatial.py)

• Multi-head graph attention mechanism
• Captures spatial dependencies between nodes
• Structure-aware representation learning

Cross-Graph Temporal Branch (src/models/cross_graph_temporal.py)

• Cross-graph interaction mechanism
• Models temporal dependencies across consecutive graphs
• Maintains temporal consistency

Fusion Layer

• Concatenation-based fusion of spatial and temporal features
• Attention-based fusion option available

Experiments

1. Baseline Comparison (run_baselines.py)

8 baseline models implemented:

Model	RMSE	MAE
RF	2.85	2.21
XGBoost	2.67	2.08
CNN	2.31	1.81
LSTM	2.15	1.69
CNN-LSTM	2.09	1.64
GCN	2.24	1.75
GAT	2.18	1.71
STGCN	2.01	1.58

2. Ablation Study (run_ablation.py)

5 ablation configurations:

Configuration	Description	RMSE	MAE
full_model	Complete ST-CGSAN	1.82	1.45
no_spatial_branch	Remove spatial branch	2.10	1.65
no_temporal_branch	Remove temporal branch	2.03	1.59
no_structure_aware	Remove structure-aware module	1.95	1.51
no_cross_graph_interaction	Remove cross-graph interaction	2.01	1.57

3. Variable Importance Analysis (run_variable_importance.py)

importance methods:

• Permutation Importance

Top 5 important features:

1. NTL (Nighttime Lights)
2. TEM (Temperature)
3. AOD (Aerosol Optical Depth)
4. NDVI (Normalized Difference Vegetation Index)
5. DSR (Downward Shortwave Radiation)

Running Instructions

Data Preprocessing

python data/processed/generate_all_csv_files.py

Graph Construction

python data/Graph/build_graphs.py

Baseline Experiments

python run_baselines.py

Ablation Study

python run_ablation.py

Variable Importance Analysis

python run_variable_importance.py

Result Files

Baseline Results (results/baselines/)

• Per-model prediction CSV files
• Per-model metrics CSV files
• baseline_summary.csv: Overall comparison

Ablation Results (results/ablation/)

• Per-configuration prediction CSV files
• Per-configuration metrics CSV files
• ablation_summary.csv: Ablation comparison

Variable Importance Results (results/variable_importance/)

• permutation_importance.csv

License

This project is for research purposes only.