STAD-Imputer

STAD-Imputer: A Spatio-Temporal Adaptive Diffusion Framework for Highly Sparse Remote Sensing Data Imputation

Remote-sensing variables are essential for monitoring coastal marine ecosystems, yet frequent cloud contamination often causes extreme data sparsity with missing rates exceeding 90%. STAD-Imputer is a unified conditional diffusion framework that addresses this challenge through three adaptive modules – ATI, ANA, and AHM.

---

1. Architecture

---

2. Installation

# Recommended: create a fresh conda environment
conda create -n stad python=3.9 -y
conda activate stad

# Install dependencies
pip install -r requirements.txt

---

3. Datasets

All data used in this work are publicly available through online sources. The observation datasets were 8-day averaged Level 3 mapped products from Moderate Resolution Imaging Spectroradiometer (MODIS) Aqua projects with a spatial resolution of 4 km. You can select the data with *.8D.*.4km.nc as filter.

We also uploaded the datasets on Zenodo at https://doi.org/10.5281/zenodo.14724760. Then,

mv data.zip /path/to/STAD-Imputer/
mkdir /path/to/STAD-Imputer/data
unzip data.zip -d /path/to/STAD-Imputer/data

Remote-Sensing Datasets

• SST4 (Sea Surface Temperature)
• PAR (Photosynthetically Active Radiation)
• Chl-a (Chlorophyll-a)

---

5. Quick Start

Datasets (SST4 / PAR / Chl-a)

# SST4 (Pearl River Estuary)
python train.py \
    --data_root /path/to/zone_sst4_data \
    --area PRE \
    --datasets_type sst4 \
    --epochs 500 \
    --batch_size 1 \
    --missing_ratio 0.9 \

# PAR
python train.py --data_root /path/to/zone_par_data \
                --area PRE --datasets_type par --epochs 500

# Chlorophyll-a
python train.py --data_root /path/to/zone_chla_data \
                --area PRE --datasets_type chla --epochs 500

---

6. Configuration

Argument	Default	Description
--missing_ratio	0.9	Fraction of observations randomly masked during training
--num_steps	50	Diffusion denoising steps
--num_samples	10	Monte-Carlo samples at inference
--ATI_tcn_layers	2	Number of stacked ATI blocks
--ATI_dilation_choices	1,2,4,8	Dilation rates for local TCN experts
--ANA_k_phys	8	Physical neighbour count
--ANA_k_feat	8	Semantic neighbour count
--ANA_num_prototypes	32	Number of learnable prototype vectors
--AHM_num_experts	8	Number of sparse spatial experts
--AHM_top_k	3	Top-k expert routing
--AHM_num_scales	3	Graph aggregation scales
--balance_weight	0.01	MoE load-balance loss coefficient

---

7. Evaluation Metrics

At each test epoch the following metrics are reported:

• [Real] MAE, RMSE, MAPE (physical units)
• [Real] R², SSIM, CRPS