Bridging sparse observations and global wetland methane emissions

We reconstruct global, 0.5°×0.5°, monthly wetland CH₄ emission fields from sparse eddy-covariance (EC) observations using STFNet, a spatiotemporal neural network combining a convolutional encoder, Fourier Neural Operator (FNO) down/up-sampling, multi-scale CBAM attention, and a ConvGRU temporal core. The model is first pre-trained to reconstruct the full field from sparsely sampled inputs, then fine-tuned on the grid cells of 27 FLUXNET-CH4 sites (aggregated to 20 cells, "EC20"). Contents 1_Benchmark_data/ — benchmark CH₄ field, masks, and EC tower cell coordinates (benchmark_CH4.npz). 2_Pretraining/ — pre-training script and its input tensors (training/validation data and the sampling mask). 3_Pretrained_model/ — pre-trained STFNet weights. 4_Finetune_and_inference/ — EC fine-tuning, inference, and evaluation scripts; the fine-tuned model; the EC20 mask; and the test set. 5_Active_learning/ — iterative active-learning code for optimizing the sampling sites. README.md — full description of files, data formats, reproduction steps, and instructions for adapting the model (adding covariates, using real EC observations). All scripts run locally on a single GPU (Python 3, PyTorch/CUDA, NumPy, scikit-learn, matplotlib, tqdm). The provided pre-trained and fine-tuned weights let users skip training and reproduce the reported test-set result directly: running run_inference.py followed by evaluate.py yields an aggregated test-set R² ≈ 0.946. Environment. The reported results were produced with Python 3.8.3, PyTorch 2.0.1, CUDA 11.8, and cuDNN 8.7, on a single NVIDIA L40S GPU (44 GB VRAM). A large-memory GPU is recommended, as the model operates on full global 0.5° (360×720) fields over 6-month sequences. The code also runs on newer PyTorch versions. Please see README.md inside the archive for full details.