README.md

metadata

language: en
license: mit
tags:
  - precipitation
  - convlstm
  - multitask-learning
  - climate
  - vegetation
  - amazon
model-index:
  - name: MultiTask ConvLSTM w/veg inputs
    results:
      - task:
          type: time-series-forecasting
          name: Precipitation Prediction
        dataset:
          name: ERA5-Land Amazon Basin (2021–2023)
          type: reanalysis
        metrics:
          - type: mean_squared_error
            value: 0.28
          - type: spearman_correlation
            value: 0.87
          - type: pearson_correlation
            value: 0.79
          - type: kendall_tau
            value: 0.7
          - type: nash_sutcliffe_efficiency
            value: 0.62
          - type: f1
            value: 0.82
          - type: accuracy
            value: 0.9
          - type: precision
            value: 0.9
          - type: ROC-AUC
            value: 0.97
          - type: recall
            value: 0.75

MultiTask ConvLSTM for Precipitation Prediction

This repository contains two MultiTask ConvLSTM models:

• veg/: Model trained with vegetation input variables
• noveg/: Model trained without vegetation input variables

Both directories include:

• convlstm.py: base ConvLSTM layers
• model.py: MultiTask ConvLSTM model definition
• example_inference.py: inference script
• data/: example .pth files (test)

These scripts are provided for reproducibility of the model architecture and workflow. Exact runtime and performance may vary depending on hardware.

Example Data

We provide a large test .pth files so you can immediately run the inference script without preprocessing. These files are already preprocessed and normalized from the ECWMF REA5 reanalysis data.

Each .pth file loads as a list of batches:

• X_batch: shape (B, T_in, C_in, H*W)
• y_batch: shape (B, T_out, C_out, H*W)
• y_zero_batch: shape (B, T_out, C_out, H*W)

with H=81, W=97. Inside evaluate(...), these are reshaped to (B, T, C, H, W).

---

How to Use

Ensure all files are in the correct directory then run the example_inference.py file.

1 Get the repo

git clone https://huggingface.co//MultiTaskConvLSTM cd MultiTaskConvLSTM

2 Install minimal deps

pip install -r requirements.txt

3 Run inference (choose one variant)

python veg/example_inference.py

or

python noveg/example_inference.py

Citation If you use this model, please cite: > Lilly Horvath-Makkos (2025). [title] [journal] BibTeX:

bibtex @article{horvathmakkos2025, title={Title}, author={Horvath-Makkos, Lilly}, journal={Journal}, year={2025} }