README.md

---
license: cc-by-4.0
language:
- en
pretty_name: 200,000 Years of Weather & Quasi-Stationary Forest Dynamics (Beech, Pine, Spruce)
tags:
- forest
- weather
- climate-impacts
- time-series
- ecology
- simulation
- multimodal
- FORMIND
- AWE-GEN
size_categories:
- 10K<n<100K
task_categories:
- time-series-forecasting
- tabular-regression
---

# 200,000 Years of Weather and Quasi-Stationary Tree Dynamics Simulation for Beech, Pine, and Spruce Forests

A large synthetic benchmark coupling a stochastic **weather generator (AWE-GEN)** with a
process-based **forest gap model (FORMIND)** to study how weather time series drive
forest biomass mortality. It is designed as a test bed for machine-learning models that
map weather (and forest structure) time series to an ecological impact — a genuinely
**multi-modal, time-series → regression** setting.

Species covered: **European beech, Scots pine, and Norway spruce.**

This is a **richer, higher-resolution** version of an earlier monthly-averaged
simulation: the weather and forest dynamics here are provided at **daily** resolution
(raw) and as a **5-day / pentad** aggregate (processed), rather than monthly.

## Overview

Weather is generated with the hourly stochastic weather generator **AWE-GEN**; its
aggregated **daily** output (precipitation, temperature, radiation) drives the
individual-based forest gap model **FORMIND**. The dataset provides annual **forest
biomass mortality rates** together with per-year **histograms of five structure
variables** — age, stem volume, leaf area index (LAI), height, and diameter. All data is
stored as **HDF5** files.

> **Temporal resolution.** `raw_simulation/` and `processed/` are **the same
> simulation** at two aggregation levels — the raw FORMIND output is **daily**, and the
> processed splits are its **5-day (pentad) aggregate**, prepared for model training.

## Repository structure

```
forest_mortality/
├── raw_simulation/          # FORMIND output, DAILY resolution (20-member ensemble / species)
│   ├── beech/               #   beech_dynMort_0.h5 … beech_dynMort_19.h5
│   ├── pine/                #   pine_dynMort_0.h5  … pine_dynMort_19.h5
│   └── spruce/              #   spruce_dynMort_0.h5 … spruce_dynMort_19.h5
│
└── processed/               # ML-ready PENTAD (5-day) aggregate, split (beech & pine)
    ├── train_MBR_beech_pentad_3_years_10000ha.h5
    ├── val_MBR_beech_pentad_3_years_10000ha.h5
    ├── test_MBR_beech_pentad_3_years_10000ha.h5
    ├── train_MBR_pine_pentad_3_years_10000ha.h5
    ├── val_MBR_pine_pentad_3_years_10000ha.h5
    ├── test_MBR_pine_pentad_3_years_10000ha.h5
    ├── bins_Xs_train_beech.npy    # structure-variable histogram bin edges
    └── bins_Xs_train_pine.npy
```

### `raw_simulation/`
Raw FORMIND simulation output at **daily** resolution, organised per species. Each
species folder holds a **20-member ensemble** (`*_dynMort_0.h5` … `*_dynMort_19.h5`) of
dynamic-mortality runs on a simulated 10,000 ha stand — the full simulated forest
dynamics (per-year structure histograms and mortality) before any train/val/test
partitioning.

### `processed/`
Model-ready **pentad (5-day aggregate)** splits for **beech** and **pine**, named
`{split}_MBR_{species}_pentad_{n_years}_years_10000ha.h5`:

- **`MBR`** — the prediction target: (mortality) biomass rate.
- **`pentad`** — weather aggregated to 5-day steps.
- **`3_years`** — length of the input weather window per sample.
- **`10000ha`** — simulated stand area.

Each HDF5 file provides the arrays consumed by the training pipeline:

| Array | Meaning |
|-------|---------|
| `Xd`  | dynamic weather inputs (pentad precipitation, temperature, radiation time series) |
| `Xs`  | static / structural forest-state features |
| `Y`   | target biomass mortality rate |

The `bins_Xs_train_*.npy` files hold the histogram **bin edges** for the structure
variables (age, stem volume, LAI, height, diameter), used to reproduce / interpret the
`Xs` histograms.

## Intended uses

- Benchmarking sequence models (Transformers, RNNs, TCNs, etc.) on long weather
  time series with an ecological regression target.
- Studying **multi-modal** learning: combining dynamic weather with static forest
  structure.
- Investigating how weather variability propagates to forest mortality under
  quasi-stationary conditions.

## Quick start

```python
import h5py

# processed pentad split
with h5py.File("processed/train_MBR_beech_pentad_3_years_10000ha.h5", "r") as f:
    print(list(f.keys()))          # inspect available arrays
    # Xd, Xs, Y = f["Xd"][:], f["Xs"][:], f["Y"][:]

# one raw (daily) ensemble member
with h5py.File("raw_simulation/spruce/spruce_dynMort_0.h5", "r") as f:
    print(list(f.keys()))
```

```python
# download the whole dataset locally
from huggingface_hub import snapshot_download
snapshot_download("mohitanand/forest_mortality", repo_type="dataset")
```

## Models & tools

- **AWE-GEN** — hourly stochastic Advanced WEather GENerator.
- **FORMIND** — process-based, individual-based forest gap model.

## Authors

- **Mohit Anand**
- **Jakob Zscheischler**

## License

Released under **CC-BY-4.0**.