---
license: mit
tags:
  - soil-science
  - pedotransfer-function
  - water-retention
  - attention-mechanism
  - ensemble
  - tensorflow
  - geoscience
pipeline_tag: tabular-regression
library_name: tensorflow
datasets:
  - custom
language:
  - en
---

# HABIT — Hierarchical Attention-Based Inference with Transfer Learning

A pre-trained ensemble model for predicting soil water retention curves from basic soil properties.

**Paper:** Ghezzehei TA (2025). Interpretable Soil Water Retention Prediction Using Hierarchical Attention Networks with Uncertainty Quantification. *Water Resources Research*. [DOI pending]

**Training data & code:** [Dryad repository](https://datadryad.org/share/LINK_NOT_FOR_PUBLICATION/iXOXxn5h-Hdq9m7aA3cTLgGMAHiHMre2HxZItyI-Ykc)

## What HABIT does

HABIT predicts volumetric water content (θ) at any water potential (ψ) from soil properties. Unlike parameter-based approaches (e.g., van Genuchten fitting), HABIT predicts retention curves **directly** — no ill-posed inverse problem.

### Input → Output

| Input | Required? | Units |
|---|---|---|
| Sand, Silt, Clay | **Yes** | fraction (0–1) |
| Bulk density | Optional | g/cm³ |
| Organic carbon | Optional | fraction (0–1) |
| Saturated hydraulic conductivity | Optional | cm/day |
| Water potential(s) | **Yes** | kPa |

**Output:** Volumetric water content (cm³/cm³) with ensemble uncertainty (mean ± std from 20 models).

### Automatic input adaptation

HABIT uses a single model architecture trained hierarchically. Provide whatever soil properties you have — the model adapts automatically:

- **Texture only** → equivalent to Stage 0 (R² ≈ 0.78)
- **Texture + BD** → equivalent to Stage 1 (R² ≈ 0.85)
- **Texture + BD + OC** → equivalent to Stage 2 (R² ≈ 0.86)
- **All properties** → Stage 3 (R² ≈ 0.92)

More properties = better predictions, but texture alone already works.

## Quick start

```python
from habit_inference import HABITPredictor, load_ensemble

# Load the 20-member ensemble
predictor = load_ensemble(stage=3)

# Predict for a single soil
import pandas as pd
soil = pd.DataFrame({
    'soil_id': ['my_soil'],
    'sand': [0.40],
    'silt': [0.35],
    'clay': [0.25],
    'bd': [1.35],       # optional
    'oc': [0.012],      # optional
    'ksat': [25.0]      # optional, cm/day
})

predictions = predictor.predict(soil)
# Returns: soil_id, water_potential_kPa, water_content_mean, water_content_std
```

### From CSV

```python
predictor.predict_from_csv('input.csv', 'output.csv')
```

### Command line

```bash
habit-predict --input soils.csv --output predictions.csv
```

## Model architecture

HABIT uses property-specific encoders, cross-attention layers for property interactions, multi-head attention over soil properties and water potentials, and a monotonic output layer that enforces physically correct behavior (water content decreases with increasing tension).

**Key design:** Hierarchical transfer learning trains the model sequentially — texture first, then adding bulk density, organic carbon, and Ksat — so a single set of weights handles any combination of available inputs via masking.

### Architecture parameters

| Parameter | Value |
|---|---|
| Embedding dimension | 192 |
| Attention heads | 4 |
| Monotonic basis functions | 40 |
| Dropout rate | 0.15 |
| Ensemble members | 20 |

## Repository contents

```
weights/
  member_01.h5 ... member_20.h5   # Hierarchical ensemble weights (19 MB each)
model/
  habit.py                         # Model architecture
  layers.py                        # Custom layers (attention, monotonic)
config/
  model_config.json                # Architecture hyperparameters
  scaler_params.json               # Feature scaling parameters (robust scaling)
```

## Performance

Evaluated on held-out test set (cluster bootstrap, 1000 iterations):

| Configuration | R² | RMSE (cm³/cm³) | MAE (cm³/cm³) |
|---|---|---|---|
| HABIT Stage 0 (texture) | 0.779 [0.737, 0.817] | 0.067 [0.060, 0.074] | 0.049 [0.044, 0.055] |
| HABIT Stage 1 (+BD) | 0.846 [0.748, 0.906] | 0.056 [0.044, 0.070] | 0.039 [0.033, 0.049] |
| HABIT Stage 2 (+OC) | 0.862 [0.781, 0.920] | 0.052 [0.040, 0.066] | 0.038 [0.030, 0.047] |
| HABIT Stage 3 (+Ksat) | 0.923 [0.899, 0.944] | 0.043 [0.036, 0.050] | 0.030 [0.026, 0.035] |
| Rosetta Model 2 (texture) | 0.009 | 0.141 | 0.113 |
| Rosetta Model 3 (+BD) | 0.511 | 0.099 | 0.075 |

## Training data

2,577 soil samples compiled from two international databases (Hohenbrink et al. 2023; Gupta et al. 2022), with 44,636 water retention measurements after adaptive thinning. See the [Dryad repository](https://datadryad.org/share/LINK_NOT_FOR_PUBLICATION/iXOXxn5h-Hdq9m7aA3cTLgGMAHiHMre2HxZItyI-Ykc) for the full dataset and training code.

## License

MIT (code and weights). Training data: CC BY 4.0.

## Citation

```bibtex
@article{ghezzehei2025habit,
  title={Interpretable Soil Water Retention Prediction Using Hierarchical Attention Networks with Uncertainty Quantification},
  author={Ghezzehei, Teamrat A.},
  journal={Water Resources Research},
  year={2025},
  note={DOI pending}
}
```

## Contact

Teamrat A. Ghezzehei — [taghezzehei@ucmerced.edu](mailto:taghezzehei@ucmerced.edu)
Life and Environmental Sciences, University of California, Merced