Point-based Rocky Terrain Prediction

Aim

To predict the presence of rocky terrain using topography and remote sensing data. This information includes:

1. Topography Features (aspect, elevation, flow length, plan curvature, profile curvature, slope, tan curvature, twi)

2. Remote Sensing Features (vegetation index, moisture index, bulk density, soil organic carbon, clay, sand)

Note: Every coordinate is considered in isolation and hence this model follows a point-based approach.

Environment Setup

• Clone this repository.

git clone https://github.com/SoilForestHealth/rocky-terrain-classifier.git

• Install all packages in the requirements.txt file.

pip install -r requirements.txt

• The following directory structure is required for the code in this repository to work properly:

├── data
│   ├── LSF_Grid_Soil_Data_2025_Summer.xlsx
│   ├── LSF_Topography_Covariates_2025_Summer.csv
│   ├── metadata
│   │   ├── data.json
│   │   ├── system.json
│   │   └── tune.json
│   ├── raw_field_summer_2025_covariates_combined.csv
│   └── s2_cloudless_covariates_field_summer_2025_combined.csv
├── main.py
├── pipeline
│   ├── evaluate.py
│   ├── model.py
│   ├── preprocess.py
│   └── select.py
└── requirements.txt

4 directories, 13 files

• To execute the code in this repository, run main.py file. Ensure you are in the root directory of the repository.

python3 main.py

• Feel free to raise an issue if there are any problems with the repository!

Results

To evaluate the models for the imbalanced binary classification problem, we use macro_avg_recall and f2_score.

$f2\_score = \frac{5 \cdot P \cdot R}{4 \cdot P + R}$

$macro\_avg\_recall = \frac{1}{2} \sum_{i=1}^{2} R_i$

Model	macro_avg_recall	f2_score	wandb sweeps
Logistic Regression	0.557	0.557	wandb
Decision Tree	0.654	0.658	wandb
Random Forest	0.723	0.731	wandb
Extra Trees	0.696	0.700	wandb
Gradient Boosting Trees	0.724	0.724	wandb

The overall model comparison dashboard is here.