Interpretable Multi-Source Machine Learning for Urban Classification in Data-Scarce Regions

Conference Poster

The final conference poster presented at Deep Learning Indaba 2026 is available in the poster/ directory.

Overview

This repository accompanies the accepted paper:

Interpretable Multi-Source Machine Learning for Urban Classification in Data-Scarce Regions

presented at the Deep Learning Indaba (DLI) 2026.

The study develops an interpretable machine learning framework for urban classification by integrating multi-source remote sensing, topographic, demographic, and night-time light datasets. Three machine learning models—Random Forest, XGBoost, and a Neural Network—are evaluated, while SHAP (SHapley Additive exPlanations) is employed to provide transparent explanations of model predictions.

The repository contains the complete Jupyter notebook, supporting figures, and the resources required to reproduce the analyses presented in the paper.

---

Repository Structure

interpretable-urban-classification/
│
├── notebooks/
│   └── Urban_Classification_Workflow.ipynb
│
├── data/
│   ├── urban_2000.csv
│   ├── urban_2005.csv
│   ├── urban_2010.csv
│   ├── urban_2015.csv
│   ├── urban_2020.csv
│   ├── urban_2024.csv
│   └── Lagos_full_stack_100m.tif
│
├── figures/
│   ├── Figure1_*.png
│   ├── Figure2_*.png
│   ├── Figure3_SHAP_Beeswarm.png
│   ├── Figure4_SHAP_Feature_Importance.png
│   ├── Figure5_SHAP_Dependence_NightLights.png
│   └── Figure6_Urban_Probability_2020.png
│
├── paper/
│
├── poster/
│
├── results/
│
├── README.md
├── LICENSE
├── requirements.txt
├── environment.yml
└── CITATION.cff

---

Study Area

The study focuses on Lagos, Nigeria, one of Africa's fastest-growing metropolitan regions. Urban classification was performed using multi-temporal Earth observation and socio-economic datasets spanning 2000–2024.

---

Predictor Variables

The machine learning models were developed using ten predictor variables:

Variable	Description
Blue	Landsat Blue Band
Green	Landsat Green Band
Red	Landsat Red Band
NIR	Near Infrared
SWIR	Shortwave Infrared
NDVI	Normalized Difference Vegetation Index
NDBI	Normalized Difference Built-up Index
NightLights	Night-time Light Intensity
Population	Population Density
Elevation	Digital Elevation Model

The response variable is:

• Urban
  • 0 = Non-Urban
  • 1 = Urban

---

Machine Learning Models

Three supervised learning models were evaluated:

• Random Forest
• XGBoost
• Neural Network (PyTorch)

Model interpretability was achieved using SHAP (SHapley Additive exPlanations).

---

Results

The Random Forest and XGBoost models achieved the strongest predictive performance, with ROC–AUC values exceeding 0.94.

SHAP analysis identified NightLights, NDBI, Population, and NDVI among the most influential variables contributing to urban classification.

---

Reproducing the Analysis

Clone the repository:

git clone https://github.com/YOUR_USERNAME/interpretable-urban-classification.git

Create the Conda environment:

conda env create -f environment.yml

Activate the environment:

conda activate geoenv

Launch Jupyter Lab:

jupyter lab

Open:

notebooks/Urban_Classification_Workflow.ipynb

Run all cells sequentially.

---

Citation

If you use this repository, please cite:

Ugege, P. (2026). Interpretable Multi-Source Machine Learning for Urban Classification in Data-Scarce Regions. Deep Learning Indaba (DLI 2026).

(Citation will be updated upon publication.)

---

Contact

Peter Ugege

Forestry Research Institute of Nigeria (FRIN)

Email: [add preferred contact email]

---

Acknowledgements

The authors acknowledge the Deep Learning Indaba (DLI) community for promoting open and reproducible artificial intelligence research across Africa.