🌍 Land Use/Land Cover (LULC) Classification using Random Forest

Python | Rasterio | Scikit-Learn | GeoPandas

This project performs a supervised Land Use/Land Cover (LULC) classification using a Random Forest machine learning model applied to multispectral Landsat imagery. Digital Elevation Model (DEM) and slope layers are incorporated as auxiliary predictor variables to improve classification accuracy.

The workflow extracts pixel values under point-based training samples, builds a classification model, evaluates its performance, and generates a classified LULC raster. This methodology is well suited for applications in:

• Remote sensing
• Environmental and ecological mapping
• Hydrological studies
• Urban and regional planning
• Agricultural monitoring

---

✅ Features

• Load multispectral imagery (Landsat)
• Integrate DEM and slope layers as additional predictors
• Extract training data from point-based shapefile
• Train a Random Forest classifier
• Perform train/test accuracy assessment
• Predict LULC for the full study area
• Export a classified GeoTIFF raster
• Visualize the final LULC map

---

📦 Prerequisites

✔ Required Python Libraries

pip install geopandas rasterio numpy pandas scikit-learn matplotlib

---

📁 Input Data Requirements

All input datasets must follow these conditions:

---

1. Coordinate Reference System (CRS)

The following must use the same CRS:

• Landsat composite raster
• DEM raster
• Slope raster
• Training point shapefile

CRS mismatches will result in incorrect pixel extraction.

---

2. Spatial Resolution Requirements

DEM and slope layers must be:

• Resampled to match the Landsat pixel resolution (~30 m)
• Aligned so that all rasters share the same grid structure

---

3. Spatial Extent

DEM and slope must be:

• Clipped to the Landsat extent
• Having matching rows, columns, and geotransform

This ensures all layers stack correctly.

---

🎯 Training Sample Requirements (Point-Based Sampling)

The classification workflow uses point training samples, where each point represents a known land-cover class.

✔ Training Data Format

• Shapefile (.shp)

• Geometry: POINT

• Attribute field:

  • Class → integer code representing the land-cover class

• CRS identical to all rasters

Example Attribute Table

Point_ID	Class	Geometry
1	1 (Water)	POINT(x, y)
2	3 (Vegetation)	POINT(x, y)
3	5 (Barren)	POINT(x, y)

Each point is used to extract:

Band1, Band2, Band3, Band4, Band5, Band6, Band7, DEM, Slope

Resulting Training Table Format

Class	Band1	Band2	Band3	Band4	Band5	Band6	Band7	DEM	Slope
Barren	123	98	76	45	23	12	5	255	14
Vegetation	45	63	88	123	145	110	95	300	9
Water	10	20	25	30	15	10	8	201	2

This feature matrix is built automatically by the script.

---

🧭 Why Use Point Samples?

• Avoids boundary errors from polygons
• Ideal for pixel-based machine learning workflows
• Ensures exact spatial correspondence between training data and raster grid
• Faster and more memory-efficient
• Works cleanly with Rasterio and NumPy

---

📁 Recommended Project Structure

project/
├── data/
│   ├── landsat_composite.tif
│   ├── dem_resampled.tif
│   ├── slope_resampled.tif
│   ├── training_samples.shp
│   └── ...
├── src/
│   └── classify_lulc.py
└── README.md

---

▶️ Running the Classification Script

1. Update file paths:

landsat_path = '/path/to/landsat.tif'
train_shp_path = '/path/to/training_points.shp'
dem_path = '/path/to/dem_resampled.tif'
slope_path = '/path/to/slope_resampled.tif'

2. Execute the script:

python classify_lulc.py

3. Output Files Generated:

• LULC_2016.tif → Classified raster
• Accuracy metrics (precision, recall, F1-score)
• Matplotlib visualization of classification

---

📊 Output Example

• GeoTIFF classified map saved to disk
• Displayed map using a color-coded scheme
• Accuracy evaluation for model validation

---

🧠 Why Random Forest?

Random Forest is a powerful classifier for remote sensing because:

• Handles nonlinear relationships
• Works well with high-dimensional data
• Robust against noise
• Requires no assumptions about data distribution
• Provides high accuracy for mixed Land Cover data

---

💡 Potential Extensions

• Add vegetation/water indices (NDVI, NDWI, SAVI)
• Add terrain-based predictors (TWI, TRI, HAND)
• Add GLCM texture metrics
• Tune model using GridSearchCV
• Generate probability/confidence maps
• Add lineament density as a predictor
• Perform temporal LULC change analysis

---