Crop Classification with Deep Learning

Sentinel-2 Satellite Imagery | Elgabel Region, Sudan

---

9 Models | 24 Spectral Features | 100% Accuracy | 104 km² Classified

Multi-class crop classification in the Elgabel Region, Sudan using Sentinel-2 satellite imagery and multiple machine learning / deep learning models. The pipeline covers data acquisition (Google Earth Engine), exploratory analysis, classical ML training (scikit-learn, XGBoost), PyTorch deep learning, and wall-to-wall satellite image classification.

---

Table of Contents

• Study Area
• Workflow
• Results
• Data Exploration
• Model Training Details
• Data Description
• Model Architectures
• Project Structure
• Setup
• Usage
• Google Earth Engine
• Contributing
• Citation
• Changelog
• License

---

Study Area

Elgabel Region, Sudan — an agricultural zone where the following five crop/land-cover classes are mapped using Sentinel-2 imagery from Q1 2020:

Class ID	Name	Color	Samples	Percentage
0	Cotton		337	1.4%
1	Wheat		7,901	32.2%
2	Fallow		11,150	45.4%
3	Grass		5,024	20.5%
4	Water		144	0.6%

Total: 24,556 labeled samples | Imbalance ratio: 77.4:1 (Fallow vs Water)

---

Workflow

graph TD
    A["1. Data Acquisition<br/><i>Google Earth Engine</i>"] -->|Sentinel-2 Composite + Indices| B
    B["2. Data Exploration<br/><i>01_data_exploration.py</i>"] -->|24,556 samples / 24 features| C
    C["3. Preprocessing & ML Training<br/><i>02_preprocessing_and_model.py</i>"] --> D
    C --> E

    subgraph sklearn ["Scikit-learn / XGBoost Models"]
        D["MLP · MLP+SMOTE<br/>XGBoost · XGBoost+SMOTE<br/>Random Forest"]
    end

    D -->|Best: Random Forest| F["4. Apply ML to Image<br/><i>03_apply_to_image.py</i>"]

    subgraph pytorch ["PyTorch Deep Learning Models"]
        E["SpectralMLP · SpectralCNN1D<br/>SpectralHybrid · SpectralAttention"]
    end

    E -->|"<i>04_pytorch_models.py</i>"| G["5. Apply DL to Image<br/><i>05_apply_pytorch_to_image.py</i>"]

    F --> H["Classification Map<br/>104 km² · 5 classes"]
    G --> H

    style A fill:#4CAF50,stroke:#333,color:#fff
    style B fill:#2196F3,stroke:#333,color:#fff
    style C fill:#FF9800,stroke:#333,color:#fff
    style D fill:#9C27B0,stroke:#333,color:#fff
    style E fill:#E91E63,stroke:#333,color:#fff
    style F fill:#00BCD4,stroke:#333,color:#fff
    style G fill:#00BCD4,stroke:#333,color:#fff
    style H fill:#F44336,stroke:#333,color:#fff

Loading

---

Results

Scikit-learn & XGBoost Models

Model	Accuracy	Macro F1	Weighted F1	Kappa
MLP Neural Network	99.98%	0.9984	0.9998	0.9997
MLP + SMOTE	99.98%	0.9984	0.9998	0.9997
XGBoost	99.98%	0.9998	0.9998	0.9997
XGBoost + SMOTE	99.98%	0.9998	0.9998	0.9997
Random Forest	100.00%	1.0000	1.0000	1.0000

PyTorch Deep Learning Models

Model	Accuracy	Precision	Recall	Macro F1	Kappa	Parameters
SpectralMLP	99.96%	0.9942	0.9996	0.9969	0.9994	78,853
SpectralCNN1D	99.49%	0.9756	0.9964	0.9854	0.9922	40,069
SpectralHybrid	100.00%	1.0000	1.0000	1.0000	1.0000	43,717
SpectralAttention	100.00%	1.0000	1.0000	1.0000	1.0000	106,309

Classification Maps

Random Forest (sklearn) — Satellite Image Classification

SpectralHybrid (PyTorch) — Satellite Image Classification

Area Statistics (SpectralHybrid Model)

Class	Area (ha)	Area (km²)	Percentage
Cotton	463.0	4.63	4.4%
Wheat	1,373.9	13.74	13.2%
Fallow	5,839.9	58.40	55.9%
Grass	2,672.5	26.73	25.6%
Water	89.3	0.89	0.9%

---

Data Exploration

Class Distribution

Feature Distributions by Class (click to expand)

Key Features — Box Plots (click to expand)

Feature Correlation Matrix (click to expand)

Highly correlated feature pairs (|r| > 0.99):

• CIgreen ↔ GCVI: 1.000
• EVI ↔ MSAVI: 0.999
• EVI ↔ SAVI: 0.999
• MSAVI ↔ SAVI: 0.998
• B7 ↔ B8A: 0.995
• NDRE ↔ SAVI: 0.994
• B2 ↔ B3: 0.994

---

Model Training Details

Sklearn — MLP Training Curves (click to expand)

Sklearn — Confusion Matrices (All 5 Models) (click to expand)

XGBoost — Feature Importance (click to expand)

Per-Class Accuracy Comparison (All Sklearn Models) (click to expand)

PyTorch — Training & Validation Loss Curves (click to expand)

PyTorch — Confusion Matrices (All 4 Models) (click to expand)

Classification Map Visualizations (click to expand)

Random Forest (sklearn)

SpectralHybrid (PyTorch)

---

Data Description

24 spectral features derived from a Sentinel-2 Q1 2020 composite:

Category	Features
Spectral Bands (10)	B2, B3, B4, B5, B6, B7, B8, B8A, B11, B12
Vegetation Indices (14)	NDVI, EVI, SAVI, NDRE, NDRE2, GNDVI, NDMI, BSI, MNDWI, LSWI, GCVI, WDRVI, CIgreen, CIrededge, MSAVI

---

Model Architectures

Scikit-learn / XGBoost (Step 02)

Model	Architecture	Imbalance Strategy
MLP Neural Network	24 → 256 → 128 → 64 → 32 → 5	Class weights
MLP + SMOTE	24 → 512 → 256 → 128 → 64 → 32 → 5	SMOTE oversampling
XGBoost	500 trees, depth=8, lr=0.05	Class weights
XGBoost + SMOTE	500 trees, depth=8, lr=0.05	SMOTE oversampling
Random Forest	500 trees, balanced	Balanced class weights

PyTorch Deep Learning (Step 04)

Model	Description	Key Features
SpectralMLP	Deep MLP (24 → 128 → 256 → 128 → 64 → 5)	BatchNorm, Dropout(0.3), FocalLoss
SpectralCNN1D	1D CNN (32 → 64 → 128 filters)	Treats spectral bands as 1D signal
SpectralHybrid	CNN + MLP dual-branch fusion	Local + global feature extraction
SpectralAttention	Transformer-style	CLS token, 4 heads, 2 layers, GELU

All PyTorch models use: FocalLoss (gamma=2), AdamW (weight_decay=1e-4), ReduceLROnPlateau, early stopping (patience=15), SMOTE-balanced data. Trained on NVIDIA RTX 3050 (4GB VRAM).

---

Project Structure

View full directory tree (click to expand)

crop-classification-deep-learning/
├── .github/
│   ├── ISSUE_TEMPLATE/
│   │   ├── bug_report.md
│   │   └── feature_request.md
│   ├── workflows/
│   │   └── ci.yml
│   └── PULL_REQUEST_TEMPLATE.md
├── docs/                          # Result visualizations
│   ├── class_distribution.png
│   ├── feature_distributions.png
│   ├── feature_boxplots.png
│   ├── correlation_matrix.png
│   ├── training_history.png
│   ├── confusion_matrices.png
│   ├── feature_importance.png
│   ├── per_class_accuracy.png
│   ├── pytorch_training_curves.png
│   ├── pytorch_confusion_matrices.png
│   ├── classification_visualization.png
│   ├── truecolor_vs_classification.png
│   ├── pytorch_classification_visualization.png
│   └── pytorch_truecolor_vs_classification.png
├── gee/
│   └── crop_classification_gee.js
├── 01_data_exploration.py
├── 02_preprocessing_and_model.py
├── 03_apply_to_image.py
├── 03_diagnose_bands.py
├── 04_pytorch_models.py
├── 05_apply_pytorch_to_image.py
├── .gitignore
├── CHANGELOG.md
├── CITATION.cff
├── CONTRIBUTING.md
├── LICENSE
├── Makefile
├── README.md
├── environment.yml
├── fix_jupyter_kernel.bat
└── requirements.txt

---

Setup

Option A: Conda (recommended)

conda env create -f environment.yml
conda activate geodl

Option B: pip

pip install -r requirements.txt

Option C: Makefile

make setup        # conda
make setup-pip    # pip

Register Jupyter kernel (optional)

fix_jupyter_kernel.bat

3. Obtain satellite data

Use the Google Earth Engine script in gee/crop_classification_gee.js to export:

• S2_composite_24bands_2020_Q1.tif — 24-band Sentinel-2 composite
• S2_key_indices_2020_Q1.tif — Key spectral indices
• crop_training_data_5classes_2020.csv — Labeled training samples

Place the exported files in the project root directory.

---

Usage

Run the scripts in order:

# Step 1: Explore the training data
python 01_data_exploration.py

# Step 2: Train scikit-learn and XGBoost models
python 02_preprocessing_and_model.py

# Step 3: Apply best ML model to full satellite image
python 03_apply_to_image.py

# Step 4: Train PyTorch deep learning models (GPU recommended)
python 04_pytorch_models.py

# Step 5: Apply best PyTorch model to full satellite image
python 05_apply_pytorch_to_image.py

Or run the entire pipeline with one command:

make run-all

Utility Scripts

• 03_diagnose_bands.py — Diagnose band ordering between training data and GeoTIFF
• Run make help to see all available commands

---

Google Earth Engine

The gee/crop_classification_gee.js script handles data acquisition:

1. Defines the study area (Elgabel Region, Sudan)
2. Filters Sentinel-2 Surface Reflectance (Q1 2020, <20% cloud cover)
3. Applies cloud masking using the SCL band
4. Computes a cloud-free median composite (10 spectral bands)
5. Calculates 14 spectral indices (NDVI, EVI, SAVI, NDRE, GNDVI, NDMI, BSI, MNDWI, LSWI, GCVI, WDRVI, CIgreen, CIrededge, MSAVI)
6. Samples training points from labeled polygons (5 classes)
7. Exports the 24-band composite and training CSV to Google Drive

---

Contributing

Contributions are welcome! Please read the Contributing Guidelines before submitting a pull request.

---

Citation

If you use this project in your research, please cite it:

@software{osman_crop_classification,
  title  = {Crop Classification with Deep Learning},
  author = {Osman},
  url    = {https://github.com/Osman-Geomatics93/crop-classification-deep-learning},
  license = {MIT}
}

You can also use the "Cite this repository" button on GitHub (powered by CITATION.cff).

---

Changelog

See CHANGELOG.md for version history and release notes.

---

License

This project is licensed under the MIT License. See LICENSE for details.