This repository contains a comprehensive data analytics project for the analysis of pasture biomass production in the Brazilian Cerrado using satellite remote sensing and laboratory chemical data. The project employs descriptive, predictive, and prescriptive analytics to provide actionable insights for agricultural management and environmental sustainability.
Created by: Yazan AlAtout
Course: Data Analytics
Topics: Agriculture · Analytics · Data Visualization · Machine Learning · Neural Networks · Satellite Data · Vegetation Analysis
- Descriptive Analytics Insights
- Predictive Analytics
- Prescriptive Analytics
- Key Recommendations
- Repository Structure
- Technologies Used
- Reference Materials
The analysis identified critical rainfall events throughout the year:
- Peak rainfall dates: 2022-12-14, 2023-01-11, 2023-01-26, 2023-02-09, 2023-03-01, and 2023-04-06
- Highest rainfall recorded: 2022-06-22 (unexpected peak outside typical seasonal patterns)
- Insight: Strong seasonal patterns can be leveraged for targeted agricultural planning
A comprehensive analysis of the relationship between Enhanced Vegetation Index (EVI), Biomass, and Rainfall reveals critical insights:
Key Finding: "Planting fewer plants during heavy rainfall periods is better than planting during heavy rain"
- High rainfall can reduce vegetation index due to cloud cover and waterlogging
- Optimal biomass development occurs under moderate rainfall conditions
- This insight directly impacts planting schedules and resource allocation
Low Acid Detergent Fiber (ADF) and Neutral Detergent Fiber (NDF) are critical indicators for forage quality:
- Better digestibility and higher energy content
- Optimal forage intake for improved livestock performance
- Essential for maximizing animal productivity
Geographic locations identified with optimal low ADF/NDF ratios have been mapped for targeted grazing management.
Most frequent DMC range: 26%-35%
This moderate balance of nutrients and moisture indicates:
- High-quality fresh forage
- Adequate nutritional value for grazing livestock
- Good overall pasture condition and management
Analysis of satellite spectral bands (B1, B8, B8A, etc.) reveals:
Key Findings:
- Distinct density peaks in each band indicate specific reflectance ranges dominate
- B8 and B8A bands show broader peaks, suggesting higher variability (vegetation/surface water indicators)
- Most frequent B8 location: (-20.446630, -54.839741) - appeared 4 times, indicating high reflectance in the 3500-4500 Hz range
Linear Relationships Discovered:
-
Dew Point vs Longwave Radiation: Positive correlation
- Higher humidity amplifies the greenhouse effect
- More water vapor = greater infrared radiation absorption
- Creates warmer, more favorable conditions for biomass growth
-
Solar Radiation vs Evapotranspiration: Positive correlation
- Higher solar radiation increases water loss through evaporation and plant transpiration
- Critical for water resource management and agricultural planning
Clear positive relationship between rainfall and Environmental Suitability Index:
- Higher rainfall = more stable environment for animal survival
- Essential for sustainable grazing management
Critical Discovery: Areas without animals (S1, S2 sample types) showed significantly higher biomass values
Impact of Grazing on Biomass:
- Restricted areas (No animals): Higher biomass due to undisturbed growth
- Grazed areas (With animals): Lower biomass due to vegetation consumption
- Biomass difference: Approximately 25-40% higher in animal-free zones
Legend:
- 🟤 White areas: Where animals are allowed
- 🟢 Green areas: Where animals are not allowed
- 🔵 Bubble size: Represents biomass volume
| Model | Feature Selection | R² Score | MAE | Epochs | Notes |
|---|---|---|---|---|---|
| LSTM (Multiple Random States) | None | 91.48% | 491.20 | 1000 | 🏆 BEST MODEL |
| LSTM | None | 87.20% | 901.68 | 700 | Good performance |
| LSTM with SelectKBest | SelectKBest (k=10) | 63.84% | 1521.86 | 500 | Poor feature selection |
| Linear Regression | Lasso | 60.56% | — | — | Limited performance |
| Linear Regression | SelectFromModel | 52.27% | — | — | Poorest performance |
| Time Series (Exponential Smoothing) | None | — | — | — | Alpha = 0.4-0.6 optimal |
- Best R² Score: 60.56%
- Methodology: 5-Fold Cross-Validation with multiple random states
- Conclusion: Moderate predictive power; Lasso outperformed SelectFromModel
Alpha Parameter Analysis:
- Alpha = 0.9: Overfits to recent data; not suitable for future forecasting
- Alpha = 0.4-0.6: Optimal range capturing patterns without overfitting
- Key Insight: Lower alpha values preserve long-term trends better
Performance Metrics:
- R² Score: 91.48% ✓
- MAE: 491.20 ✓
- Epochs: 1000
- Training: Multiple random states (5 iterations)
Why This Model Excels:
- Captures temporal dependencies in biomass data
- MinMaxScaler normalization optimizes neural network learning
- Multiple random state training ensures model consistency
- Larger epoch count allows better convergence
- R² Score: 63.84%
- MAE: 1521.86
- Issue: Selected only sample type features, lost important environmental variables
- Lesson: Feature selection can be counterproductive if it removes critical predictive variables
The developed LSTM model with 91.48% R² accuracy enables:
✅ Advance Biomass Forecasting: Predict highly productive days weeks in advance
✅ Risk Identification: Identify areas likely to underperform
✅ Harvest Planning: Optimize timing for maximum yield and quality
✅ Resource Allocation: Direct management efforts to high-potential areas
✅ Seasonal Strategy Development: Plan planting and grazing schedules based on forecasts
The Genetic Algorithm was applied to find optimal weights for the Environmental Stability Index by minimizing Mean Squared Error (MSE).
Algorithm Steps:
Parents are selected based on fitness values, preferring higher-performing individuals.
- Example parents: [0.279, 0.052, 0.151] and [0.434, 0.270, 0.853]
Genetic material is exchanged between parents at random points to create offspring.
Random modifications introduce genetic diversity and explore new solution spaces.
- Before: [0.279, 0.145, 0.072]
- After: [0.279, 0.145, 0.148]
Identical individuals are removed to maintain population diversity.
- Duplicates removed in final iteration: 7
Fitness values determine how well each individual solves the optimization problem.
Example Fitness Values:
Individual [0.0127, 0.0520, 0.0719] → Fitness: 10.18
Individual [0.0127, 0.1446, 0.2019] → Fitness: 1.52
Individual [0.0330, 0.3654, 0.0719] → Fitness: 48.24
Individuals are ranked by fitness; top performers advance to the next generation.
After 50 iterations of the Genetic Algorithm:
| Weight | Value | Interpretation |
|---|---|---|
| w₁ (Rainfall) | 0.27936118 | ~28% importance |
| w₂ (Max Temperature) | 0.14457443 | ~14% importance |
| w₃ (Vegetation Index - EVI) | 0.07190834 | ~7% importance |
Key Insight: Rainfall is the most critical factor (2× more important than temperature, 4× more important than vegetation index) for environmental stability in the Brazilian Cerrado.
The optimized weights enable:
🎯 Feature Importance Ranking: Understand which factors have greatest impact on environmental stability
🎯 Resource Prioritization: Focus management efforts on high-impact variables (rainfall management)
🎯 Strategic Planning: Develop policies emphasizing rainfall conservation and water management
🎯 Scenario Analysis: Model different rainfall scenarios to predict environmental outcomes
🎯 Risk Mitigation: Prepare contingencies for rainfall variability
- Optimize Planting Schedules: Avoid heavy planting during peak rainfall periods; focus on moderate rainfall conditions
- Target High-Quality Forage Areas: Prioritize grazing in identified low ADF/NDF zones for optimal livestock nutrition
- Implement Rotational Grazing: Restrict animals from optimal biomass areas periodically to maintain productivity (S-type areas show 25-40% higher biomass)
- Monitor DMC Levels: Maintain forage in the 26-35% DMC range for optimal nutritional quality
- Prioritize Water Management: Rainfall is the dominant factor (w₁=0.279) for environmental stability
- Manage Evapotranspiration: Monitor solar radiation and implement irrigation strategies during low-rainfall periods
- Track Environmental Indicators: Use Environmental Suitability Index to predict animal survival rates and grazing capacity
- Deploy LSTM Model: Use the 91.48% accuracy model for biomass forecasting 1-2 weeks in advance
- Identify High-Risk Areas: Flag zones predicted to underperform and allocate additional resources
- Optimize Harvest Timing: Coordinate harvest schedules based on biomass forecasts to maximize yield quality
satellite-biomass-analytics-main/
├── README.md # This file
├── cover-image.png # Header image used in the README
├── images/ # Exported notebook figures
│ ├── rainfall-analysis.png
│ ├── evi-biomass-rainfall-bubble.png
│ ├── ndf-adf-scatter.png
│ ├── optimal-forage-locations.png
│ ├── dmc-frequency-distribution.png
│ ├── spectral-bands-density.png
│ ├── dew-point-longwave-radiation.png
│ ├── solar-radiation-evapotranspiration.png
│ ├── rainfall-environmental-stability.png
│ └── biomass-animal-distribution.png
├── descriptive_analytics.ipynb # Descriptive analysis notebook
├── predictive_analytics.ipynb # Predictive analysis notebook
├── Complete_Dataset_updated.csv # Source dataset
├── sat-img.png # Satellite image overlay used in the map figure
└── Perspective Analytics/ # Genetic algorithm notes and outputs
- Jupyter Notebook (99.6% of codebase)
- Python (0.4% of codebase)
- Pandas - Data manipulation and analysis
- NumPy - Numerical computing
- Matplotlib & Seaborn - Data visualization
- Scikit-learn - Machine learning algorithms
- TensorFlow/Keras - Deep learning framework
- LSTM (Long Short-Term Memory) - Time series prediction
- Scikit-learn - Linear regression, feature selection (Lasso, SelectFromModel, SelectKBest)
- Genetic Algorithm - Optimization for prescriptive analytics
- MinMaxScaler - Feature normalization for neural networks
- SelectKBest, SelectFromModel, Lasso - Feature selection techniques
- Cross-Validation - Model evaluation (5-Fold K-Fold)
- Sentinel-2 Satellite Bands (B1, B8, B8A, etc.)
- Enhanced Vegetation Index (EVI) - Vegetation health indicator
- Geographic Analysis - Latitude/Longitude mapping
[1] Schick, B. (2023). Pasture and Forage Minute: Understanding ADF and NDF, Hay Quality After Calving. CropWatch, University of Nebraska-Lincoln. [Online] Available at: https://cropwatch.unl.edu/2023/pasture-and-forage-minute-understanding-adf-and-ndf-hay-quality-after-calving/
[2] DairyNZ (2023). Dry Matter Content Guidelines for Pasture Quality. [Online] Available at: www.dairynz.co.nz
- Python 3.8+
- Jupyter Notebook
- Required packages (see
requirements.txt)
# Clone the repository
git clone https://github.com/abstract-inf/satellite-biomass-analytics.git
cd satellite-biomass-analytics
# Install dependencies
pip install -r requirements.txt
# Launch Jupyter
jupyter notebook- Start with
notebooks/01_descriptive_analytics.ipynbfor data exploration - Move to
notebooks/02_predictive_analytics.ipynbfor model development - Conclude with
notebooks/03_prescriptive_analytics.ipynbfor optimization
Yazan AlAtout - yazanatout1@gmail.com
This project is provided as-is for educational purposes.
- Brazilian Cerrado pasture data providers
- Satellite remote sensing data (Sentinel-2)
- Laboratory chemical analysis data sources
- Course instructors and advisors in Data Analytics: Eng. Bassam Kasasbeh