🌾 Emmet County Crop-Type-Prediction-using-Machine-Learning 🛰️

🔍 Project Overview

This repository contains a land cover classification project for Emmet County, Michigan, using two distinct approaches:

1. Random Forest Classification (Machine Learning)
2. Probability-Based Transition Model

The project leverages historical satellite imagery from 2008–2021 to predict 2021 land cover, with a focus on corn and soybean rotations.

🗂 Data & Assumptions

• Input Data: TIFF raster images from 2008–2021 (30m resolution, EPSG:32615).
• Metadata: crop_metadata.json maps crop codes to names, with a focus on corn (1) and soybean (5).
• Labels: 2021 pixel classes serve as ground truth, with 2020 held out for validation.
• Class Reclassification: All land covers outside corn/soybean collapsed into an "Other" category.
• Sampling Strategy: Spatially balanced 10x10 grid sampling, selecting 10% of pixels evenly across the county.
• Feature Engineering: Temporal land cover values from the past 14 years.

🧠 Methods

Method 1: Random Forest Classifier

1. Data Preparation

   • Extracted pixel values from 2008–2020.
   • Split dataset into train (50%), validation (20%), test (30%).

2. Model Training & Optimization

   • Grid search to optimize max_depth and class_weight to handle class imbalance.
   • Feature selection based on historical importance.

3. Evaluation & Prediction

   • Best model selected via Precision-Recall AUC (PR-AUC).
   • Applied trained model to classify every pixel for 2021 land cover.

Method 2: Probability-Based Transition Model

1. Transition Probabilities Computation

   • Grouped land cover values over past 7 years.
   • Calculated historical transition probabilities for 2021 predictions.

2. Validation & Selection of Best History Length

   • Tested multiple history lengths (3, 5, 7 years).
   • Selected best-performing sequence based on PR-AUC analysis.

3. Final Prediction & Evaluation

   • Applied maximum probability classification rule.
   • Compared accuracy, precision, recall, and PR-AUC with Random Forest.

📈 Results

🌳 Random Forest Model (RF_65)

• Validation PR-AUC: 1.000 (best performing model).
• Test Accuracy: 99.64%
• Class-wise performance:
  • Corn: Precision 99.55%, Recall 99.71%
  • Soybean: Precision 99.57%, Recall 99.54%
  • Other: Precision 99.85%, Recall 99.67%
• Feature Importance: 2015, 2017, 2018 were most influential.

📊 Probability-Based Model (7-Year History)

• Validation PR-AUC: Stabilized at 7 years.
• Test Accuracy: 91.99%
• Class-wise performance:
  • Corn: Precision 88.49%, Recall 92.7%
  • Soybean: Precision 90.62%, Recall 88.3%
  • Other: Precision 98.76%, Recall 94.79%

🚧 Challenges & Future Refinements

🛠️ Challenges Faced

• ⚠️ Resolution Mismatch: The 2007 dataset (56m) was dropped for consistency
• ⚠️ Class Simplification: Non-corn/soy crops grouped into "Other"
• ⚠️ Temporal Gaps: Annual snapshots limit finer-scale transitions
• ⚠️ Stationarity Assumption: Model assumes historical transition rules remain unchanged

🔥 Future Improvements

• 🚀 Hybrid Model: Mixing Random Forest learning with probability-based classification for efficiency & accuracy
• 🗺️ Spatial Analysis: Introducing geospatial trends to refine predictions
• 📊 Dynamic Transition Modeling: Moving toward adaptive transition probabilities rather than fixed history window

How to Use This Repository

1. Clone the Repository

   git clone https://github.com/your-username/emmet-landcover.git
   cd emmet-landcover

2. Install Dependencies

3. Run Random Forest Model

4. Run Probability-Based Model

5. Analyze Results

   • Outputs are saved as classified TIFFs and evaluation reports.

👥 Contributors

• Calvin Samwel Swai – Remote Sensing & Machine Learning Scientist Candidate

🌟 Thank you for checking out this project! 🌍 Feel free to open issues, contribute, or reach out if you have feedback.