Crop Classification DL Project

This repository contains the code and data pipeline for classifying crop types in Cyprus using Sentinel-2 time-series data and street-level imagery. Two hybrid deep-learning models are implemented:

• HybridResNetLSTM: Combines a ResNet-50 image encoder with an LSTM-based time-series branch.
• HybridResNetTempCNN: Combines a ResNet-50 image encoder with a 1D temporal CNN branch.

Table of Contents

• Project Structure
• Data
• Installation
• Usage
• Models
• Training
• Evaluation
• Contributing
• License

Project Structure

├── nasa_data/
│   ├── fs_sentinel_cyprus_2022_buffer5_interp.csv  # Sentinel-2 time-series
│   ├── meta/
│   │   ├── cyprus_2022_meta.csv                    # Field metadata with crop labels
│   │   ├── cyprus_2022_meta.shp                    # Shapefile for field geometries
│   │   └── cyprus_2022_street_level_meta.csv       # Metadata for street-level images
│   └── street_level/                               # Directory of Mapillary street images
├── CS_Project_V1
├── CS_V2
├── CS_DataHandling
├── Data_Description
├── Data_structure.md
├── requirements.txt        # Python dependencies
└── README.md               # Project overview (this file)

Data

The data pipeline uses:

1. Sentinel-2 optical & Sentinel-1 SAR time-series

   • Preprocessed to 5-day temporal resolution, with spectral bands and vegetation indices.

2. Mapillary street-level images

   • Ground-level imagery to add spatial context.

3. Metadata

   • Field identifiers, crop codes/names, area, pixel counts.

Place all data files under the data/ directory as shown above.

Installation

1. Clone this repository:

   git clone https://github.com/theBlackfish01/CropStats
   cd crop-classification-dl

2. Create a Python environment (recommended with conda):

   conda create -n crop-classify python=3.11
   conda activate crop-classify

3. Install dependencies:

   pip install -r requirements.txt

Usage

Quick Shape Check

In the Python REPL or notebook, run:

from src.data_loader import get_loaders_fast
train_loader, val_loader = get_loaders_fast(
    meta_csv   ='data/meta/cyprus_2022_meta.csv',
    sentinel_csv='data/fs_sentinel_cyprus_2022_buffer5_interp.csv',
    image_dir  ='data/street_level/',
    code_to_idx=code_to_idx_map,
    num_classes=num_classes,
    batch_size =32,
    num_workers=0
)
batch = next(iter(train_loader))
_, T_steps, C_feats = batch['series'].shape
print(f"Time steps: {T_steps}, Bands: {C_feats}")

Training

python src/train.py \
  --meta_csv data/meta/cyprus_2022_meta.csv \
  --sentinel_csv data/fs_sentinel_cyprus_2022_buffer5_interp.csv \
  --image_dir data/street_level/ \
  --batch_size 32 \
  --epochs 30 \
  --model lstm  # or "tempcnn"

Model checkpoints (.pth) will be saved in the project root.

Models

• HybridResNetLSTM: Uses a pre-trained ResNet-50 (images) + bidirectional LSTM (time-series).
• HybridResNetTempCNN: Uses a pre-trained ResNet-50 (images) + 1D temporal CNN.

Both share a fusion MLP head and support freezing the CNN backbone.

Evaluation

Validation accuracy and loss are printed each epoch, with early stopping based on validation loss. Adjust hyperparameters in src/train.py as needed.

Contributing

Contributions are welcome! Please open an issue or pull request for bug fixes, improvements, or new features.

License

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