Cropland Mapping

This project uses a deep learning framework for crop classification using multichannel inputs from Sentinel-2, combining high-resolution RGB imagery with vegetation indices (VIs) such as NDVI and GNDVI to capture both spatial and phenological crop characteristics. The model used is a convolutional long short-term memory (ConvLSTM) architecture to process time-series sequences of fused sensor data that integrates spatial features and temporal dependencies across multiple seasons. Input normalization, channel-wise sensor fusion, and data augmentation are used to mitigate class imbalance inherent in the dataset. It is evaluated on the Canadian Cropland Dataset and the model achieves ~90% classification accuracy.

Architecture

Findings

Class	Precision	Recall	F1-Score	Support
BARLEY	0.82	0.75	0.78	36
CANOLA	0.94	0.97	0.96	66
CORN	0.89	0.91	0.90	55
MIXEDWOOD	0.94	1.00	0.97	32
OAT	0.91	0.88	0.89	67
ORCHARD	0.98	0.92	0.95	51
PASTURE	0.86	0.91	0.89	47
POTATO	0.90	0.94	0.92	67
SOYBEAN	0.94	0.89	0.92	104
SPRING_WHEAT	0.56	0.71	0.62	7
Accuracy			0.91	532
Macro Avg	0.87	0.89	0.88	532
Weighted Avg	0.91	0.91	0.91	532

There is visible overfitting starting from the 4th epoch onwards which should be addressed in later works.

Project Setup

This project uses uv for package management. uv is an extremely fast Python package and project manager, written in Rust that can be used as a drop-in replacement for pip, pip-tools, pipx, poetry, pyenv, twine, virtualenv.

• uv Installation

  curl -LsSf https://astral.sh/uv/install.sh | sh

• Clone the Repository:

  git clone https://github.com/shama-llama/cropland-mapping.git
  cd cropland-mapping

• Create a Virtual Environment and Install Dependencies with uv:

  uv venv
  uv pip install -e .

• Activate the Virtual Environment:

  source .venv/bin/activate

• Launch Jupyter Notebook:

  jupyter notebook

  Navigate to the notebooks/ directory to run the analysis.

Notebooks

• cmap_dataset.ipynb: Data preparation with minor preprocessing and loading into HDF5 for better portability.
• cmap_eda.ipynb: Exploratory data analysis.
• cmap_msi.ipynb: ConvLSTM model training and evaluation.

Model

The trained ConvLSTM model weights are saved in the models/ directory.

torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model.load_state_dict(torch.load('models/best_model.pth', map_location=device))

References

A. A. B. Jacques, A. B. Diallo, and E. Lord, “The Canadian Cropland Dataset: A New Land Cover Dataset for Multitemporal Deep Learning Classification in Agriculture,” June 04, 2023, arXiv: arXiv:2306.00114. doi: 10.48550/arXiv.2306.00114.

M. O. Turkoglu et al., “Crop mapping from image time series: Deep learning with multi-scale label hierarchies,” Remote Sensing of Environment, vol. 264, p. 112603, Oct. 2021, doi: 10.1016/j.rse.2021.112603.

D. Darwish, “Improving Techniques for Convolutional Neural Networks Performance,” European Journal of Electrical Engineering and Computer Science, vol. 8, pp. 1–16, Jan. 2024. Number: 1.

J. M. Johnson and T. M. Khoshgoftaar, “Survey on deep learning with class imbalance,” Journal of Big Data, vol. 6, pp. 1–54, Dec. 2019. Number: 1 Publisher: SpringerOpen.

C. Shorten and T. M. Khoshgoftaar, “A survey on Image Data Augmentation for Deep Learning,” Journal of Big Data, vol. 6, pp. 1–48, Dec. 2019. Number: 1 Publisher: SpringerOpen.

License

This project is licensed under the terms of the MIT open source license.