-
Recreating code for the paper "Deep Learning Based Multi-Temporal Crop Classification"
-
There are a few minor differences - Zhong et al. use 46 element vector as time series, I use MODIS 16 day composite NDVI data (23 entries per year).
-
cnn-network.py
contains code for 1d cnn network described in the paper -
lstm-network.py
contains code for lstm described in the paper -
evaluate_baselines.py
contains code for XGBoost, SVM and RandomForest classification -
data_generator.py
assumes that each time series is stored as a separate csv with the NDVI values stored under the column 'NDVI'. It does nearest neighbor interpolation to fix missing values. -
Python version and required packages: Python3, Keras, Tensorflow, Scikit
-
Use Hufkens et al. to download MODIS data.
References:
- [1] Hufkens K. (2017) A Google Earth Engine time series subset script & library. DOI: 10.5281/zenodo.833789
- [2] Zhong, Liheng, Lina Hu, and Hang Zhou. "Deep learning based multi-temporal crop classification." Remote sensing of environment 221 (2019): 430-443.
Acknowledgements:
- Bharathkumar Ramachandra (tnybny)