to tag satellite data chips with information on atmospheric conditions and land use/land cover by performing image multioutput classification
- 3-band, 8-bit satellite data chips of size 256x256 pixels, being parts of Planet's 'visual product' created based on imagery from ISS and Flock 2 satellite, characterized by GSD of app. 3.7 m, saved in .jpg format [1]
- training data annotation .csv file [1]
data collected:
- over 'Amazon basin which includes Brazil, Peru, Uruguay, Colombia, Venezuela, Guyana, Bolivia, and Ecuador' [1]
- 'between January 1, 2016 and February 1, 2017' [1]
part 1: EDA and data pre-processing
- Downloading the source data using Kaggle API and unzipping it
- Analyzing the training annotation file
- Analyzing the training images
part 2: Baseline model
- Feature engineering
- Stratified training/validation data split
- Classification
- Accuracy assessment
part 3: PyTorch CNN models - in progress