Training temporal Convolution Neural Networks (CNNs) on satellite image time series for yield forecasting.

• Prerequisites
• Technologies
• Setup

To set up the environment:

git clone https://github.com/waldnerf/yieldCNN.git
cd yieldCNN
conda env create -f environment.yml
conda activate leanyf

If a GPU is available, then type in: pip install tensorflow-gpu==2.3.0

You can monitor GPU usage with the following command: watch -n 1 nvidia-smi

Below is the directory tree structure of the project. To run the project, one only needs the raw_data folder to be populated. Other folders are created automatically.

leanyf
├── data
│       ├── meta
│       └── params
│           └── Archi_{model_name}_{normalisation}_{input_size}
│                └── crop_{index}
│                    ├── month_{forecast}
│                        └── best_model
├── figures
└── raw_data

Aggregated results are stored in data/model_evaluation_{input_data}CNN.csv.

As a helper, certain variables have been hardcoded in mysrc/constants.py.

Scripts were named as explicitly as possible. Here is a short description of the main ones:

• preprocess_1D_inputs.py: Convert tabular data into arrays usable by a 1D CNN. Arguments are defined in mysrc/constants.py.
• preprocess_2D_inputs.py: Format tabular data into arrays usable by a 2D CNN and save them as pickles with the dill package. It is recommended to re-run this step on the machines where the code will run.
• optimise_so_1D_architectures.py: Full model optimisation and evaluation for 1D input data.
• optimise_so_2D_architectures.py: Full model optimisation and evaluation for 2D input data.
• optimise_so_2D_architectures_data_augmentation_data_from_Aug.py: Full model optimisation and evaluation for 2D input data with augmentation.
• optimise_so_p2D_architectures.py: Full probabilistic model optimisation and evaluation for 2D input data. Needs to be updated.
• launcher_1D.sh: Submit optimise_so_1D_architectures.py on AWS.
• launcher_2D.sh: Submit optimise_so_12D_architectures.py on AWS.

Models have been developed in the deeplearning folder:

• architecture_complexity_1D.py: Architecture definition of 1D models
• architecture_complexity_2D.py: Architecture definition of 1D models
• architecture_complexity_p2D.py: Architecture definition of 2D probabilistic models
• architecture_features.py: Defining tensorflow.keras architecture, and training the models

• merge preprocessing
• normalisation per province Normalization (2021-07-27) Two options now:

1. read raw data and normalize min max over the whole data set after train/test/val split
2. read normalized (by histo image) and normalize min max over the whole data set after train/test/val split (note that this latter norm has no effect because all data are already 0-1) Data generator delas with both norm and unorm because in any case normalise per image to add error (then norm back to orginal units) In option 2 each image gets the same scale of values (0-1). This means that for a given region, both a good year and a bad year (in terms of yield) will have some grid cells with 1 (the most represented profile). Another option can be tested:
3. read data normalized by region (all histos of a region) Note: after this option 3, the norm min max over the whole data set after train/test/val split should be turned off.

• Validation, Validation is now on one single year (no inner loop). We may consider increasing the number of years
• Trend, Trend data (in a way or another) could be passed after CNN

• Dr. Franz Waldner
• Dr. Michele Meroni