Code for training a NN for Sentinel-2 and Google images on nightlights (poverty proxy).

We train CNN to predict nightlights intensity from Google Maps and Sentinel-2 satellite images. The trained models are used for extracting features in the HRM project. The work is based on transfer-learning papers from N Jean et al. and A Head et al..

The process to obtain these models is done in 4 steps:

• get the nightlights intensity as a raster. We use Google Earth Engine to get nightlights layer from NOAA - VIIRS and mask it with the JRC - GHS settleemnt layer. The script can be found here.
• bin the nightlights into 3 classes: high intensity, medium intensity and low intensity. The first 2 steps can be done using the code in src/nightlights_prep.py.
• pull the images from Google Maps static API and Sentinel-2. The code that handles this step is master_getdata.py.
• train the models using train.py, GPU may be necessary, we use AWS ec p2 instance. We trained for every datasource both a simple CNN with 3 conv and 1 dense hidden layers and a VGG16 with block 5 and dense layer trainable. The models are in (https://github.com/WFP-VAM/HRM_NN_Training/blob/master/src/models.py).