Youcef Kacer

Machine Learning & Computer Vision engineer

Human Density Prediction

This project aims at predicting human density combining Landsat-8 satellite images and Machine Learning.

Landsat-8 satellites images

Images are taken directly from U.S. Geological Survey website. One can query images with criteria like :

• cloud covering
• date range
• day or night

Here after, a screenshot of France query

and the 70 resulting datasets (thumbnails) projected, queried with minimum possible cloud covering (<20%), day acquisition and between May to September 2013.

Vegetal indice extraction

Each dataset is composed of 11 bands (1 pixel = 30x30 meters). NDVI (Normalized Difference Vegetal Indice) is a combination of band 4 (red) and band 5 (red-edge) that make vegetation in evidence according to (values span from -1 and 1) :

• NDVI < 0 : water, snow, ice, cloud
• NDVI = 0 : ground without vegetation
• NDVI > 0 : ground with vegetation

We then expect high values (rich vegetation) to explain low density areas, while low values (poor vegetation) should explain high densities. Below, RGB and NDVI images for different densities:

Paris - 21153 habs/km²

Lyon - 10117 habs/km²

Amiens - 2698 habs/km²

Abbeville- 914 habs/km²

Histogram of NDVI (1024 bins) is then finally taken as city descriptor. Superposition of various histograms shows clearly a relation between NDVI behavior and density :

Importing Labels

Precise densities can be taken from French official institute census (INSEE). This census is from 2013, which perfectly matches our Landsat-8 datasets. We categorize densities in order to deal with a classification problems (6 categories)

• category 1 : density between 0 and 500 habs/km²
• category 2 : density between 500 and 2000 habs/km²
• category 3 : density between 2000 and 5000 habs/km²
• category 4 : density between 5000 and 10000 habs/km²
• category 5 : density between 10000 and 13000 habs/km²
• category 6 : density greater than 13000 habs/km²

Machine Learning

We can now go through supervised learning. France data has been used for training/validation. Belgium, Netherlands and Switzerland have been used for testing model generalization.

Neural Network - Testing - Belgium

ground truth

prediction

Neural Network - Testing - The Netherlands

ground truth

prediction

{:class=”img-responsive

Neural Network - Testing - Switzerland

ground truth

prediction

Code is avalaible in my github at following link. There is a complete tutorial on how to make a density prediction of any geographic area. Japan is taken as example.