This is a project code for the paper: Predicting Urban Vitality from Satelite imagery.

• Background
• Install
• Structure
• Usage
  • Generator
• Badge
• Maintainers
• Contributing
• License

All the geospatial data in this project are reprojected to WGS 84 / UTM zone 32N EPSG:32632, as it is the default for Italy and it works in meter units.

In this project, we experimented with predicting both vitality for both imagelets and whole distritcts.

Requirements for this project include:

pandas
geopandas
scikit-learn 
xgboost

code

• satellite_imagery
• training_data
• prediction
• general
• bing

data

• boundaries
• satellite_imagery
• labels

preprocessed

• training_data

results

• XGBoost

1. Download satellite imagery around selected cities (Sentinel-2) and preprocess the imagery into .geotiff with EPSG:32632 projection. Then place the .geotiff under data/training_imagery/satellite/.

2. Download shapefiles with adminstrative borders of the selected cities and place the .shp files under data/boundaries/districts/. NOTE: We have found that it is best to download the whole regions from Italy in which the respective cities are and then extract each city's shape file. If you will do the same, the regional shapefiles are under data/boundaries/blocks/italy7s/ while the code to extract each city's shapefile is under code/general/extract_city_shapes.py. Upon running this code for each city, we have the respective shapefiles saved (in the proper projection) under data/boundaries/blocks/city. To obtain the city shapefiles with district borders that we need, the respective shapefles with block-level borders should be dissolved on the 'ACE' variable. The resulting shapefiles are under data/boundaries/districts/.

3. Download JJ labels from the repository of the paper The Death and Life of Great Italian Cities: A Mobile Phone Data Perspective and place under data/labels/.

1. Crop each satellite image to the area encompassing the city shapefile. This is done using code/satellite_imagery/crop_raster_using_vector.py and will save the cropped .geotiffs under data/satellite_imagery/crop/2A_source_raster/.
2. Produce training data (imagelets) by splitting the cropped satellite images into 64x64 px pieces. This is done in two steps: first, using code/satellite_imagery/split_raster.py, we create imagelets as .geotiffs, i.e., we preserve their spatial information, under data/satellite_imagery/2A_imagelets/. Second, we produce .rgb imagelets, i.e., those that are suitable for deep learning extractors, using code/satellite_imagery/jpeg_from_tiff.py -- this saves our results under data/satellite_imagery/2A_imagelets_jpgs/.

3. Produced .rgb imagery is parsed using VGG-16 extractor (TODO code/training_data/feature_extractor/) and resulting feature vectors are saved under preprocessed/features/imagelets. Since the size of feature from the extractor ranges from 2048 to 4096, we need to reduce them. The code using PCA to reduce this set of features to a predefined number of components (we experimented with number of features ranging from 8 to 64) is under code/training_data/features/imagelets/PCA_features.py. This saves the imagelet features under preprocessed/training_data/features/imagelets/.

4. Moreover, we extract geo-coordinates of the centroid for each imagelet. This is done using code/training_data/features/imagelets/extract_geocoordinate_features.py and the result is saved also under preprocessed/training_data/features/imagelets/.

5. To obtain district-level features from imagelet features, we first need to find a link between imagelets and districts. This is done using code/training_data/district_imagelets_link/district_labels_to_imagelets.py. The resulting imagelet district labels are found under preprocessed/training_data/district_imagelets_link. Using this mapping, the corresponding imagelet features are aggregated using the code under code/training_data/features/districts/aggregate_features_districts.py.

6. Moreover, in simiarity to geo-features for imagelets, we also extract centroid of each district using code/training_data/features/districts/extract_centroids_features.py and the results are saved under preprocessed/training_data/features/districts/.

7. We need labels for both classification and regression for each imagelet/district. For districts, this is a more straightforward process: each district takes its regression label from the original label data under data/labels/ (using code/training_data/labels/districts/regression_district_labels.py), while its classification label is defined based on the quartile/tertile its regression label takes among all the district labels (using code/training_data/labels/districts/classification_district_labels.py). The resulting labels are saved under preprocessed/training_data/labels/districts..

8. For imagelets, the situation is a bit more complex.

   Each imagelet will get the regression label (for each of the JJ variables) based on the district in which the imagelet is located. There are two ways to define whether the imagelet is located in a district. Due to the small distrcit sizes and the low resolution of the Sentinel imagery we work with, the imagelets will rarely fall completely within a single district. Hence, the two possible approaches are:

   1. imagelet is located within a district with which it overlaps with more than the 50% (THRESHOLD approach) of its area with the districts area, 2. imagelet is assigned to the district with which it has the highest overlap (MAXIMUM approach). This allows to label more imagelets, as the strict criteria in the first case omits many imegelets that overlapped with several districts and with none of them with more than 50% of its size.

   For classificaton labels, we merge the districts into classes (the quartiles for each variable) and then disolve the shapefile along such a defined class. We then look for each imagelet, in which area (distrct group) it is located, again having two possible defitions as above.

The steps to achieve these 4 possible types of labeling are:

1. produce shapefiles from imagelets using code/satellite_imagery/polygonize_raster.py. This saves a new set of files under data/satellite_imagery/2A_imagelet_shapes/. We do this since it is easier to calculate the overlap with district sizes in the next step using such shapefiles than rasters of imagelets. 2) For a finer granularity labelling, we use code/district_data/training_data_assign_labels_districts.py, which assigns each imagelet('s shape) to the district. The LABELING_METHOD in this script can be set to "threshold" or "maximum" depedning on which of the two above discussed approaches to labelling we take. The "threshold" method is more strict. This script saves files with imagelet names and their district labels under preprocessed/district_labels/.

1. Predict imagelet/district class for each of the JJ labels, including the vitality, either directly or indirectly.

@sanja7s