ischlo / ldm

Deep learning for land development

Intro

This repository holds a deep learning model built with pytorch.

Background

Land use patterns can be extremely complex, on the one hand depending on the natural features of the environment, such as rivers, mountains etc… on the other reflecting how humans got to use a specific parcel of land other the years, decades and even centuries. This model is developed to predict a set of variables about a location, based on the knowledge of the variables in the neighborhood.

Data

We use as input a set of rasters, each modelling a feature of the environment (flood risk, population, slope etc…) covering the region of Oxfordshire.

Sampling

A sampling procedure was developed in order to create a training, validation and testing data sets. It consists in drawing from the whole grid an individual cell and its nearest neighbors. So for a set of $N_f$ features (or raster) over a grid of $N_x\times N_y$ cells, and for n-neighborhoods of cells, we get a pool of $(N_x-2n)\times(N_y-2n)\times N_f$ samples which are then at random assigned to the training, validation and testing sets at random with probability of $.7,.15,.15$ respectively.

Using it

It is in its current state.