Spatial deaggregation of agricultural statistics using constrained cross entropy minimisation

Abstract

Every year, millions of hectares of forests disappears in the tropics, something that has impacts on both local and global scale. Locally, the forest loss impacts ecosystem services such as water, energy and food security. Globally, the tropical forest loss releases large amounts of carbon dioxide into the atmosphere, making it the second largest driving factor of climate change, after combustion of fossil fuels. In research aiming to catalogue the main driving factors behind deforestation linked to land usage, there is a need to spatially downscale agricultural statistics from large statistical reporting units into smaller, in order to increase the accuracy of the driver identification. This thesis will investigate how one feasibly can do this deaggregation. Specifically, we have used constrained cross entropy minimisation, a method which aims to minimise the difference between a target distribution and a prior assessment of the spatial distribution of land usage, while also taking certain limiting constraints into account. For our investigations, we have chosen to focus on the Brazilian state Mato Grosso, a region that has experienced deforestation due to the spread of agricultural land uses. The prior will be created from land cover maps, generated from satellite imagery. Here we evaluated different prior preprocessing methods, finding that a rescaling of the land cover classifications using the land cover map’s confusion matrix was the method with the most promising result, capturing the general shape of the true distribution while also having a reasonable area distribution. It was however also noted that this method has a difficulty, common when dealing with land cover classification, in separating certain similar land uses from one another. It is suspected that this is due to the quality of the prior, and it would be interesting to investigate further how adjustments to the prior creation would improve the results found in this thesis.