Informations et ressources scientifiques
sur le développement des zones arides et semi-arides

Accueil du site → Master → Pays Bas → 2021 → Land Use and Land Cover Changes in Monduli District, Tanzania : Analysis of multiple classification methods and satellite sensors in order to perform a multi-temporal post-classification change detection analysis in a difficult to map semi-arid savannah landscape

Utrecht University (2021)

Land Use and Land Cover Changes in Monduli District, Tanzania : Analysis of multiple classification methods and satellite sensors in order to perform a multi-temporal post-classification change detection analysis in a difficult to map semi-arid savannah landscape

Rosmalen, R.C. van

Titre : Land Use and Land Cover Changes in Monduli District, Tanzania : Analysis of multiple classification methods and satellite sensors in order to perform a multi-temporal post-classification change detection analysis in a difficult to map semi-arid savannah landscape

Auteur : Rosmalen, R.C. van

Université de soutenance : Utrecht University

Grade : Master Thesis 2021

Résumé
In this study Land Use and Cover changes in Monduli district, Tanzania, are analysed from 2019 to 1985 with the use of a post classification change detection technique. In Monduli there are several aspects that complicate the classification of the landscape with remote sensing : A semi-arid climate with wet and dry seasons, similarity of spectral signatures of savannah vegetation types, fuzzy transition zones and a small heterogenous agricultural system. As a result of the difficulty of the study area for remote sensing previously conducted studies were unable to map the surface correctly (van den Bergh, 2016 ; Verhoeve, 2019). In this study measurements have been taken to combat the aforementioned issues and improve the accuracy of the classifications : the number of inputs for the classifier from the ground truth dataset has been increased, the classifiers are trained on each image separately, accuracies have been calculated for each classification, ancillary data and indexes are added and Sentinel 2 imagery (10m spatial resolution) has been incorporated, next to Landsat imagery (30m spatial resolution). Unsupervised ISODATA and supervised maximum likelihood and random forest classification methods have been applied. Sentinel did not result in higher accuracies because of the lower number of spectral bands available. However, of the random forest classifications with Landsat imagery four classifications reached an overall accuracy higher than 0.746 and were used for the change detection analysis. From 1985 to 2019 classes that increased are agriculture (+2.5%), built environment (+0.3%) water (+0.4%) and barren (+1.0%). However, barren shows some fluctuation over the years. As a result of the increase of these classes vegetation (woody savannah, savannah, open shrubland, closed shrubland and grassland) decreased with 2.7%. Additionally, a decrease in forests of 0.4% can be observed which is, next to cloud cover, primarily the result of an increase in woody savannah followed by closed shrub and agriculture.

Présentation

Version intégrale (28 Mb)

Page publiée le 31 mars 2022