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

Accueil du site → Master → Finlande → Classification of agricultural crops of the Taita Hills, Kenya using airborne AisaEAGLE imaging spectroscopy data

University of Helsinki (2014)

Classification of agricultural crops of the Taita Hills, Kenya using airborne AisaEAGLE imaging spectroscopy data

Piiroinen, Rami

Titre : Classification of agricultural crops of the Taita Hills, Kenya using airborne AisaEAGLE imaging spectroscopy data

Auteur : Piiroinen, Rami

Université de soutenance : University of Helsinki,

Grade : Master 2014

Résumé
Land use practices are changing at a fast pace in the tropics. In sub-Saharan Africa forests, woodlands and bushlands are being transformed for agricultural use to produce food for the rapidly growing population. Although food production is crucial for the survivability of the people the uncontrolled expansion of agricultural land at the expanse of natural habitats may in the longer term decrease food production due to disturbances in water balance, increased land erosion and eradication of natural habitats for pollinators. Before the impacts of land use/land cover changes on the ecosystem can be studied the study area needs to be mapped. The study area of this thesis is located in the Taita Hills, Kenya. In previous studies the land use/land cover was mapped on higher hierarchical level in classes such as agricultural land, forest and bushland. In this thesis high spatial and spectral resolution AisaEAGLE imaging spectroscopy data was used to map the common agricultural crops found in the study area. Ground reference data was collected from 5 study plots located in the study area. Over 50 plant species were mapped but only 7 of these were used in the classification. The AisaEAGLE data was acquired in January–February of 2012 and was radiometrically, geometrically and atmospherically corrected. Minimum noise fraction (MNF) transformation was applied to the data to reduce the noise and the dimensionality. Optimal number of MNF bands was defined based on analysis of the information content of the bands. The classification was done with support vector machine (SVM) algorithm using radial basis function (RBF) kernel. Gamma, penalty and probability threshold parameters for the classifier were defined based on analysis of different combinations of these values. The analysis showed that gamma and penalty values had only minor impacts on the classification result. Based on the analysis an optimal threshold level was defined where pixels that were not likely to belong to any of the classes were left unclassified while maximum number of the known targets were correctly classified. Study area was classified with the optimal threshold value 0.90. Classification with threshold value 0.00 was done for reference. The overall accuracies for the classified pixels were 91.52% and 99.70% for the classifications done with probability threshold values 0.00 and 0.90. As the threshold was increased to 0.90 61% of the pixels were left unclassified. At the optimal threshold level between classes misclassifications were almost completely removed whereas the total number of correctly classified testing samples decreased. Applying MNF transformation to the data before the classification increased the overall accuracy from 80.58% to 91.52% while other parameters stayed the same. Results of this thesis showed that SVM classifier used with MNF transformation yielded high overall accuracies for the crop classifications. Adjusting the probability threshold to an optimal level was important since the study area was heterogeneous and only fraction the species were classified. For further applications the possibilities of object-based classification should be considered. The results of this thesis will be shared with the Climate Change Impacts on Ecosystem Services and Food Security in Eastern Africa (CHIESA) –project.

Présentation

Version intégrale (6,34 Mb)

Page publiée le 25 avril 2018