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

Titre : Mapping of dry savannah tree species using object oriented classification and high resolution imagery in Serowe, Botswana

Auteur : Kimani Jacob Ndirangu

Etablissement de soutenance : International Institute for Geo-Information Science and Earth Observation (ITC)

Grade : Master of Science in Geo-Information Science and Earth Observation 2004

Résumé
Reliable and accurate classification of dry savannah tree species, essential for transpiration up scaling for water management, has become a major challenge in semi arid areas such as Botswana. The major hurdles identified are pixel based classifiers and low resolution remotely sensed data. Therefore, the main objectives of this study were to assess the ability of two object-oriented classification techniques, eCognition and Feature Analyst, in mapping tree species using different high resolutions of airborne multi-spectral images (30cm, 60cm and 1m) ; and to determine factors associated with the distribution of trees and other vegetation types ; among fire, grazing and soil. A comparison was also done between eCognition classifications of Pan-sharpened IKONOS and airborne data, both of 1m resolution. Kappa statistics was used to determine the accurate technique and the optimal resolution. For grazing, herd’s size per cattle post and grazing radius were used to create weighted buffers that were correlated with NDVI values. The same NDVI values were correlated with fire evidence point map created from presence or absence of fire data collected from sample plots. Soil, particle size variability of <125μ m fraction, which determines its water holding capacity was used. The fraction’s variability was tested in three depths of 0 to 10, 20 to 30 and 60 to 70 (all in cm). Kappa accuracy values for eCognition ranged from 0.79 to 0.94, averaging at 0.89, 0.86 and 0.87 for 30cm, 60cm and 1m resolution respectively. The highest value, in Feature Analyst technique was 0.74 with the lowest being 0.29 with averages at 0.73, 0.49 and 0.54 for 30cm, 60cm and 1m resolution respectively. Though a Z-test showed no significance difference between the two relatively high Kappa values from the techniques, eCognition was more reliable when other factors were considered. However, in both techniques, 30cm and 60cm resolutions, gave the highest and the lowest Kappa accuracy values respectively. Though a one-way-ANOVA gave a standard deviation of zero (0) value among the Kappa accuracy values for the three resolutions in eCognition, other factors considered, 30cm was the most accurate. IKONOS classification yielded an average of 0.54 Kappa accuracy value. The results revealed a number of factors that determined the accuracy and reliability of the classification. These included the season of data acquisition ; diversity of tree species to be mapped ; closeness of different species in their spectral and biophysical parameters ; tree configuration ; and image preprocessing techniques (e.g. Brovy transformation in case of IKONOS). While grazing intensity had an inverse relationship with NDVI values, fire occurrence had a poor one ; being a constant factor affecting the whole area equally. The fine fraction in the sandy soils had a low variability (± 2.74%) with an average of 17.6% of the soil. The extremes were 5.5% and 25%. The results indicated no variation of the fraction within the depths considered . Long and medium spatial variability were 38%, and 28.5% respectively with a 33.5% unexplained short variability. This fraction decreases quadratically in a north-east direction in the study area. The spatial distribution of vegetation is shown to be influenced by grazing intensity and fire ; with a possible influence from the variability in soil particle size

Version intégrale (ITC) -> http://www.itc.nl/library/papers_20...