Super typhoon Haiyan makes landfall
Super typhoon Haiyan
Subscribe here to receive special images

VEGETATION - 2000

Lake Maggiore - Italy, 3-6 April 2000


Fire patches in natural vegetation in southern Africa

Swinnen E. *, Verwimp R. **, Gulinck H. ***
* KULeuven / VITO (since 15/9/1999)
** Ground for GIS, KULeuven
*** KULeuven

Paper (pdf file, 223 k)

The objective of this study is to explore the feasibility of SPOT-VEGETATION for burnt surface mapping in savanna ecosystems. The study site is located in the Chobe area, northern Botswana. It is mainly covered by a mosaic of open and dense woody savanna and woodland. A time series with an interval of 2 weeks of VEGETATION images over one dry season (1998) is used for the signature analysis and the change monitoring. End member analysis is applied on a single date image with aid of a high resolution SPOT-XS image of even date and classified by means of field survey recordings. The spectral behaviour of burnt and unburnt savanna is examined in the RED, NIR and SWIR bands on a single image and for a time series. The reflectance of a burnt surface is lower than for savanna: a decrease in reflectance of 25-30% in NIR and 15-20% in SWIR was observed immediately after burning. The signal variability is largest in SWIR compared to NIR and RED. An increase of variability is observed in all examined bands immediately after a fire occurred. NIR has the most discriminative power. The change detection analysis focussed on mapping fire scars (change) and savanna (no change). Two change algorithms are employed to extract the change information: standardised differencing and 2-dimensional principal component analysis (PCA) (time1 vs time2). A statistical threshold is applied to classify the output of these algorithms to a burnt/non-burnt information layer. Accuracy is assessed using the field work recordings. Both change algorithms are successfully applied. All kappa coefficients are higher than 80%. No significant difference can be proved between standardised differencing and PCA. Although the slower decrease of the SWIR band after the burn-event the change monitoring process performed no significant better results for NIR. For areal extent measurement, two sub-pixel classifiers are examined, linear spectral unmixing (LSU) and artificial neural networks (ANN). The former is based on typical class signatures, but omitting the spectral heterogeneity within one class for a given band, whereas the latter takes this heterogeneity into account by training the network with examples. Different band combinations are explored. Results are evaluated by comparing the areal estimates of the classes, the rank correlation coefficient with the reference high resolution classification and the distributions of the errors. Both methods give satisfactory results, although ANN performs slightly better. Band combinations including NIR always yield better results. When SWIR is added, the results are less accurate for LSU, because of the large variability of the reflectance of both classes in this spectrum. The best combination is RED-NIR for both techniques. Rank correlations obtained range between .85 and .90 for LSU and exceed .90 for ANN. Largest estimation errors are logically found for the mixed pixels. The result of ANN contains more small errors than LSU (resp. 75% and 35% of the pixels with an error <5%). VEGETATION images prove to be an excellent tool for monitoring fire scars and complementary themes as flood monitoring.