Geoinformation, monitoring and assessment Environment

Updated December 1998

Remote Sensing Centre Series

Use of NOAA remote sensing data for assessment of the forest area of Liberia

By H.-J. Stibig, R. Baltaxe
21 pp, 12 figures
RSC Series No. 66, FAO, Rome 1993.


Summary

In the context of developing a practicable and cost effective method for obtaining a country's forest area by remote sensing, the computer processing of NOAA AVHRR HRPT data covering Liberia was investigated. The only cloud-free scene then recorded turned out to be severely and unevenly affected by atmospheric haze. To mitigate the effects of this, the country was divided into six areas (strata) of more uniform haze conditions. Pixel DN values were obtained for forest and adjoining formations on transects within each stratum, for the first four AVHRR channels and three transforms: NDVI (2-1/2 + 1), IND3 (3-2/3+2), IND4 (4-2/4+2). After analysing the transects and comparing them with the available reference data (a mixture of large scale colour composite Landsat TM and MSS images for 1989 and 1986 respectively), channels 2,3,4 and IND3 were retained for processing. This was done by applying three methods to each stratum - thresholding, maximum likelihood classification using clustering signatures (hybrid), m.1.c. using training area signatures - and directed at separating the 5 main classes distinguishable on the Landsat images: Closed forest, Disturbed forest, Shifting cultivation and regrowth, Cultivation and Other.

Mainly due to the haze, the separation of Forest and Disturbed forest within some of the strata was not consistent for the three methods. The data were, therefore, processed to separate only Forest from the other classes. Results were evaluated by analysis of their internal consistency and comparison with the Landsat images. The difference between the three methods, for Forest area as percentage of total stratum area, ranged between 3.2% and 8.4%. The Forest area for the whole of Liberia obtained by thresholding and by the hybrid classification differed by only 6 sq km. However, with the third method the difference amounted to 1922 sq km, or 4.1% of the average of the largest and smaller forest area obtained by the three methods - 46,525 sq km. Agreement between the results and the TM images was judged to be generally good (the MSS images were extremely poor), although lack of ground truth and other factors militated against obtaining forest area from the images. The same scene was available in the form of 4 km GAC data and the same six strata as used for the HRPT data were thresholded with IND3 to separate the same classes, with satisfactory results.

Thresholding was found to be the most practical and directly interactive method. The IND3 transform displayed the highest sensitivity to variations of vegetation cover apparent on the Landsat images and lent itself well to thresholding. Although no quantitative assessment of the results was possible, the evaluation which could be made was taken to indicate that even under the unfavourable conditions of this investigation, and for the purposes of global monitoring, the HRPT data provided an acceptably reliable figure for one class of Forest over the country as a whole. The same appeared to be true for GAC data, though with lower accuracy as was to be expected. It was also maintained that, under more favourable conditions, more than one class of Forest could be reliably mapped for Liberia by the method favoured here.


The above publication is available from:
Chief, FAO/SDRN
Viale delle Terme di Caracalla
00100 Rome, Italy
(e-mail: changchui.he@fao.org)



Back to Top FAO Homepage