Stima e spazializzazione di attributi forestali tramite il metodo k-NN: distretto di Mocuba Mozambico
Author(s)
Raposo Pereira, Carla Maria
Date Issued
March 14, 2006
Type
Doctoral Thesis
Abstract
Forest inventory covering vast areas provides a high number of variables and information that need to be processed from plot level to the overall inventoried area. The increasing availability of remotely sensed data has made possible to obtain a synoptic view over large areas, although with less accuracy than the field measurements. The introduction of georeferenced sample plots in forest inventories created the possibility of combining field data to satellite images, that is multisource forest inventories, and to expand the information from plot level to the overall image.
In this study, forest variables were estimated by means of Landsat ETM images and existing forest inventory field plots data. The study was carried out for Mocuba district in Zambézia province – Mozambique, characterized by typical african miombo forests with predominance of Julbernardia globiflora and Brachystegia sp., covering an area of about 500 000 hectares. The non-parameteric k-nearest neighbour (k-NN) estimator was used for estimating and mapping of three selected forest variables: total volume, commercial volume and forest density. The method has been applied successfully in temperate and boreal forests but only a few studies regarding species richness prediction can be found in the literature for tropical and sub-tropical regions.
To estimate the selected forest variables, k-NN parameters were firstly calibrated to the district data in order to determine the most appropriated distance type and number of neighbours. Results obtained are encouraging and in line with those mentioned in several literature and studies, mostly for Nordic countries. When the k-NN estimator and the central point/pixel value was used to estimate forest variables high relative errors were obtained, from 76 to 99%. When the average values of the pixels within sample plots with 30 meters buffer area was employed to estimate the forest variables the relative error dropped to 47 - 63%. Errors obtained are far higher than those allowed for forest inventory (usually less than 20%) and are not accurate enough for forest management purposes at district and concessions levels. However, the mapping of forest variables with a known error is a powerful instrument for strategic decision making. Forest variables estimates derived by the method can easily produced and used in areas where no other information exists or for non-sampled areas.
Additional information
Dottorato di ricerca in Scienze e tecnologie per la gestione forestale e ambientale
File(s)![Thumbnail Image]()
Name
cmraposo_tesid.pdf
Size
2.79 MB
Format
Adobe PDF
Checksum (MD5)
d3b12e43b3262ae3191027b27bb784db