SPOT-5 Spectral and Textural Data Fusion for Forest Mean Age and Height Estimation

Precise estimation of the forest structural parameters supports decision makers for sustainable management of the forests. Moreover, timber volume estimation and consequently the economic value of a forest can be derived based on the structural parameter quantization. Mean age and height of the trees are two important parameters for estimating the productivity of the plantations. This research aims to estimate mean height and age of a Pinus radiata plantation using SPOT-5 textural and spectral data derived from multi-spectral and panchromatic images, respectively. The study site for this research consisted of a 5000 ha Pinus radiata plantation from 35◦ 23/ 35// S to 35◦ 29/ 58// latitude, and 147◦ 58/ 48// E to 148◦ 04/ 02// E longitude, near the town of Batlow in the Hume Forestry Region, NSW Australia. Tree age classes ranged from 10-20 years, 21-30 years and more than 30 years. A total of 63 plots with radii varied from 7 m to 20 m to ensure there were a minimum of 15 trees per plot in variable stocking classes, were randomly surveyed in June and July 2008 by the New South Wales Department of Industry and Investment (IINSW) and Forests New South Wales (FNSW). The effects of forest boundaries, road edges and the other irrelevant features on the estimations were eliminated through buffering prior to locating the plot centres. Tree heights and DBH were measured for 978 trees. Tree heights were measured twice using Vertex hypsometer to increase the accuracy of the measurement. After calculating mean height of the plots, a regression equation was derived based on the relationship between tree age and mean height of the plots. Using this equation and available age of each segment, the mean height was calculated for 278 segments of this plantation. Spectral data includes band reflectance, vegetation indices, and principal component analysis were derived using multispectral image for each segment. Gray level co-occurrence matrix was calculated in four different angles and window sizes to extract the textural data from panchromatic image. Prior to modeling, random forest feature selection method was applied on the spectral and textural data, individually and together to determine the most important features for estimating mean height and age at segment level. Multiple-linear regression (MLR) was applied to model mean height and age using textural and spectral data. Also, spectral and textural attributes were fused at feature level using two approaches. In the first approach, the spectral and textural attributes were used together as inputs for MLR. In the second approach, ratio of spectral and textural attributes were calculated for feeding MLR. The results indicated that there is not significant difference between the models derived from spectral attributes of multispectral data and those derived from textural attributes of panchromatic data. Moreover, it was shown that the models derived from the ratio of the spectral and textural data with age estimation error of 17% and height estimation error of 13% performed better than the other models.