Estimating the Average Age and Height of the Trees using SPOT-5 Panchromatic and Multi-spectral Image Fusion

Sustainable management of the forests requires satellite data at a large scale. This research aims to exploit pixel-based image fusion methods including principal component analysis (PCA) transformation, wavelet transformation, PCA/Wavelet transformation to improve the estimation accuracy of the mean height and age of a Pinus radiata plantation using SPOT-5 panchromatic and multi-spectral images at segment level. Therefore, the average height and age of the trees is measured within 61 plots in a Pinus radiata plantation in NSW, Australia. After applying pre-processing on the images, the spectral information including reflectance and vegetation indices along with textural information derived from gray level co-occurrence matrix for four window sizes and orientations are extracted from multispectral and panchromatic images, respectively. The same information is extracted from the fused images. It is shown that the textural information derived from the fused images performs more efficient than the textural information derived from panchromatic images to estimate average height and age. The results indicate that the models derived from PCA/Wavelet-based fused images with estimation error of 16% and 11% for age and height, respectively, perform better than the models derived from the data extracted from the images fused by the other fusion methods.

Extended Abstract

1-Introduction

Efficient management of the forests, especially those which are commercially important, requires update information about the structural parameters. Having reliable estimation of structural parameters enables estimation of forest productivity, prediction and modeling of the forest tensions, and prediction of environmental problems of the forests. Recently, remote sensing techniques are used as a cost-effective alternative to estimate forest structural parameters. So far, different methods were proposed for image fusion at pixel-level and all these methods are based on transferring the spectral and textural information of the input images to the fused image with the minimum changes. There have been many studies which compared the performance of the pixel-level image fusion methods; however, there are not many studies on the assessment of the fused image for the estimation of forest structural parameters. Therefore, this study aims to assess the capability of SPOT-5 fused images to estimate the average age and average height of Pinus radiata trees at segment level. 

2-Materials and Methods

The study area is a 5000 ha Pinus radiata plantation in the vicinity of Batlow at New South Wales, Australia. In September 2008, 61 plots were randomly collected at the study area. During this campaign, height and diameter at breast height (DBH) were measured and collected for 978 trees in the plots. The information of the years when the trees were planted was available. Accordingly, the average age of the trees were calculated for each plot and consequently the calculated mean ages were extended to the segments. SPOT-5 panchromatic and multispectral data were used in this study. Dark object subtraction-3 (DOS-3) was applied on multispectral and panchromatic images to reduce the effect of the atmosphere. Then, multispectral and panchromatic data were fused using three pixel-level image fusion methods including principal component analysis (PCA)-based image fusion, wavelet-based image fusion, and wavelet-based PCA image fusion. Afterwards, the spectral and textural attributes were derived from the fused and original images and they were used to fit the models to estimate the average age and average height of the segments. In order to avoid multicollinearity and model overfitting, it was necessary to remove some of the attributes for which the inter-correlation is high. For this purpose, random forest feature selection (RFFS) method was used. After applying RFFS, the attributes which were remained by RFFS were used as input data for multiple-linear regression to develop the models for estimating mean age and mean heights of the trees at segment-level.    

3-Results and Discussion

According to the results, the attributes derived from the images fused by wavelet-based PCA method performed better than the others fusion methods to estimate average height and age of trees. The main reason for the good performance of the fused image derived from wavelet-based PCA method is that this method improves the spatial information of the multispectral images better than the other image fusion methods. Moreover, it changes spectral information of the fused images less than the other fusion methods, since only the approximation image derived from wavelet-transformed panchromatic data is replaced by the first principal component and the coefficients of this data is preserved. This partial replacement led to the less spectral distortion in the fused image compared to that derived from the other fusion methods. The models derived from textural attributes extracted from the fused images estimated average age and height with higher accuracy compared to the models derived from the textural attributes extracted from the panchromatic image. The type of spectral data used for calculating textural information is very important. The panchromatic data are derived in a wide spectral band and consequently contain less spectral information compared to the fused multi-spectral data with a spatial resolution of 2.5 m. In the other words, the fused multispectral data inherits spectral data and structural data from multispectral images and panchromatic image, respectively.

4-Conclusion

In this research, three pixel-level image fusion methods were compared to identify which result is more appropriate fused image to estimate mean height and age of Pinus radiata plantation. It was shown that the models derived from spectral and textural information of the images fused by wavelet-based PCA method performed better than the other methods to estimate the average height and age. There is not any significant difference among the performances of the models derived from textural information of the images fused by different methods. Moreover, the textural information derived from fused multispectral images, performed better than those derived from panchromatic data to estimate the average age and height of trees.