Retrieving forest structure variables from Very High Resolution satellite images using an automatic method

Over the last decade, a growing number of Very High Resolution (VHR) remote sensing data from various spatial sensors has become available. They provide interesting information for forest inventory applications thanks to the strong relationship between forest spatial structure and image texture at stand level. The main goal of this study is to find the most relevant image features to describe forest structure from very high resolution satellite imagery. The emphasis is placed on the automatisation of this process, exploiting both spectral and spatial information. Compared to High Resolution images, VHR imagery provides a very rich textural information that has to be thoroughly investigated for the forest structure characterization. Our approach is based on linear regressions between the forest structure variables to be estimated and various spectral and Haralick's texture features (derived from Grey Level Co-occurrence Matrix). The main drawback of this well-known texture representation is the underlying parameters (window size, displacement length and orientation, quantification level) which are extremely difficult to set due to the spatial complexity of the forest structure. To tackle this major issue, probably the main cause of poor texture analysis in practice, we propose an automatic feature selection process whose originality lies on the use of test frames of adequate forest samples where the structure variables were measured from ground. This method, inspired by camera calibration protocols, selects the best image descriptors via statistical modelling, exploring a wide range of parameter values. Hence, just a few samples are required to build up the test frames but allow a fast assessment of thousands of descriptors, given the large number of tested combinations of parameters values. This method has been successfully applied to the modelling of 7 typical forest structure variables (age, tree height, crown diameter, diameter at breast hight, basal area, density and tree spacing) over maritime pine stands in South West of France from QuickBird panchromatic and multi-spectral images. Overall results are good, multi-spectral and panchromatic images show similar performances to provide well-suited features to estimate the forest variables. The coefficient of determination, R², of the best models for 6 of the forest variables of interest, estimated from the test frames, ranges from 0.89 to 0.97. Only the basal area was weakly correlated to the considered image features (0.64). To improve the results, various linear combinations of panchromatic descriptors or multi-spectral descriptors or a combination of both, were tested using multiple linear regressions. As collinearity is a very perturbing problem in multi-linear regression, this issue is carefully addressed. Different variables subset selection methods are tested. A stepwise method, derived from LARS (Least Angular Regression), turned out the most convincing, significantly improving the quality of estimation for all the forest structure variables, including the basal area (R²>0.98). The best estimation results are obtained from subsets combining multi-spectral and panchromatic features, highlighting the potential of a multi-scale approach for retrieving forest structure variables from VHR satellite images.