Evaluation of statistical methods to estimate forest volume in a mediterranean region

The use of three estimation methods was investigated for mapping forest volume over a complex Mediterranean region (Tuscany, central Italy). The first two methods were based on the processing of satellite images, specifically a summer Landsat Thematic Mapper scene. From this scene, information about forest volume was extracted through a nonparametric approach [k-nearest neighbor (k-NN)] and by means of locally calibrated regressions. The last method considered, kriging, instead used only the spatial autocorrelation of tree volume relying on geostatistical principles. The experiments performed demonstrated that, at the original sampling density, the three methods produced nearly equivalent accuracies. This was no more the case when reducing the sampling density to various levels. Whereas, in fact, this reduction marginally affected the performances of the two remote-sensing-based methods, it dramatically degraded that of kriging. Additionally, the investigation showed how per-pixel estimates of error variance were obtainable also by k-NN and locally calibrated regression procedures, in analogy with the same property of kriging. Such estimated error variances were utilized to optimally integrate the outputs of the methods based on remotely sensed data and spatial autocorrelation. In all cases, the integrated estimation outperformed the single procedures. These results are relevant to develop an operational strategy for mapping forest attributes in complex Mediterranean areas