Two operative forest site class estimation methods utilizing satellite images have been developed for forest income taxation purposes. For this, two pixelwise classification methods and two post-processing methods for estimating forest site fertility are compared using different input data. The pixelwise methods are discriminant analysis, based on generalized squared distances, and logistic regression analysis. The results of pixelwise classifications are improved either with mode filtering within forest stands or assuming a Markov random field type dependence between pixels. The stand delineation is obtained by using ordinary segmentation techniques. Optionally, known stand boundaries given by the interpreter can be applied. The spectral values of images are corrected using a digital elevation model of the terrain. Some textural features are preliminary tested in classification. All methods are justified by using independent test data.
A test of the practical methods was carried out and a cost-benefit analysis computed. The estimated cost saving in site quality classification varies from 14% to 35% depending on the distribution of the site classes of the area. This means a saving of about 2.0–4.5 million FMK per year in site fertility classification for income taxation purposes. The cost savings would rise even to 60% if that version of the method were chosen where field checking is totally omitted. The classification accuracy at the forest holding level would still be similar to that of traditional method.
The PDF includes a summary in Finnish.
A new sampling design, the local pivotal method (LPM), was combined with the micro stand approach and compared with the traditional systematic sampling design for estimation of forest stand variables. The LPM uses the distance between units in an auxiliary space – in this case airborne laser scanning (ALS) data – to obtain a well-spread sample. Two sets of reference plots were acquired by the two sampling designs and used for imputing data to evaluation plots. The first set of reference plots, acquired by LPM, made up four imputation alternatives (varying number of reference plots) and the second set of reference plots, acquired by systematic sampling design, made up two alternatives (varying plot radius). The forest variables in these alternatives were estimated using the nonparametric method of most similar neighbor imputation, with the ALS data used as auxiliary data. The relative root mean square error (RelRMSE), stem diameter distribution error index and suboptimal loss were calculated for each alternative, but the results showed that neither sampling design, i.e. LPM vs. systematic, offered clear advantages over the other. It is likely that the obtained results were a consequence of the small evaluation dataset used in the study (n = 30). Nevertheless, the LPM sampling design combined with the micro stand approach showed potential for improvement and might be a competitive method when considering the cost efficiency.