Maps of standing timber volume provide valuable decision support for forest managers and have therefore been the subject of recent studies. For map production, field observations are commonly combined with area-wide remote sensing data in order to formulate prediction models, which are then applied over the entire inventory area. The accuracy of such maps has frequently been described by parameters such as the root mean square error of the prediction model. The aim of this study was to additionally address the accuracy of timber volume classes, which are used to better represent the map predictions. However, the use of constant class intervals neglects the possibility that the precision of the underlying prediction model may not be constant across the entire volume range, resulting in pronounced gradients between class accuracies. This study proposes an optimization technique that automatically identifies a classification scheme which accounts for the properties of the underlying model and the implied properties of the remote sensing support information. We demonstrate the approach in a mountainous study site in Eastern Switzerland covering a forest area of 2000 hectares using a multiple linear regression model approach. A LiDAR-based canopy height model (CHM) provided the auxiliary information; timber volume observations from the latest forest inventory were used for model calibration and map validation. The coefficient of determination (R2
= 0.64) and the cross-validated root mean square error (RMSECV
= 123.79 m3
) were only slightly smaller than those of studies in less steep and heterogeneous landscapes. For a large set of pre-defined number of classes, the optimization model successfully identified those classification schemes that achieved the highest possible accuracies for each class.