We focus on the simultaneous use of several sources of information for the evaluation and benchmarking of land surface models with special attention to the multi-scale temporal structure of both observed and model computed signals. We explore the usefulness and diagnostic power of several quantitative measures of similarity based on single- and multi-variable traditional statistics, statistical learning theory, and set theory. These measures are contrasted against traditional single stream-valued measures computed at different levels of temporal/spatial aggregation and/or frequency to show their increased diagnosing power. Three land surface models are used in the study (BATS, SiB3 and Noah) to provide a representation of different conceptualizations and modeling methods in semi-arid locations to highlight issues related to distinct seasonal behavior. The uncertainty associated with model parameters and forcing errors is intrinsically considered in the evaluation procedures. We provide suggestions for model benchmarking/evaluation that are consistent with the standards used in systems theory and computer science.