Forest complexity modelling and mapping with remote sensing and topographic data: a comparison of three methods

Abstract

The concept of forest complexity has been recently adopted to represent the multiple horizontal and vertical forest structure and composition attributes that support ecological functions in a single measure or index. The index variables are often selected based on known associations with types of habitat and biodiversity potential. In modelling and mapping of forest complexity using geospatial data, three multivariate methods have been evaluated in previous studies: (i) defining an additive complexity index by manual a priori selection and combination of a set of field variables followed by regression-based modelling of the index against geospatial data; (ii) as the previous method, but where a complexity index is defined a priori using principal components analysis (PCA) of the field data; and (iii) direct modelling of a set of field variables against a set of geospatial variables using techniques such as redundancy analysis (RDA). The objective of this study was to compare these methods through an assessment of model quality and their relative merits and limitations in implementation. In the rural municipality of Chelsea, Quebec, 70 field plots were established, and 24 forest structure and composition variables were measured that had between-variable correlations of less than 0.8. Spectral and spatial Quickbird imagery information and topographic data were used to derive complexity models using the three methods. The manual additive index and the RDA-derived index had similar validation errors relative to their index means (24.3% and 22.3%, respectively). However, the RDA index was based on the full set of 24 field variables, which had a total structure-composition variance greater than the manual additive index comprised of a subset of 10 of those variables. Thus, the RDA index was deemed to more comprehensively represent forest structure and composition than the additive index. RDA also produced outputs that were richer in information content, showing associations between individual variables and plots, as well as forest–environmental gradients. The PCA method was useful for evaluating environmental gradients in the field data, but dimensionality was too high to provide a single useful complexity index for modelling with geospatial data. The additive and RDA index models were used to produce maps of predicted forest complexity to aid in biodiversity survey planning and habitat conservation efforts within the municipality.