Combining Environmental, Multispectral, and LiDAR Data Improves Forest Type Classification: A Case Study on Mapping Cool Temperate Rainforests and Mixed Forests.

Predictive vegetation mapping is an essential tool for managing and conserving high conservation-value forests. Cool temperate rainforests (Rainforest) and cool temperate mixed forests (Mixed Forest, i.e., rainforest spp. overtopped by large remnant Eucalyptus trees) are threatened forest types in the Central Highlands of Victoria. Logging of these forest types is prohibited; however, the surrounding native Eucalyptus forests can be logged in some areas of the landscape. This requires accurate mapping and delineation of these vegetation types. In this study, we combine niche modelling, multispectral imagery, and LiDAR data to improve predictive vegetation mapping of these two threatened ecosystems in southeast Australia. We used a dataset of 1586 plots partitioned into four distinct forest types that occur in close proximity in the Central Highlands: Eucalyptus, Tree fern, Mixed Forest, and Rainforest. We calibrated our model on a training dataset and validated it on a spatially distinct testing dataset. To avoid overfitting, we used Bayesian regularized multinomial regression to relate predictors to our four forest types. We found that multispectral predictors were able to distinguish Rainforest from Eucalyptus forests due to differences in their spectral signatures. LiDAR-derived predictors were effective at discriminating Mixed Forest from Rainforest based on forest structure, particularly LiDAR predictors based on existing domain knowledge of the system. For example, the best predictor of Mixed Forest was the presence of Rainforest-type understorey overtopped by large Eucalyptus crowns, which is effectively aligned with the regulatory definition of Mixed Forest. Environmental predictors improved model performance marginally, but helped discriminate riparian forests from Rainforest. However, the best model for classifying forest types was the model that included all three classes of predictors (i.e., spectral, structural, and environmental). Using multiple data sources with differing strengths improved classification accuracy and successfully predicted the identity of 88% of the plots. Our study demonstrated that multi-source methods are important for capturing different properties of the data that discriminate ecosystems. In addition, the multi-source approach facilitated adding custom metrics based on domain knowledge which in turn improved the mapping of high conservation-value forest.