Scale challenges in inventory of forests aided by remote sensing

The impact of changing the scale of observation on information derived from forest inventories is the basis of scale-related research in forest inventory and analysis (FIA). Interactions between the scale of observation and observed heterogeneity in studied variables highlight a dependence on scale that affects measurements, estimates, and relationships between inventory data from terrestrial and remote sensing surveys. This doctoral research defines "scale" as the divisions of continuous space over which measurements are made, or hierarchies of discrete units of study/analysis in space. Therefore, the "scale of observation" (also known as support) refers to that integral of space over which statistics are computed and forest inventory variables regionalized. Given the ubiquitous nature of scale issues, a case study approach was undertaken in this research (Articles I-IV) with the goal to provide fundamental understanding of responses to the scale of observation for specific FIA variables. The studied forest inventory variables are; forest stand structural heterogeneity, forest cover proportion and tree species identities. Forest cover proportion (or simply forest area) and tree species are traditional and fundamental forest inventory variables commonly assessed over large areas using both terrestrial samples and remote sensing data whereas, forest stand structural heterogeneity is a contemporary FIA variable that is increasingly demanded in multi-resource inventories to inform management and conservation efforts as it is linked to biodiversity, productivity, ecosystem functioning and productivity, and used as auxiliary data in forest inventory. This research has two overall aims: 1. To improve the understanding of the association between the scale of observation and observed heterogeneity in inventory of forest stand structural heterogeneity, forest-cover proportions, and identification of tree species from a combination of terrestrial samples and remote sensing data. 2. To contribute knowledge to the estimation of scale-dependence in inventory of forest stand structural heterogeneity, forest-cover proportions, and identification of tree species from a combination of terrestrial samples and remote sensing data. Different scales of observation were considered across the four case studies encompassing individual leaf, crown-part or branch, single-tree crown, forest stand, landscape and global levels of analysis. Terrestrial and remote sensing data sets from a variety of temperate forests in Germany and France were utilized across case studies. In cases where no inventory data were available, synthetic data was simulated at different scales of observation. Heterogeneity in FIA variable estimates was monitored across scales of observation using estimators of variance and associated precision. As too much heterogeneity is hardly interpreted due to a low signal to noise ratio, object-based image analysis (OBIA) methods were used to manage heterogeneity in high resolution remote sensing data before evaluating scale dependence or scaling across observed scales. Similarly, ensemble classification techniques were applied to address methodological heterogeneity across classifiers in a case study on classification of two physically and spectrally similar Pinus species. Across case studies, a dependence on the scale of observation was determined by linking estimates of heterogeneity to their respective scales of observation using linear regression and a combination of geo-statistics and Monte-Carlo approaches. In order to address scale-dependence, thresholds to scale domains were identified so as to enable efficient observation of studied FIA variables and scaling approaches proposed to bridge observations across scales. For scaling, this research evaluated the potential of different regression techniques to map forest stand structural heterogeneity and tree species wall-to-wall from remote sensing data. In addition, radiative transfer modelling was evaluated in the transfer between leaf and crown hyperspectra, and a global sampling grid framework proposed to efficiently link different stages of survey sampling. This research shows that the scale of observation affected all studied FIA variables albeit to varying degrees, conditioned on the spatial structure and aggregation properties of the assessed FIA variable (i.e. whether the variable is extensive, intensive or scale-specific) and the method used in aggregation on support (e.g. mean, variance, quantile etc.). The scale of observation affected measurements or estimates of the studied FIA variables as well as relationships between spatially structured FIA variables. The scale of observation determined observed heterogeneity in FIA variables, affected parameter retrieval from radiative transfer models, and affected variable selection and performance of models linking terrestrial and remote sensing data. On the other hand, this research shows that it is possible to determine domains of scale dependence within which to efficiently observe the studied FIA variables and to bridge between scales of observation using various scaling methods. The findings of this doctoral research are relevant for the general understanding of scale issues in FIA. Research in Article I, for example, informs optimization of plot sizes for efficient inventory and mapping of forest structural heterogeneity, as well as for the design of natural resource inventories. Similarly, research in Article II is applicable in large area forest (or general land) cover monitoring from sampling by both visual interpretation of high resolution remote sensing imagery and terrestrial surveys. This research is also useful to determine observation design for efficient inventory of land cover. Research in Article III contributes in many contexts of remote sensing assisted inventory of forests especially in management and conservation planning, pest and diseases control and in the estimation of biomass. Lastly, research in Article IV highlights scale-related effects in passive optical remote sensing of forests currently understudied and can ultimately contribute to sensor calibration and modelling approaches.