475
agricultural landscapes (Haboudane et al. 2003). In its simplest form, validation is
a pixelwise comparison of presumably high accuracy (often in-situ) data with the
remotely sensed or modeled data, using an x-fold validation approach (splitting the
in-situ data into training and validation data) if some of the in-situ data are needed
for model development or downscaling. For an overview of more complex techniques see, for example, Montesano et al. (2016) for Landsat-derived tree cover,
Lesiv et al. (2016) for crowdsourced forest cover, Joshi et al. (2016) for optical- and
radar-derived land use, and Sun et al. (2017) for in-situ validated land cover.
With the rapidly growing availability of RS and auxiliary data, validation can
become a time-consuming and complex task. Thus, increasingly, web-based validation systems are being developed that integrate big data access and storage, adjustment, and different intercomparison and validation techniques simultaneously (e.g.
Sun et al. 2017).
One of the potential issues with the data fusion and assimilation methods described
above is that they are often applied globally without testing whether variables and
correlations remain stable in space and time (Comber et al. 2012). This is a recognized
problem, and solutions have been proposed for over a decade (e.g., geographically
distributed correspondence matrices, Foody 2005) with new approaches being continually proposed. Often these are specific for certain applications, such as net primary
production (Wang et al. 2005), epidemiology (Khormi and Kumar 2011), biomass
(Propastin 2012) or population segregation (Yu and Wu 2013). One recently proposed
more generic approach is that of locally geographically weighted correspondence
matrices, which combine categorical difference measures (Pontius and Milones 2011,
Pontius and Santacruz 2014) with spatially distributed kappa coefficient, user, and
producer accuracy estimates—with code to run these tests in R being available, e.g.,
see packages gwxtab, differ and RSLcode (available on github (https://github.com/
lexcomber/RSLcode)) (Comber et al. 2017). All of these draw attention to the fact
that, even after performing data cleaning, fusion, assimilation, and validation steps,
local approaches, data, and techniques cannot necessarily be directly transferred from
one location and spatiotemporal resolution to another.
17.6 Conclusions
We are living in an increasingly data-rich world, in which not only more but also
more accurate and reliable data are available on many aspects related to plant biodiversity, both from remotely sensed as well as in-situ measurements. Increasingly,
limitations and potential circularities inherent to these data are acknowledged, often
aided by the provision of associated estimates of uncertainties and dedicated intercomparison studies. Techniques are being developed that enable even nonexperts to
account for and learn from these. Even in the case of data limitations, however, our
ability to map all aspects of biodiversity over large spatial and temporal scales has
increased exponentially over the last decade, and monitoring and understanding
ecosystem functions is easier than ever before. Nevertheless, significant challenges
17 Integrating Biodiversity, Remote Sensing, and Auxiliary Information for the Study…
agricultural landscapes (Haboudane et al. 2003). In its simplest form, validation is
a pixelwise comparison of presumably high accuracy (often in-situ) data with the
remotely sensed or modeled data, using an x-fold validation approach (splitting the
in-situ data into training and validation data) if some of the in-situ data are needed
for model development or downscaling. For an overview of more complex techniques see, for example, Montesano et al. (2016) for Landsat-derived tree cover,
Lesiv et al. (2016) for crowdsourced forest cover, Joshi et al. (2016) for optical- and
radar-derived land use, and Sun et al. (2017) for in-situ validated land cover.
With the rapidly growing availability of RS and auxiliary data, validation can
become a time-consuming and complex task. Thus, increasingly, web-based validation systems are being developed that integrate big data access and storage, adjustment, and different intercomparison and validation techniques simultaneously (e.g.
Sun et al. 2017).
One of the potential issues with the data fusion and assimilation methods described
above is that they are often applied globally without testing whether variables and
correlations remain stable in space and time (Comber et al. 2012). This is a recognized
problem, and solutions have been proposed for over a decade (e.g., geographically
distributed correspondence matrices, Foody 2005) with new approaches being continually proposed. Often these are specific for certain applications, such as net primary
production (Wang et al. 2005), epidemiology (Khormi and Kumar 2011), biomass
(Propastin 2012) or population segregation (Yu and Wu 2013). One recently proposed
more generic approach is that of locally geographically weighted correspondence
matrices, which combine categorical difference measures (Pontius and Milones 2011,
Pontius and Santacruz 2014) with spatially distributed kappa coefficient, user, and
producer accuracy estimates—with code to run these tests in R being available, e.g.,
see packages gwxtab, differ and RSLcode (available on github (https://github.com/
lexcomber/RSLcode)) (Comber et al. 2017). All of these draw attention to the fact
that, even after performing data cleaning, fusion, assimilation, and validation steps,
local approaches, data, and techniques cannot necessarily be directly transferred from
one location and spatiotemporal resolution to another.
17.6 Conclusions
We are living in an increasingly data-rich world, in which not only more but also
more accurate and reliable data are available on many aspects related to plant biodiversity, both from remotely sensed as well as in-situ measurements. Increasingly,
limitations and potential circularities inherent to these data are acknowledged, often
aided by the provision of associated estimates of uncertainties and dedicated intercomparison studies. Techniques are being developed that enable even nonexperts to
account for and learn from these. Even in the case of data limitations, however, our
ability to map all aspects of biodiversity over large spatial and temporal scales has
increased exponentially over the last decade, and monitoring and understanding
ecosystem functions is easier than ever before. Nevertheless, significant challenges
17 Integrating Biodiversity, Remote Sensing, and Auxiliary Information for the Study…