473
A common application of fusion is an increase of spatial or spectral resolution,
either accounting for limitations in the available data or imitating high-cost systems
using low-cost alternatives, e.g., in precision agriculture. For example, Zeng et al.
(2017) developed a system imitating very high spatial resolution hyperspectral
measurements such as those required for the calculation of some vegetation indices
(VI) using low-cost UAV-mounted sensors and fusing multispectral imagery with
spectrometer data using Bayesian imputation and principal component analysis.
Data fusion can also be applied when linking RS data to in-situ data. In their recent
review, Lesiv et al. (2016) compared different algorithms fusing RS with crowdsourced data for forest cover mapping, including geographically weighted logistic
regression (GWR), naïve Bayes, nearest neighbor, logistic regression, and classification and regression trees (CART), finding GWR to perform slightly better where
input data were disparate.
In its simplest form, fusion can be a basic overlay of high- (spectral/spatial/temporal) resolution data over low-resolution data. However, as Lesiv et al. (2016) and
others have shown, it is worth comparing different fusion techniques. Several studies have performed such comparisons but mainly with respect to land cover classification and specific to certain sensors and spatiotemporal scales (e.g., Caruana and
Niculescu-Mizil 2006, Clinton et al. 2015). Consequently, Liu et al. (2017) recommend routine use of statistical comparisons between different fusion techniques
(e.g., a Wilcoxon signed-ranks test for two algorithms or a Friedman test with Iman
and Davenport extension if more than five algorithms are compared) to detect the
optimal solution for a given application.
17.5.2 Assimilation
In essence, data assimilation is an extension of data fusion, linking noisy RS measurements with the outputs from imperfect numerical models to optimize estimates
of measures that are not directly observable from RS [e.g., for detailed, high spatiotemporal drought monitoring (Ahmadalipour et al. 2017) or, in the example given in
Fig. 17.3, to derive biomass estimates using a combination of canopy height, digital
terrain, and aboveground biomass (AGB) modeling]. Advantages of data assimilation include enhanced quality control, the ability to take into account errors and
uncertainties in data and models simultaneously, gap filling in data-poor locations
and where insufficient temporal information is available, and improved parameter
estimation in models.
However, data assimilation can also result in circular and inconsistent analyses.
The end user needs to be aware that many remotely sensed variables [e.g., leaf area
index (LAI)] are based on models incorporating ancillary information and are thus
not independently retrieved. In our example, land cover might already be used as an
information layer to tune the RS LAI retrieval within the data assimilation step.
Thus, using LAI as biodiversity variable and adding land cover as an explanatory
variable could be problematic (inconsistent if from different sources or circular).
17 Integrating Biodiversity, Remote Sensing, and Auxiliary Information for the Study…
A common application of fusion is an increase of spatial or spectral resolution,
either accounting for limitations in the available data or imitating high-cost systems
using low-cost alternatives, e.g., in precision agriculture. For example, Zeng et al.
(2017) developed a system imitating very high spatial resolution hyperspectral
measurements such as those required for the calculation of some vegetation indices
(VI) using low-cost UAV-mounted sensors and fusing multispectral imagery with
spectrometer data using Bayesian imputation and principal component analysis.
Data fusion can also be applied when linking RS data to in-situ data. In their recent
review, Lesiv et al. (2016) compared different algorithms fusing RS with crowdsourced data for forest cover mapping, including geographically weighted logistic
regression (GWR), naïve Bayes, nearest neighbor, logistic regression, and classification and regression trees (CART), finding GWR to perform slightly better where
input data were disparate.
In its simplest form, fusion can be a basic overlay of high- (spectral/spatial/temporal) resolution data over low-resolution data. However, as Lesiv et al. (2016) and
others have shown, it is worth comparing different fusion techniques. Several studies have performed such comparisons but mainly with respect to land cover classification and specific to certain sensors and spatiotemporal scales (e.g., Caruana and
Niculescu-Mizil 2006, Clinton et al. 2015). Consequently, Liu et al. (2017) recommend routine use of statistical comparisons between different fusion techniques
(e.g., a Wilcoxon signed-ranks test for two algorithms or a Friedman test with Iman
and Davenport extension if more than five algorithms are compared) to detect the
optimal solution for a given application.
17.5.2 Assimilation
In essence, data assimilation is an extension of data fusion, linking noisy RS measurements with the outputs from imperfect numerical models to optimize estimates
of measures that are not directly observable from RS [e.g., for detailed, high spatiotemporal drought monitoring (Ahmadalipour et al. 2017) or, in the example given in
Fig. 17.3, to derive biomass estimates using a combination of canopy height, digital
terrain, and aboveground biomass (AGB) modeling]. Advantages of data assimilation include enhanced quality control, the ability to take into account errors and
uncertainties in data and models simultaneously, gap filling in data-poor locations
and where insufficient temporal information is available, and improved parameter
estimation in models.
However, data assimilation can also result in circular and inconsistent analyses.
The end user needs to be aware that many remotely sensed variables [e.g., leaf area
index (LAI)] are based on models incorporating ancillary information and are thus
not independently retrieved. In our example, land cover might already be used as an
information layer to tune the RS LAI retrieval within the data assimilation step.
Thus, using LAI as biodiversity variable and adding land cover as an explanatory
variable could be problematic (inconsistent if from different sources or circular).
17 Integrating Biodiversity, Remote Sensing, and Auxiliary Information for the Study…