Source: EE Publishers, Article: PositionIT
The effects of emerging environmental problems associated with climate change and human-induced land cover change often directly impede livelihoods of the rural populace in Africa. Erratic rainfalls, flooding, forest cover loss, and land degradation are causing unstable agricultural yields and incomes. This article illustrates examples, possibilities and future perspectives from current earth observation (EO) research to address landscape dynamics in relation to emerging environmental issues in Africa.
Two experimental EO examples from eastern Africa are showcased; a multi-sensor approach to map vegetation productivity decline over eastern Africa, and an integrative (multi-data) approach to map the spatial distribution of flowering plants at a local to landscape scale. In example one, vegetation productivity, mapped at a regional scale using 250 m MODIS NDVI imagery, is related to very high resolution (VHR) imagery in Google Earth. The MODIS-based productivity declines could be effectively linked to land transformation processes (i.e. deforestation) using the VHR imagery. Moreover, using rainfall trends from passive radar observations, climate induced change could be largely disentangled from the human-induced change. The results of the second example are instigated for the quantification of pollination effects and to sustain healthy honey bee colonies. We suggest that integrative, that is multi-sensor and multi-scale, monitoring of flowering plants is possible. The up- or down-scaling and data fusion techniques such as layering or convolution techniques have to be probed separately for every site and application. The use of innovative, effective and fast field-based data collection methods such as smartphone “crowdsourcing‟ is of special relevance to data scarce environments. Smart data integration techniques are paramount in seamless and “use case” orientated monitoring of emerging environmental issues in Africa.
Fig. 1: Vegetation productivity decline mapping workflow using multiple datasets (V7 – version 7; MOD. – MODIS).
Increasing demand for natural resources and agricultural products in Africa often leads to unsustainable land use intensifications, conversions of natural areas to croplands [1] and land productivity losses [2]. Climate change often exacerbates human- induced change [3], and the distinction between human and climate induced change still poses a research challenge [4, 5]. These cumulative land changes affect the ability of the agro-ecological system to render important ecosystem services (ESS) such as products from forests for food and health, biodiversity for pollination and soil fertility for sustainable food production [4].
Earth Observation (EO) assessments, defined as the amalgamation of remote sensing observations, in situ measurements and spatial modeling or geographical information systems (GIS) datasets, can be used to cover larger areas and render a continuum of seamless observations to map land surface fluxes related to human-induced change and climate change [6, 7]. Moreover EO datasets have the potential to link qualitative and quantitative data, field and remote sensing and assimilate human and biophysical data in an integrative spatial working framework [8]. This would specifically mean using a multiple datasets at various spatial scales and linking these using sophisticated scaling, up- or down-scaling, or assimilation models. In scaling and assimilation models the coexistence of qualitative and quantitative datasets, and non-linear interactions of the features, variables and processes would have to be considered [9]. The challenge is the seamless integration and the degree of overlay of the various datasets. In this article we will not deal with the definition and complexity of scale and scaling dimensions; we will instead focus on linking different datasets for integrative mapping, and illustrate this by giving two examples.
Within EO concepts spatial datasets are utilised to assess the spatial and temporal heterogeneity of the land surface that is mostly a result of human and environmental interactions. The spatial structuring of the data is largely determined by autocorrelations; features or land classes that are in close spatial proximity to another can be assumed to exhibit similar behaviour or dynamics [8]. However, the agro-ecological landscape in Africa is highly fragmented and dynamic, largely due to small scale farming, high land use intensities and erratic rainfalls [10]. This complexity calls for an integrative that is multi-temporal and multi-scale mapping approach. The data assimilation model, specifically whether an up-scaling or down-scaling is preferred and which data fusion model is used, largely depends on the site (local) heterogeneity and its dynamics, and more profane data availability and costs [11].
In this study we showcase two innovative examples for integrative EO approaches and discuss their workflows and data assimilation techniques in the context of addressing emerging environmental issues in Africa. One of the example pertains to mapping human-induced land productivity decline. In the first example we use time-series optical satellite data on vegetation productivity, at 250 m geometric resolution, as well as time-series radar (satellite) data on rainfall at 25 km geometric resolution. In the other example airborne 0,6 m hyperspectral data is used to map the spatial distribution of flowering plants. We further propose to upscale the spectrally “pure” end members from the airborne (hyperspectral) imagery to 2 m resolution Worldview-2 satellite (multi-spectral) imagery. Continuous cross-verification is proposed using in situ collected smartphone data.
Fig. 2: Land productivity decline map (2000 to 2012) derived from MODIS NDVI and TRMM rainfall metrics.
Land productivity mapping over East Africa
Datasets and methodology
The workflow of combining optical MODIS (Moderate Resolution Imaging Spectroradiometer) NDVI (Normalised Differential Vegetation Index) time-series data with passive rainfall data from TRMM is shown in Fig. 1. Corrected 16-day composite images of NDVI from the MODIS (MOD13Q1 product, collection 5) [12] at spatial resolution of 250 m for the years 2000 to 2012 served as an input for the analysis. Monthly rainfall data from 2000 to 2012, at 25 km resolution, for the corresponding period, from the Tropical Rainfall Measuring Mission (TRMM) passive radar instrument (product 3B43) were acquired and geo-located to the MODIS data [6]. The geo-location entailed the downscaling of the TRMM rainfall data to MODIS vegetation productivity time-series data using nearest neighbourhood re-sampling. For both datasets per pixel trends as well as significances of these trends were computed [13]. The two datasets were then overlaid using straightforward layering and the corresponding per pixel negative rainfall trends from TRMM were masked in the MODIS data to exclude mapping areas where water scarcity can largely be assumed to be the driver of vegetation change.
Furthermore, we computed per pixel rain use efficiencies (RUE) using the ratio between monthly NDVI means, from optical MODIS data, and monthly rainfall means, from passive radar TRMM [14]. The RUE were computed to map severity levels of vegetation productivity decline for the observation period. The monthly RUE can be used to normalise the effects of rainfall variability in the vegetation productivity signal when interpreting vegetation productivity trends [14, 15] .
The MODIS-based productivity decline dataset map was linked to very high resolution data in Google Earth (“cross verification” in Fig. 1), since historic datasets are freely available for some sites to verify causes for the regional mapped vegetation decline.
Results and discussion
The experimental results of using the integrative mapping approach (Fig. 1) to map and characterise land productivity shows that overall accuracies of 82% for mapping deforestation can be achieved. The overall accuracy score was attained from comparing the 250 m MODIS- based land (vegetation) productivity decline data to very high resolution (bi-temporal) imagery in Google Earth. Land productivity “hot spots”, that is areas that exhibit severe productivity loss from the RUE computation (red areas in Fig. 2), were mostly found in southern Ethiopia, eastern Uganda and central Kenya. Fig. 2 shows a sub-section of the vegetation productivity map that covers Kenya and adjoining countries. Yellow areas depict pixels where vegetation productivity losses were “moderate”.
The land productivity decline areas in Fig. 2 were “rainfall normalised” using the independently collected passive radar observations from TRMM. By essentially linking the remote sensing vegetation productivity trends (i.e. NDVI data from MODIS) to concurrently available data metrics on rainfall from passive radar remote sensing observations (i.e. TRMM), integrative assessments of vegetation responses to rainfall patterns can be effectively performed (using for instance RUE to map change “hot spots”). Understanding the rainfall-vegetation relationship is critical specifically for dryland areas in Africa due to their natural vulnerability to rainfall variability and the adverse effects of “climate shocks” [5].
Fig. 3: Integrated flower mapping workflow landscape (LS), ecosystem (ES), hyperspectral (HS), multi-spectral (MS).
The assimilation approach, hierarchical layering of the two low to moderate resolution datasets, is straightforwardly applied, using freely available data. Thus land dynamics processes at a local to landscape scale cannot be considered, such as decisions at a village level for instance [8]. This information loss is however offset by the permissible temporal resolution, between 16 to 30 days, of the two datasets and by linking the results to the qualitative information on land transformation processes from VHR imagery in Google Earth. The link to the Google Earth data helps to confirm that most land productivity decline areas are caused by human-induced land transformation processes such as deforestation or conversions of natural wetlands to croplands.
Integrative mapping of the floral cycle in Africa
Datasets and methodology
Fig. 3 shows the integrative flower mapping workflow proposed to map the abundance, distribution and floral cycle of flowering plants in the landscape. In situ smartphone data is suggested for validation. The scaling possibility is given by the spectroscopic concept of spectral unmixing using “pure as possible” spectral endmembers (EM) of features or materials within an image [16]. EM from a high resolution and hyperspectral image (or even field spectra) can be fitted that is up-scaled to a multi-spectral resolution image for landscape scale and repetitive monitoring (Fig. 3) [9]. Likewise, “pure” EM of “smaller than the pixel resolution” features are used to downscale hyperspectral imagery producing per pixel abundances of features that occur at a sub-pixel level. Using EM and spectral unmixing is a physically based scaling approach for integrating data from different satellite sensors [11].
In the experimental (example) study Airborne 0,6 m AISA Eagle hyperspectral airborne imagery (400 – 1000 nm spectral range) was captured over a site in Kenya during the maximum flowering period. The aim was to probe whether the 0,6 m HS imagery can be up- and down- scaled for integrative that is multi-sensor flower mapping and to monitor the floral cycle in Africa (regarding the proposed workflow in Fig. 3). The input imagery, i.e. 0,6 m AISA hyperspectral data, was pre-processed to surface reflectance and spectral endmembers were derived on the imagery for flowering as well as non-flowering plants (“pre-processing” and “surface reflectance” in Fig. 3). Geotagged smartphone data from the field was used to cross verify the location of flowering plants for EM extraction on the AISA imagery. EM spectra from 0,6 m AISA eagle was spectrally fitted (scaled) to simulate a 2 m Worldview-2 image using spectral weighting and spatial nearest neighbourhood re-sampling. Spectral unmixing was used on both datasets, using the respective EM as input that is “training” spectra, to produce abundance maps for flowering and non-flowering plants.
Theoretically, field spectral signatures could also have been collected for surface reflectance calibration (Fig. 3). Likewise it is recommended to use “minimum noise fraction” transform (MNF) or principle component analysis [17] as pre-processing steps (“spectral processing” in Fig. 3) in order to reduce autocorrelations and dimensions in the data and be able to derive representative and “pure” endmembers for spectral unmixing. Alternatively a data fusion technique, for instance wavelet transformation, between sensors data can be employed [11] or more straightforwardly spectral weighting functions can be applied [18]. Finally, local and landscape mapping of the floral cycle is possible, if smartphone data is integrated in periods where no imagery capturing is possible or feasible. Moreover if structural and biochemical land surface variables such as leaf area index (LAI) are modeled within a radiative transfer model, augmented with spatial explicit remote sensing information, ecosystem process modeling is possible (Fig. 3) [19].
Results and discussion
Fig. 4a and b below show the experimental flower mapping results. Fig. 4a shows the 0,6 m AISA hyperspectral data unmixing result. Fig. 4b is the abundance map of the simulated 2 m Worldview-2 image after the AISA derived EM were up-scaled and spectrally fitted to the spectral and geometric resolution of the Worldview-2 image. In general the spatial patterns in the two images are similar. Flowering activity, yellow in Fig. 4a and b, is scattered similarly in both images, whilst non-flowering chlorophyll active vegetation, green colour, was found mostly within riverside vegetation in both datasets respectively. The information about the distribution of flowering plants in the landscape can be linked to pollen and honey flow measures for bee health assessment studies [20]. The bee related indicators would have to be collected at apiaries that are within the study area. In this study the link between bee health and landscape phenology (flowering) was not performed.
Fig. 4. EM abundances maps from spectral unmixing. Flowering plants are shown in yellow and non-flowering ‘green’ trees and shrubs in green.
Fitting the AISA derived EM to Worldview-2 data for wide-area mapping is more feasible, since Worldview-2 data can cover larger areas systematically and repeatedly. However we did not account for the spectral variability of the EM if the area of observation, and with that the landscape complexity (heterogeneity), was enlarged [18]. We also did not apply the spectra to an image taken some years back due to the lack of historical data and the high temporal dynamics of flowering. For monitoring of flowering activity, using several images of various spatial resolutions in time sequence, a data fusion technique, such as the convolution image fusion procedure, can also be used for seamless mapping. Although this is a proven technique [11], it has never been applied to a subtle phenological phenomenon such as flowering. Also, data fusion on multi-source instance imagery produces a fused image in which the physical information content is lost [11].
In this experimental study we suggest a spectral weighting and EM based spectral unmixing approach (up-scaling) for integrative mapping, since we are also interested in repetitive mapping of the floral cycle using various high resolution imagery datasets on a landscape scale. The smartphone tagging methodology in the field proved to be valuable in verifying EM on the hyperspectral (AISA) image. Using smartphone data feeds to monitor phenological responses such as flowering over larger and remote area, also at various times within the flowering period, would essentially help to fill in data gaps and thus effectively improve floral cycle mapping in Africa.
Overall conclusion
Both experimental examples show that various independent information (data) levels are needed in a data scares and highly complex environment, as such Africa, for wide-area mapping of emerging environmental issues. For land productivity mapping, no assertive conclusions about drivers of land change would have been possible without cross versification of the results with high resolution reference data on land transformation processes and TRMM rainfall “normalisation”. The flowering mapping example and discussion shows that the multi-data assimilation approach needs to consider data availability, needs and costs as well as site characteristics including the temporal characteristics of the floral cycle. A physically-based spectral unmixing approach, using EM, is suggested for up- and down-scaling to map the floral cycle using multiple sensors. Since landscapes in Africa are highly heterogeneous and dynamic, more research is needed to probe the variability of the EM in the spectral sub-space for accurate scaling.
We also suggest including in situ data on the floral cycle from smartphone geotagged information to fill temporal data gaps in remote sensing effectively. First results using smartphones in the field for flower mapping are encouraging; however the smartphone field mapping method needs to be probed for its potential to capture floral cycles more systematically at key sites. This can possibly be achieved by developing a mobile smartphone application that will minimise data entry efforts in the field and facilitate the link to a given database on flowering plants automatically.
Ultimately the geo-spatial assessment results show that if data assimilation is performed wisely, integrative EO assessment of environmental problems over larger areas is feasible, with the results being highly relevant for mitigation and adaptation efforts, including early warning systems.
