In East Africa and the Sahel, livestock herding is the principal livelihood for over 40 million people and represents up to 60 percent of agricultural GDP. Despite the considerable contribution of the livestock sector to local economies and to household welfare, pastoralist populations remain extremely vulnerable to climate shocks. Drought is the largest cause of livestock mortality, often resulting in food insecurity and poverty traps for pastoralists in the region. The increase of climate variability combined with rising population pressure and rangeland degradation makes typical herding strategies, such as migration, less effective and strains informal traditional coping mechanisms. Early detection of drought extent and severity coupled with financial mechanisms to trigger response actions is thus critical to mitigate drought impacts.

Earth Observation (EO) data are increasingly used in the framework of micro-, meso- and macro-level agricultural index-insurance and disaster risk finance initiatives in Africa as primary input data to design insurance-indices to assess drought impacts and trigger financial response mechanisms. Despite the strong demand and the potential benefits of such initiatives, there are still only few examples of operational implementation and significant uptake of EO-based products. This is, among other reasons, because of the limited number of studies and methodological frameworks to assess the accuracy of different EO-based methodologies in the context of agricultural index-insurance (e.g. quantification of the basis risk, occurring when insurance payouts depend on an index that is not perfectly correlated with the actual losses experienced).

This study analyzes and compares different data sources and index design options for drought index-insurance on the extensive pastoral areas of East African and Sahelian drylands (i.e. arid and semi-arid savannas). Specifically, we compare spatially- and temporally-aggregated indices based on time series of Water Satisfaction Index (WSI) and the Normalized Difference Vegetation Index (NDVI, from different data sources). We analyze the effect of different index calculation procedures (i.e. seasonality, spatial aggregation, anomaly determination, etc.) to assess and discuss the advantages and limitations of each design option.
We then compare our aggregated drought indices with existing national-level long-term datasets on forage biomass availability and drought and food security impacts to assess their accuracy in detecting the spatial and inter-seasonal variability of drought events.

The results indicate that WSI- and NDVI-based indices are relatively highly correlated (r between 0.7 and 0.8). However, NDVI-based indices seem more accurate in the assessment of major drought events and less prone to significant mismatches between the index and the observed drought impacts on rangeland resources. In addition, WSI performances are sensitive to the quality of the input rainfall estimates dataset. How NDVI or WSI are used for the drought index design also has a significant influence on the overall performance of the product and must be carefully evaluated taking into consideration the trade-offs between accuracy and critical operational needs of index-insurance (e.g. financial implementation, simplicity, regulations, etc.).

This study has important implications for the implementation of multi-level index-insurance and disaster risk financing initiatives in the pastoral areas of Africa. It also demonstrates the fundamental need of a stronger involvement of the EO scientific community in the design of innovative agricultural index-insurance products using state-of-the art technologies and methods, and in supporting the establishment of frameworks for defining international quality standards.

Acknowledgements
This project is supported by African Risk Capacity (ARC) - A Specialized Agency of the African Union

Area A - Board 45

Global information on land cover (GLC) is of special importance for environmental change studies, but also to governments and international initiatives engaging in land resource management and sustainable development. GLC data exist at various spatial resolutions for applications on global as well as regional scales depending on application type. Studies have shown that land cover classification results show significant differences in respect to thematic accuracy. Automated approaches provide efficient and economic solutions but are generally characterised by low classification accuracy. The poster shows results of a quantitative and qualitative comparison of selected land cover products for Namibia. These include medium to low resolution products such as AVHRR DISCover, MODIS MCD12Q1 and CCI–LC (Climate Change Initiative Land Cover) as well as higher resolution products such as NLC-2010-SII (Namibia Land Cover) and the recently published prototype classification Sentinel 2 Land Cover map of Africa 2016 of the European Space Agency (ESA) with a 20 m resolution mapping approach. The results were compared with own classifications for four test sites (50 km x 50 km) using images of Landsat 8 (OLI) with a spatial resolution of 30 m.
The incomparability of land cover data through use of different procedures, tools and thematic nomenclatures (classification schemes/legends) is a major challenge. By assigning a common legend to all datasets, the investigated classifications became comparable showing significant differences in terms of quantity and quality. These vary depending on land cover type and spatial resolution. The Sentinel 2 Land Cover classification definitely shows the highest overall accuracy of all freely available products but lies significantly below the accuracy of the own test sites despite their lower resolution. This finding suggests that other factors such as spatial coverage, tools and procedures influence the accuracy of land cover products as well.

Area A - Board 47

Introduction

Biodiversity loss is a major concern of this century. Protected areas have been established to protect threatened species in sensitive regions. However, monitoring the ecological outcomes of these large areas is challenging, because carrying field surveys in these remote and difficult to access areas is fastidious. One example to this is the Pendjari National Park (Benin), which is the largest remaining preserved savannah ecosystem in West Africa. The park holds natural habitats which are critical for a large number of species, yet threatened by anthropogenic pressure and climate change [1]. However, there is currently no map detailing at fine scale the distribution of the natural habitats, which could help the conservationists quantify the impacts of land cover change on biodiversity in the park.
Recently launched Sentinel-1 and Sentinel-2 constellations, delivering imagery with a high revisit frequency, could be used to fulfill this need, thanks to their high spatial resolution and the complementarity between optical and radar data. But despite the free availability of the images, their contribution to map natural savannah vegetation has so far not been assessed [2]. The aim of this study is thus to evaluate the potential of Sentinel-1 and Sentinel-2 time series to map natural habitats in a large savannah protected area of high biodiversity interest at a high spatial resolution.

Study area

Pendjari National Park covers an area of about 2,800 km², dominated by grasslands, shrub savannah, savannah woodlands, gallery forests and riparian vegetation around water bodies. It is subject to a tropical Sudanese-Guinean climate marked by a rainy season lasting from June to October, with an average of 1,100 mm annual precipitation.

Reference data

Natural habitats present in this park were categorized into six classes: grassland savannah, shrub savannah, savannah woodland, forest, bare soil and water bodies. Ground truth vegetation data were collected in 2017 around seventy GPS points. Polygons were then digitized following a small and homogeneous area around these points using Bing Aerial Photos® as well as Sentinel-2 imagery for 2017. More polygons were digitized by photointerpretation, ensuring to have a balanced reference dataset. In total, the reference dataset is composed of 182 polygons.

Remote sensing data and methods

We used different sets of input data to compare the contribution of Sentinel-1 and Sentinel-2 time series to map the savannah vegetation.

For Sentinel-2, the images acquired over Pendjari National Park in 2017 with less than 50% cloud cover were kept. In total, 35 acquisitions were used. The Level-L1C images (orthorectified and radiometrically corrected) were processed to Level-L2A (corrected of atmospheric effects) using the Sen2Cor processor. A binary mask of noisy/non-noisy pixels was extracted from the scene classification provided with L2A images at 20m resolution, and it was then resampled to a 10m spatial resolution. A temporal gap filling was applied using the Orfeo ToolBox (OTB) application on the time series associated with each band at 10m (B02, B03, B04, B08) and each band at 20m resampled to 10m (B05, B06, B07, B8A, B11, B12), to fill the gaps due to noisy data. Then, the NDVI (Normalized Difference Vegetation Index) and the NDWI (Normalized Difference Water Index) were computed.

For Sentinel-1, all the images acquired in 2017 over the same ascending orbit were used, for a total of 29 images. We used Level-1 GRDH (Ground Range Detected High resolution) products recorded in interferometric wide swath mode. The images were calibrated to convert them to radar backscattering coefficient σ0 for both polarizations VV and VH, and orthorectified using the OTB application with an output spatial resolution of 10m. Finally, an adaptive temporal filtering was applied to both polarizations σ0(VH) and σ0(VV) to reduce speckle effects.

Different stacks of time series were considered as inputs of the classification:
A- Sentinel-2: four bands (B02, B03, B04, B08);
B- Sentinel-2: ten bands (B02, B03, B04, B08, B05, B06, B07, B8A, B11, B12);
C- Sentinel-2: NDVI;
D- Sentinel-2: four bands and NDVI;
E- Sentinel-2: NDVI and NDWI;
F- Sentinel-1: σ0(VH), σ0(VV), σ0(VH)/σ0(VV);
G- Sentinel-2 and Sentinel-1: A stacked with F;
H- Sentinel-2 and Sentinel-1: B stacked with F.

Classifications were performed using the Random Forest algorithm provided by OTB applications. A Monte Carlo procedure with 20 repetitions was followed to split randomly the reference data into training (70%) and validating (30%) subsets. The split was performed at the polygon level in order to have an independent set of pixels between the training and the validation steps. The resulting classification map was sieved to reduce the salt and pepper effect. The confusion matrix was computed based on the validation subset. Overall accuracies (OA) were estimated from the confusion matrices issued from the 20 repetitions. To compare the performances of each pair of input configuration, a Wilcoxon rank-sum test between each pair of distributions of OA was run.

Results

On average, the four bands at 10m of Sentinel-2 provide the best results in terms of classification overall accuracy (input A, mean OA = 76.9%). The lowest performance is observed for Sentinel-1 data alone (input F, mean OA = 68%). The combination of Sentinel-2 and Sentinel-1 data (input G, mean OA = 76.8% and input F, mean OA = 76.7%) provides equivalent results to Sentinel-2 data alone.
All the pairs of distributions of OA are not significantly different except for the distribution associated with Sentinel-1 data (input F) which is significantly lower than the others (p-value < 0.01).
The best OA was obtained with input D (87.7%).
The individual class accuracies showed that in general, grass savannah, savannah woodland, forest, water and bare soil were well detected, while shrub savannah was confused with grass savannah or savannah woodland.

Discussion & Conclusion

This study aimed at evaluating the contribution of Sentinel-1 and Sentinel-2 time series to map the natural resources in a savannah landscape. The best results in terms of classification accuracy were obtained using Sentinel-2 data alone. Sentinel-1 data provided significantly lower results. The combination of Sentinel-2 and Sentinel-1 data did not significantly improve the results compared to Sentinel-2 data alone.
Obviously, radar data is sensitive to different landscape features (small relief such as thalweg, soil roughness...) that were not taken into account in the validation dataset which was only focused on land cover type. Additional ground survey would be needed to perform a deeper analysis of the complementarity between Sentinel-2 and Sentinel-1 data to map savannah features.
Using only the Sentinel-2 four bands at 10m and the NDVI, we were able to map the natural resources of Pendjari National Park at an overall accuracy of 87.7% (see attached map). We conclude that the high revisit frequency of Sentinel-2 (one acquisition every ten days over the WAP complex) makes possible a good discrimination between different types of natural savannah habitats (grass savannah, shrub savannah, savannah woodland, forest, water and bareground), despite the clouds during the rainy season occurring in this tropical area. Shrub savannah was confused with the other classes of savannah because theses classes are defined as a gradient in canopy cover that may be refined.

With this study, we provided the first map of natural habitats in the Pendjari National Park at 10 meters spatial resolution. Further processes will be performed to improve the classification accuracy, such as a better definition of the classes and a segmentation of the image followed by an object-based classification. In addition, classifications based on Landsat imagery of 2000 and before will be performed to monitor the natural resources distribution in Pendjari National Park over the time.

References

[1] Amahowe, O., Houessou, L., Ashanti, S. and Tehou, A. (2013) Transboundary protected areas management: experiences from W-Arly-Pendjari parks in West Africa. PARKS. 19. 95-105.
[2] Schulte to Bühne H. and Pettorelli N. (2017) Better together: Integrating and fusing multispectral and radar satellite imagery to inform biodiversity monitoring, ecological research and conservation science. Methods Ecol Evol.;00:1–17.

Area A - Board 48

The processes of occupation by human activities managed without planning produce great pressures on natural systems. The interaction between society and nature define the types of land use, which, define the spatial and structural pattern of landscapes. These interferences, convert extensive and continuous areas with forest cover into forest fragments, compromising their environmental functions and consequently the services provided by ecosystems. Landscapes constitute the manifestation of natural and anthropic elements and its structure is the result of physical, biological, political, economic and social interactions, fragmented or connected for different land uses. The landscape analyses are fundamental to the environmental planning in order to designate a setting that corresponds to the spatial structure, subsidizing the understanding of the impacts of anthropogenic changes in the natural processes. Anthropogenic actions influence most landscapes and the resulting mosaic is a mix of natural and anthropogenic landscape elements that vary in size, shape, pattern, and structure. Measures of landscape patterns can quantitatively analyze landscape structure. This study aimed to analyze the effects of land use on the spatial and temporal patterns in the landscape of the Tietê-Jacaré Hydrographic Basin, Brazil, from 2007 to 2017 through the application of the Environmental Quality of Water Resources Index (EQI Hydro), Vegetation Environmental Quality Index (EQI Bio) and Landscape Environmental Vulnerability Index (EVI). The Tietê-Jacaré Hydrographic Basin is located in São Paulo state - Brazil between coordinates 49º14' and 47º70' west longitude and 21º62' and 22º79' south latitude, with a population of 1,462.855 inhabitants and a total area of 1,181,090.00 hectares distributed in 37 municipalities. Landsat satellite images were used that correspond to 220/75, 220/76, 221/75 and 221/76 orbit/frame, with overpass dates of April 30th, 2007, and September 21st, 2017. For 2007, Landsat 5 TM images were used, bands 5/4/3, and for 2017 Landsat 8 OLI/TIRS images were used, bands 6/5/4. The classification of land use was based on a multi-level classification system proposed by the Land Use Technical Manual from the Brazilian Institute of Geography and Statistical. In the primary hierarchical level I, four classes was included that indicate the principal land use categories. The secondary hierarchical level II analyzed the types of land uses that were included in the first level. The tertiary hierarchical level III explained the land uses themselves. The landscape indices were used to describe the temporal landscape pattern resulting from the influence of anthropogenic land use processes in Tiete-Jacaré Hydrographic Basin for 2007 and 2017. This approach assumes that the relevance of the environmental impacts of the types of land use is associated with the vulnerability and sensitivity of environmental components. The concept of Fuzzy Logic was essential to understanding the application of landscape indices. Being an extension of Boolean logic that supports logical values intermediate between the false (0) and the true (1). It was considered 0, the minimum degree of ease and 1, a maximum degree, corresponding to predominantly altered by anthropic systems, where larger naturalistic correlate with less impacted and lower naturality systems with more impacted systems. The Vegetation Environmental Quality Index was obtained based on the analysis of area (EQI-Area), shape (EQI-Form) and distance (EQI-Distance) metrics of the native vegetation fragments. The Environmental Quality of Water Resources Index was determined based on the overlapping of the thematic maps of agricultural and non-agricultural anthropic activities obtained by reclassifying the use and land cover and drainage network. This analysis represents the distance of water resources in relation to impacting sources, considering the minimum quality grade (EQI Hydro = 0) the rivers with a distance from the impacted areas approximately zero. The maximum degree of quality (EQI Hydro = 1) corresponded to rivers with a distance from the impacted areas greater than 1000 meters. The Environmental Vulnerability Index was obtained through the correlation of the EQI-HYDRO and EQI-BIO. The greatest degree of environmental vulnerability of the landscape (ELV = 1) was ascribed to a state most susceptible to impacts, while the lowest degree of environmental vulnerability landscape (ELV = 0) is assigned to a condition with the largest capacity to absorb impacts (high resilience). A reduction in the areas of native vegetation over time was identified with loss of 2.72% in ten years, representing over 32,149ha, an expansion of agricultural areas of 2.05% (24,507.53ha) and consequently a reduction of the environmental quality in the landscape. These impacts compromise ecosystem structure and ecosystem services, e.g. through impacts on soils that support vegetation cover as the major sources of energy for terrestrial life. This predominance of agricultural crops and mainly of sugarcane coincides with the characteristics of the São Paulo state, mainly in the interior, which is the largest producer of sugarcane in Brazil. An increase in the number of fragments with low environmental attributes (EQI BIO = 0.0 to 0.2) was observed, while the fragments with intermediate qualities (EQI BIO = 0.2 to 0.6) decreased in number. The fragments with high environmental quality (0.6 to 1) were maintained in terms of area, principally due to these areas being related to water bodies and steep slopes preventing agricultural land use. These fragments are related mainly to the principal rivers in the regions as Jacaré-Pepira, Jacaré-Guaçu and Jáu river. The EQI of Water Resources follows the same trend as EQI Bio, which identified a decline in environmental quality. There is a decreased number of fragments of low quality (EQI HYDRO = 0.0 to 0.2), while the fragments of intermediate quality (EQI HYDRO = 0.2 to 0.6) increased and the fragments of high quality (0.6 to 1.0) decreased. However, the landscape showed an improvement in some regions, where fragments with very low quality were converted in low quality, but in the same way, very high quality became in medium or low quality, whereas exist only a few fragments with high quality in both scenarios. The Environmental Vulnerability Index shows a trend of expansion of anthropogenic processes in the landscape due to the reduction of vegetation fragments over time, in which the remaining natural vegetation fragments, although preserving some ecological value, mostly become worthless for biodiversity protection, serving only as ecological corridors. Lower Environmental Vulnerability Index values from 0 to 0.2 represent areas with lower vulnerabilities. Their frequency decreased from 2007 to 2017. This situation is different for the Environmental Vulnerability Index ranges from 0.6 to 1.0, which increased in frequency from 2007 to 2017. Fragments were lost over time, particularly vulnerable fragments in 2017 were extinguished and converted for other land uses. The areas with higher degrees of vulnerability are most environmentally susceptible to disturbances and have lower resilience. The entire landscape of Tiete-Jacaré Hydrographic Basin needs special attention in municipal planning, as 85% of the native vegetation area is in the Environmental Vulnerability Index interval from 0.0 to 0.6. Virtually all municipal green areas are highly vulnerable to land use impacts. The need for restoration and conservation of native vegetation areas is evident in order to reduce the degree of environmental vulnerability. Indices used were presented as important tools for the diagnosis aimed at the conservation of ecosystems. They enabled accurate analysis of each fragment related to the interconnections observed between the environmental qualities of the different landscape compartments, where the loss of natural areas directly involves loss of quality of several compartments, which are interconnected. The Tietê-Jacaré Hydrographic Basin requires an immediate plan aimed at maintaining the remnants of native vegetation, giving priority to conservation and restoration of these areas, considering not only the current state of the region but the developmental trend which has been following this. The growth of human activities and the loss of natural areas can compromise the biodiversity present in the study area, resulting in the loss of environmental functions and consequently the benefits they provide, considering that these areas serve as support for maintenance of urban and agricultural activities generally.

Area A - Board 49

The accurate mapping and quantification of above ground biomass (AGB) is required for a number of applications, including carbon accounting, _re and grazing management, amongst others. Accordingly, relating field measurements of AGB to satellite-derived indicators, most prominently the Normalised Difference Vegetation Index (NDVI) has been a feature of the remote sensing literature for over 30 years. Recently, there has been an increase in the use of machine learning methods and the incorporation of auxiliary environmental variables for spatiotemporal modelling. However, there is increasing evidence that these models may be vulnerable to artefacts of data structure, such as spatial autocorrelation and inappropriate auxiliary variables, which may hinder the development of accurate models. In this study, a robust methodology for the creation of moderate-resolution AGB estimates is presented. We obtained AGB data from an 18-year long dataset comprising 533 sites within the Kruger National Park of South Africa. We then generated a 36 1km-resolution NDVI product by downscaling the GIMMS 3g NDVI using Empirical Orthogonal Teleconnections (EOT) and the MODIS MYD13A2. AGB was then predicted based on a series of NDVI-metrics and auxiliary environmental variables in a Cubist regression model framework. Our analysis consisted of two components: i) a comparison of validation approaches, including a k-fold cross-validation (CV) and multiple spatial/temporal CVs; and ii) a variable selection component, incorporating forward feature selections (FFS) on the above validation strategies. Prediction accuracies differed considerably, with the Root Mean Squared Error ranging from 1310 to 1844 kg ha-1, depending on the variables and validation strategy employed. Errors were consistently higher with spatial or temporal validation strategies. Spatial overfitting was prominent in most models, which we attribute to spatial autocorrelation within the predictor variables. Comparatively, the NDVI-biomass relationship was highly variable between years, with unseen years being poorly modelled. This potentially results from changing species composition and moisture content on an annual basis. The FFS was effective at correcting these issues, where possible, by constructing models with appropriate variable combinations. For temporal models, the profile of auxiliary variables was increased leading to a more deterministic prediction approach. This study contributes to the growing literature highlighting the potential pitfalls of machine learning for spatiotemporal predictions, and offers strategies for their detection and mitigation.

Area A - Board 50

Woody plant encroachment has been documented for savannahs and grasslands in nearly all continents. Yet the drivers of this process remain unclear, with a range of local and global factors postulated. The traditional ecological narrative dictates that shrub encroachment is a localised phenomenon, resulting from poor land management regimes. The most frequently proposed mechanisms are overgrazing and suppression of fire, both of which are common management techniques in sub-Saharan Africa. More recently, increased focus has been directed at the role of global factors in woody cover dynamics. As savannah woody cover is constrained by both total and wet season rainfall, changes in precipitation regime have been proposed as drivers of shrub encroachment. This theory has been supported by small-scale field experiments showing shrubs disproportionally benefiting from increases in rainfall frequency, amount, and variability. A further potential global factor is the ongoing rise in atmospheric CO2 concentrations since the industrial revolution. A theoretical understanding of water limitations to woody cover in savannahs makes it reasonable to assume that CO2-driven increases would be concentrated in water-limited environments. This has been observed across South Africa using aerial photography and globally using satellite-derived Rain Use Efficiency (RUE).
Here, we combine satellite-derived fractional woody cover maps with a suite of potentially explanatory variables, to elucidate on the potential drivers and mechanisms of woody cover change, in South African savannahs. The study area consists of the Limpopo and North West Provinces in northern South Africa. These municipalities cover a plurality of the savannah biome within South Africa (193,200 km2, 49% of the total savannah area), in addition to containing 33,830 km2 of grassland.
More specifically, we test the three, abovementioned, competing hypotheses on the drivers of woody encroachment for South African savannahs. Our modelling framework was developed using Generalised Additive Models (GAM). We collated a series of 11 variables that have a hypothetical basis for explaining woody cover changes. These variables can be grouped into three categories: rainfall-derived, human, and natural factors.
Fractional woody cover changes were mapped using Landsat-derived % woody cover layers, based on the methodology developed in Higginbottom et al. (2018, ISPRS Journal Ph&RS, 139, 88-102). In summary, two five-year epochs (1984-1988 and 2008-2012) of Landsat imagery were used to generate pixel-level seasonal spectral variability metrics, at 120 m resolution. Reducing the pixel resolutions improves the classification accuracies, and is more suited for observing overall trends. High-resolution imagery were classified into woody/ non-woody masks and used as training data for a Random Forest regression for the fractional cover of each 120m-pixel. The Random Forest model was applied to the Landsat metric stacks to generate the two epochal maps. We then calculated both the absolute percentage change and the relative percentage change in woody cover between the two maps.
The fitted models had R2 values of 0.39 for absolute change and 0.41 for relative change. The results show that the modelled variables most closely matched the a priori responses of the carbon fertilisation hypothesis. In recent years, this explanation has been postulated by studies using a variety of methods to account for the observed woody encroachment. Further work in this arena is still necessary, particularly where the data sources are sub-optimal. Land-use history and rainfall dynamics are especially difficult to quantify and would require further investigation. Furthermore, additional factors, such as reactive nitrogen deposition and megafauna extinctions, are likely to be relevant but were not included in our models. If carbon fertilisation is the key driver of shrub encroachment in savannahs, it would raise concerns for future environmental change: as CO2 levels continue to rise more savannahs and grassland are likely to experience an increase in woody cover levels, which has been linked to savannah land degradation.

Area A - Board 51

Trends in vegetation greenness, measured using the Normalised Difference Vegetation Index (NDVI), are one of the most ubiquitous tools for inferring ecological change at large spatial scales. The NDVI has long been a favoured metric, owing to its simple calculation, correlation with various ecological attributes, and transferability between sensors. In particular, the Advanced Very High Resolution Radiometer (AVHRR) - derived Global Inventory Monitoring and Modelling System (GIMMS) dataset has been used extensively for environmental change purposes.
The regions most studied using GIMMS-NDVI data are the drylands of Africa. These areas are particularly well suited for this form of analysis, due to the limited historical coverage of other Earth-observation archives in the region, with the exception of South Africa. Furthermore, NDVI becomes less sensitive to vegetation in dense canopies (NDVI > 0.7) but is relatively responsive to the low biomass levels found in savannahs and grasslands. Accordingly, a considerable amount of earlier work on the use of NDVI for monitoring environmental change focussed on the Sahel region.
This study investigates the NDVI dynamics of African savannahs using a variety of time-series analysis techniques and NDVI-derived metrics. The overarching aim is to understand how vegetation dynamics have occurred and evolved during the 1982-2016 period and the ecological implications of these changes. Towards this aim, we generate two NDVI time-series from the GIMMS dataset: the annual maximum and aggregate sum values, hereafter NDVImax and NDVIsum, respectively. These series are used as inputs into monotonic linear and breakpoint regression models. The slopes and any associated breakpoints of these models are classified and examined as indicators of large-scale ecological change.
We further processed the GIMMS data to maximum monthly composites and discarded the months for the incomplete 1981 year, resulting in a 34-year (408 months) time-series. When gaps remained after compositing, values were estimated by linear interpolation across months. We calculated the median NDVI using the monthly time-series, and discarded pixels with a value of less than 0.15 or greater than 0.8 from further analysis. This removed areas with very low vegetation cover (e.g the Sahara and Namib deserts) and dense forests (e.g. the Congo and Guinean forests). The monthly time-series were aggregated into two annual metrics: NDVImax and NDVIsum. This resulted in two 34-year annual time series, to be used as inputs for the trend analyses.
We applied a standard linear regression model on both annual NDVI time-series. Firstly, a standard linear model was implemented with no breaks quantified. Secondly, a model allowing for two structural changes was applied. Breakpoints were determined first by an ordinary least-squares moving sum (MOSUM) test. When breaks were detected, the number was estimated by the Bayesian Information Criterion (BIC), and timings were set based on the residual sum of squares (RSS).
The slopes resulting from the linear models were grouped into three categories: greening, no change, or browning, after removing pixels with insignificant trends. By comparing the trend classifications from the NDVI-max and NDVI-sum models, six combinations were possible. To classify the breakpoint outputs the following procedures were applied. Firstly, all segments with a length of less than seven years or resulting from an insignificant break were discounted. Secondly, the remaining components were classified into greening, browning or no change based on the slope value. Finally, as only two breaks were allowed, a maximum of three segments was deemed possible.
According to our results, greening pixels comprised 75% of the NDVIsum slopes, and 80% of the NDVImax, whereas browning was identified in 25% and 20% of pixels, respectively. Removing trends that did not meet a P < 0.05 significance threshold eliminated some of these pixels. Pronounced increases in NDVI were observed in the Sahel and southern Africa. Browning was concentrated in east Africa, Angola, Zambia, and Mozambique, with dispersed and isolated patches in the northern Sahel. Results also show that, regardless of NDVI metric, a large majority of African drylands, across all regions, have experienced only greening trends in the 1982-2015 period: a minimum of seven years’ increase and no seven-year decrease in the NDVI metrics. Conversely, few areas displayed only browning trends (a minimum of seven years’ decrease with no increase). These pixels were geographically clustered in the central Sahel, Angola-Zambia-Mozambique, and Tanzania. Patches of trend reversals (e.g. browning to greening) were present, although not geographically extensive.
In summary, the overarching trends of African savannahs across our study period is of vegetation greening. This has occurred across all regions, even when different precipitation patterns have occurred. This would indicate the role of pan-continental driver(s). There are numerous ongoing trends which are beneficial for plant growth that could contribute to this, such as elevated CO2 levels, nitrogen deposition, and increased temperatures. Areas of browning predominantly occurred in areas where either population growth had been high, or where phenological change has curtailed the growth season.

Area A - Board 52

The analysis of vegetation structure in savanna ecosystems is an important task to assess the amount of above ground biomass (AGB) and their role as part of the global carbon cycle (Naidoo 2015). Here, height related parameters are a valuable source of information as they are fundamental variables for a large variety of study fields such as ecology, hydrology, agriculture, geology, pedology or geomorphology (Pulighe & Fava 2013). Within the Kruger National Park (KNP) freely available reference information is scarce especially when targeting reliable height data at high spatial resolution. However, these data sets are of exceptional value for studies being conducted in this area.
In order to create the first publicly available height model for the whole KNP and its surroundings to the west the GeoImaging Accelerator (GXL) software developed by PCI Geomatics was applied as the FSU Jena just recently was named a “PCI Technology Center of Excellence in Remote Sensing and Geomatics”. The overall goal is to derive digital surface models (DSM) as well as digital terrain models (DTM) for the complete mentioned area to finally produce canopy height models (CHM). For this purpose, aerial imagery was gathered from the updated Colour Digital Aerial Imagery (CDNGI 2013) archive of the South African Department for Rural Development and Land Reform captured by the Digital Mapping Camera (DMC) at a nominal spatial resolution of 25 and 50 cm. For this study we have been using existence workflows for the derivation of height models and were able to extend them to consequently improve the extraction of detailed height information.
First results indicate a good overall fit between the extracted height from the DMC data and the reference data which was acquired during a terrestrial laser scanning (TLS) field campaign with R² being greater than 0.8. With the adaption of new aerial imagery at higher resolution and further adjustments of existing algorithms height extraction will be further improved to derive KNP’s first wall-to-wall height model which may serve as a unique reference data set for studies being carried out in this complex ecosystem.

Area A - Board 53