We illustrate the steps
SeeSeeTop-Down approach
to split a complex landscape (of any size) into homogeneous units using GIS and RS information and socioeconomic surveys to study mitigation potential
(Fig. 2.1). This may be of interest, for example, where a carbon credit project
is implemented, or if a district, province, or other authority wishes to assess the mitigation potential of a number of agricultural technologies. Once the landscape boundaries
are defined, one can disaggregate the complex landscape into different units. If the landscape boundaries are not delineated, the analyst may choose to select an area that is representative of the larger region in order to extrapolate results. The landscape can be analyzed initially using a combination of RS and GIS. We suggest different approaches to disaggregate a landscape and decide where to conduct field measurements.
After selecting a landscape for assessment and developing a conceptual model of land-use and land-cover (LULC), the simplest method to identify landscape units
is the exploration and visual interpretation of satellite imagery, preferably with the best available spatial resolution and observation conditions (e.g., peak of vegetation productivity). LULC classification (using object-based approaches and VHR imagery) and landscape classification (using RS vegetation productivity parameters) are more sophisticated methods of approaching a landscape. With visual interpretation, numerous landscape features can be characterized using physical (e.g., geomorphology, vegetation, disturbance signs) and human criteria (e.g., presence of population, land-use, and infrastructure). This yields relatively large, homogeneous landscape units (e.g., describing the mosaic of LULCs in an area). By comparison, automated LULC classification yields results at a much finer spatial scale. In most cases it maps the individual fields that make up a landscape. The process of automated LULC mapping involves:
-
1.
Discriminating areas of general LULC types such as croplands or shrublands
-
2.
Characterizing structural traits of all these types
-
3.
Integrating areas and traits to identify homogeneous landscape units
The two first steps require the composition of the landscape
to be characterized (i.e., the areas under each of the field or land types according to Fig. 2.1), and their spatial configuration (i.e., the arrangement of field or land types).
In landscapes with dominant smallholder agriculture, cultivated land can be easily recognized through the presence of regular plots with homogeneous surface brightness, and minor features such as ploughing or crop lines and infrastructure. In addition, the structural heterogeneity of cultivated areas can be assessed by the geometry of the fields (size and symmetry of the shapes), the presence of productive infrastructure and signs of disruption, such as woody encroachment within fields. Land under (semi-)natural vegetation can be characterized in terms of vegetation composition (share of trees, shrubs, and grass), signs of biomass removal or the presence of barren areas, and degradation (gullies, surface salt accumulation). Finally, in order to delimit landscape units, all descriptions should be integrated in a holistic manner using, for example, Gestalt-theory
(Antrop and Van Eetvelde 2000) to identify and digitize potential discontinuities. This simple method has the potential to enhance the quality of broad-scale land-use studies, and can be performed using freely available imagery, like Google Earth, supported by online photographic archives such as “Panoramio” or “Confluence Project
” (Ploton et al. 2012).
Landscape Stratification: An Example from East Africa
The Lower Nyando region of Western Kenya, which is dominated by smallholder producers, provides an example of the proposed approach. The CGIAR Program
for Climate Change, Agriculture, and Food Security (CCAFS)
promotes climate smart agriculture in this area. To develop and test our targeting approach, we used the three methods described above: (1) visual classification using VHR imagery, (2) LULC classification using object-based approaches and VHR imagery, and (3) landscape classification using medium to coarse resolution RS vegetation productivity parameters.
Visual Classification
Using VHR Imagery
This is a quick and relatively inexpensive visual approach for exploring landscapes. The largest costs are the acquisition of the VHR
images. Based on a QuickBird® image from the dry season (1 December 2008), six landscape classes were identified (Table 2.1 and Fig. 2.2). This initial classification can be used to test whether the units are indeed related to soil emissions and mitigation potential. The landscape classification is expected to reflect differences in land productivity and GHG emissions, because it captures inherent soil and vegetation variability.
Table 2.1 List of visual classes determined for the Nyando study region, Kenya
Class delimitation criteria and mitigation opportunities are listed for each class in Table 2.1. The limits between the classes are determined by spatial changes in the detailed criteria. As expected, these changes can be abrupt or gradual, and the ability or experience of the mapper could lead to variable results.
The visual delineation may or may not coincide with regional biophysical gradients, as shown by a quick assessment of the topography of Nyando (Fig. 2.3). In our case study, the highlands coincided with areas allocated to cash crops, while the lowlands included a continuum from subsistence crops to wooded natural land types. Delineating a landscape on the sole basis of topography may be inaccurate and/or incomplete, yet the use of a digital elevation model (DEM) is an inexpensive option to simplify landscapes.
Land-Use and Land-Cover Classification Using Object-Based Approaches and VHR Imagery
The fine-scale analysis of actual LULC
SeeSeeLand-Use and Land-Cover Classification (LULC)
allows the interface between biophysical and human-induced processes to be captured. The automated methods are more complex than the visual interpretation described previously and require digital processing of remote sensing imagery. VHR satellite imagery with pixel resolution <1 m can be used for semiautomatic (supervised) mapping of LULC in heterogeneous and fine-structured landscapes with sparse vegetation cover. To make optimal use of the rich information provided by the VHR data, object-based approaches
are recommended. Compared to pixel-based approaches, object-based approaches permit the full exploitation of the rich textural information present in VHR imagery, as well as shape-related information. They also avoid “salt and pepper” effects when classifying individual pixels. Figure 2.4 summarizes the main steps of such an approach.
In a similar way to Fig. 2.2, the landscape is first segmented into small, homogeneous subunits or objects. This process is indicated in Fig. 2.4 as
image segmentation
. Input to this image segmentation is georectified, multilayered very high-resolution (VHR) satellite images. The resulting objects (also called “segments”) are groups of adjacent pixels, which share similar spectral properties, and which are different from other pixels belonging to other objects.
To segment a landscape using VHR satellite images, the so-called segmentation algorithms are used. Contrary to the visual classification approach, objects/segments are identified in a fully automated manner. Both commercial and open source solutions exist for this task. Excellent open source solutions are, for example, QGIS (www.qgis.org/), GRASS GIS (grass.osgeo.org/) and ILWIS (www.ilwis.org/).
After segmenting the image into image objects, an arbitrary number of features are extracted for each object. In Fig. 2.4, this process is labelled as
feature extraction
.
Besides spectral features, textural features, as well as shape information, can be extracted. This information is used in a subsequent step to automatically assign each object to one of the user-defined LULC classes (process labelled as Random (RF) forest classifier). To “learn” the relationship between input features and class labels, training samples with known LULC must be provided in sufficient numbers and quality using a process called training data extraction.
Because the relation between input features and class label may change depending on image location (e.g., related to terrain and elevation), a stratified classification is recommended. For this task, before starting the classification process, the entire scene is (visually) split into a few (larger) regions (or strata) that can be considered homogeneous in terms of land-cover characteristics and the physical setting of the landscape.
The stratification is usually done just after the automated image segmentation (Fig. 2.4). Of course, results from other studies can be used as well (e.g., boundaries
shown in Fig. 2.2). Figure 2.5a shows the RGB composite
of a WorldView-2 image of the Nyando study area, and Fig. 2.5b, the corresponding DEM
. In both maps, manually drawn landscape boundaries (strata) are also shown (yellow lines).
For one of the strata, Fig. 2.6a shows the available reference information obtained from fieldwork and complemented through visual image interpretation. These training samples are necessary for the RF classifier to “learn” the relationship between input features and class labels. The resulting object-based classification
is shown for this landscape unit in Fig. 2.6b. The object limits
(e.g., gray lines in Fig. 2.6a) have been automatically derived using GRASS GIS.
For the classification, several algorithms are available (e.g., maximum likelihood classifier, CART, kNN, etc.). Based on the authors’ own and published experience, we exploited a widely used ensemble classifier called “random forest” (RF)
which often yields good and robust classification results (Gislason et al. 2006; Rodriguez-Galiano et al. 2012; Toscani et al. 2013). RF
uses bootstrap aggregation to create different training subsets, to produce a diversity of classification trees, each providing a unique classification result. For example, if 500 decision trees are grown inside the RF, one will obtain 500 class labels for each object. The final output class is obtained as the majority vote of the 500 individual labels (Breiman 2001). The proportion of votes of the winning class to the total number of trees used in the classification is a good measure of confidence; the higher the score, the more confident one can be that a class is correctly classified. Similarly, the margin calculated as the proportion of votes for the winning class minus the proportion of votes of the second class indicates how sure the classifier was in their decision. Such confidence indicators are not readily obtained using visual image interpretation. RF
also produces an internal unbiased estimate of the generalization error, using the so-called “out-of-bag” (OOB
) samples to provide a measure of the input features’ importance through random permutation. Classification performance of the entire LULC map can be based on common statistical measures (overall accuracy (OA), producer’s accuracy (PA) and user’s accuracy (UA)) (Foody 2002) derived from the classification error matrix, using suitable validation samples. Figure 2.7 shows the resulting LULC map of Nyando
obtained with this object-based classification approach and using VHR imagery from WorldView-2®.
Landscape Classification Using RS Vegetation Productivity Parameters
The two previous approaches are based on static descriptions of the landscape units (or of their constituent elements) in terms of LULC. However, alternative land traits can be explored to determine homogeneous landscape units. A promising alternative is the analysis of vegetation function in terms of the magnitude and temporal variability of primary productivity (Paruelo et al. 2001). We tested this functional analysis in Lower Nyando, using the period 2000–2012. Vegetation primary productivity was assessed through the proxy variable Normalized Difference Vegetation Index (NDVI)
. This index has been of great value for biogeographical studies, allowing rough but widespread characterizations of the magnitude and temporal variability of productivity based on homogeneous measurements across wide spatial and temporal extensions and different ecosystems (Lloyd 1990; Xiao et al. 2004; Sims et al. 2006). In this example, we acquired NDVI data from the MODIS (Moderate Resolution Imaging Spectroradiometer) Terra instrument.Footnote 1 In this dataset, one image is produced every 16 days, leading to 23 images per year.
We selected from the 13-years × 23-dates database, only those values indicating good to excellent quality conditions (i.e., pixels not covered by clouds, and with a low to intermediate aerosol contamination). Then, we used the code TIMESAT v.3.1 to reconstruct temporal series (Jönsson and Eklundh 2002, 2004; Eklundh and Jönsson 2011). This tool fits smoothed model functions that capture one or two cycles of growth and decline per year. We selected an adaptive Savitzky-Golay model (Jönsson and Eklundh 2002), assuming two vegetation growth cycles per year due to the natural bimodal behavior of rains in the study region. From the reconstructed temporal series (and by means of TIMESAT and the R v.2.15 statistical software), we calculated different functional metrics depicting average annual magnitude (e.g., mean, maximum NDVI) and seasonality (e.g., coefficient of variation (CV) of available values, number of growing seasons), and interannual variability (e.g., CV of mean annual values, annual trends) (Baldi et al. 2014).
For the sake of simplicity in the Lower Nyando example, Figure 2.8 presents: (a) NDVI maximum values as a proxy for carbon stocks of cultivated and uncultivated ecosystems; (b) intra-annual CV, describing whether the productivity is concentrated in a short period or distributed evenly through the year; (c) interannual CV of mean annual values, describing long-term productivity fluctuations; and (d) the slope of the maximum annual NDVI versus time relationship (Paruelo and Lauenroth 1998; Jobbágy et al. 2002).
Figure 2.9 shows the entire temporal range for the case of maximum annual values. Combined, structural and functional assessments provide essential information about the quality of the detected field or land types to study GHG mitigation potentials. Likewise, this approach may reveal functional divergences between a single field or land type or convergences between different classes as shown in Figs. 2.8 and 2.9, with strong impacts on cascading ecosystem processes.
To identify landscape units using only functional information, we integrated functional attributes by applying an unsupervised classification procedure. In contrast with a LULC classification, we do not expect a priori conceptual scheme, both in terms of the number of classes and their identity. Functional classes often have to be split or merged to create a meaningful map, i.e., to show patterns of patches and corridors rather than isolated pixels (“salt and pepper” appearance). Using the unsupervised clustering algorithm ISODATA
(Jensen 1996), we generated a map delimitating five different classes which reached our pattern-based expectations (Fig. 2.10). This approach revealed functional divergences between single farm types or common lands (e.g., western versus eastern cultivated areas dominated by cash crops), and convergences between different classes (e.g., western mixed shrubs and cultivated land versus eastern cultivated areas dominated by cash crops), with potential impacts on cascading ecosystem processes.
In addition to the landscape analysis, other on-the-ground information is needed for the development of a representative sampling design for smallholder systems before resource-consuming measurements of soil GHG fluxes or soil carbon and nitrogen stock inventories are implemented. The characterization of farmers’ socioeconomic condition is important here, because this also affects resource management. On-farm variations in soil properties, which result from long-term differences in field management, create soil fertility gradients that may justify the use of a field typology.