Application to Zimbabwe - Gea

Remote Sensing of Environment 113 (2009) 1046–1057

Contents lists available at ScienceDirect

Remote Sensing of Environment j o u r n a l h o m e p a g e : w w w. e l s ev i e r. c o m / l o c a t e / r s e

Detection and mapping of long-term land degradation using local net production scaling: Application to Zimbabwe S.D. Prince ⁎, I. Becker-Reshef, K. Rishmawi Geography Department, University of Maryland, College Park, MD 20742-8225, USA

a r t i c l e

i n f o

Article history: Received 12 August 2008 Received in revised form 28 January 2009 Accepted 31 January 2009 Keywords: Dryland degradation Desertification Zimbabwe Communal land Net primary production (NPP) Local NPP scaling (LNS) MODIS Sustainability

a b s t r a c t Degradation of vegetation and soils in drylands, sometimes called desertification, is thought to be a serious threat to the sustainability of human habitation, but maps of the extent and severity of degradation at country and global scales do not exist. Degraded land, by definition, has suffered a change relative to its previous condition set by its climate, soil properties, topography and expectations of land managers. The local net production scaling (LNS) method, tested here in Zimbabwe, estimates potential production in homogeneous land capability classes and models the actual productivity using remotely-sensed observations. The difference between the potential and actual productivities provides a map of the location and severity of degradation. Six years of 250 m resolution MODIS data were used to estimate actual net production in Zimbabwe and calculate the LNS using three land capability classifications. The LNS maps agreed with known areas of degradation and with an independent degradation map. The principal source of error arose because of inhomogeneity of some land capability classes caused by, for example, the inclusion of local hot-spots of high production and differences in precipitation caused by local topography. Agriculture and other management can affect the degradation estimates and careful inspection of the LNS maps is essential to verify and identify the local causes of degradation. The Zimbabwe study found that approximately 16% of the country was at its potential production and the total loss in productivity due to degradation was estimated to be 17.6 Tg Cyr− 1, that is 13% of the entire national potential. Since the locations of degraded land were unrelated to natural environmental factors such as rainfall and soils, it is clear that the degradation has been caused by human land use, concentrated in the heavily-utilized, communal areas. © 2009 Elsevier Inc. All rights reserved.

1. Introduction Land degradation has mainly been studied at a local scale and from the perspective of the farmer and pastoralist, but it also has effects at the country, continental and global scales (Prince, 2002). For example, the cumulative costs of degradation in Zimbabwe through siltation of dams and waterways has been estimated to have a major impact on Gross Domestic Product (GDP) (Gore et al., 1992; Grohs, 1994). Other regional effects include: reduction of food security; disruption of the surface water balance; reduced carbon sequestration and release of carbon through soil erosion; impacts on regional climate through changes in the evaporation ratio, roughness, albedo and increased atmospheric dust loads (Reynolds & Stafford Smith, 2002). Drylands are particularly susceptible to degradation and, although they cover about 41% of Earth's land surface (Safriel & Adeel, 2005), estimates of the extent and severity of degradation vary greatly (Lepers et al., 2005). There are estimates that 70% of the world's drylands are affected and that at least one third of present deserts are

⁎ Corresponding author. Tel.: +1 301 405 4062; fax: +1 301 314 9299. E-mail address: [email protected] (S.D. Prince). 0034-4257/$ – see front matter © 2009 Elsevier Inc. All rights reserved. doi:10.1016/j.rse.2009.01.016

man-made (UNCED, 1993). Set against these high estimates are studies that find no widespread dryland degradation (desertification), at least in the Sahel (e.g. Prince et al., 1998; Niemeijer & Mazzucato, 2002). Nevertheless, most observers agree that there are significant areas of land degradation and these have large effects on the environment and human well-being. In view of the far-reaching consequences of degradation and the large areas that are said to be affected there is a need for inventories and monitoring at the country to global scales using consistent, objective, repeatable, and spatially explicit measures (Prince, 2004). Objective measurement of degradation for large areas has, however, proved extremely difficult, mainly due to multiple criteria and the lack of reliable methods (Verstraete, 1986; Prince, 2002; WMO, 2005). Existing global maps such as GLASOD (Thomas & Middleton, 1992), the USDA NRCS Desertification Vulnerability map (Eswaran & Reich, 2003), the United Nations World Atlas of Desertification (UNEP, 1997) and, more recently, Lepers et al. (2005) all depend on coarse resolution soils maps and indicate vulnerability to degradation, rather than actual degradation. Part of the definition of desertification, or dryland degradation, used by the United Nations (UN) and others is a reduction of the productive potential of the land (Reynolds, 2001). While productivity may be

S.D. Prince et al. / Remote Sensing of Environment 113 (2009) 1046–1057

measured in many ways, the growth of vegetation per unit area per unit time (net primary production, NPP) is used here. The processes that lead to land degradation involve the interaction of environmental, social, economic, and historical factors (Gore et al., 1992; Reynolds et al., 2007) and are typically induced when cropping and grazing exceed this potential which can happen, for example, during droughts. Degradation manifests itself in characteristics other than productivity, for example reduced biodiversity (Pickup, 1998; Adeel et al., 2005). Nevertheless, reduced NPP is a consistent symptom of degradation relative to the potential in the site and can be used as an index whether or not the effect on productivity is the objective. In order to detect degradation, however, a reference value for the non- or less-degraded condition is required (O'Malley & Wing, 2000; Stoms & Hargrove, 2000; Boer & Smith, 2003; Prince et al., 2007), and a fundamental difficulty is to determine that reference value (Wessels et al., 2007). Here the potential NPP, the NPP that would be expected in the absence of human land use, is used as the reference. The magnitude of the difference between actual and potential NPP provides both a quantitative measure of degradation and the associated loss of carbon fixation. Potential NPP can be estimated using mechanistic biogeochemical models but the climate and soils data that are required to drive the models are often available only at a resolution of 0.5°–2° (0.25–4×104 km2), an area that cannot contribute to policy applications at a country scale (Prince, 2002). In order to circumvent the use of coarse resolution climate and soils maps, satellite measurements of net primary production (NPP) are used here to estimate both actual and potential NPP (Prince, 2004). Techniques for measurement of net primary production (NPP) using Earth-observing satellite data were first developed in the mid 1980s (Prince, 1991) but it is only now that a satellite data archive has accumulated with a long enough record (N25 years) to allow degradation studies at appropriate time scales (Prince et al., 2000; Prince, 2002). The remotely sensed data that can be used to model productivity includes MODIS (from 2000 to present), SPOT VEGETATION (1998 to present), SeaWifs (1997 to present), MERIS (1995 to present) and NOAA Advanced Very High Resolution Radiometer (AVHRR, 1981 to present). The purpose of the present work is to assess the Local NPP Scaling (LNS) method (Prince, 2004) for quantification and mapping of degradation in areas from a few km2 to country scales. In LNS multitemporal satellite data are used to calculate the annual NPP of each pixel, then the difference between the potential and actual NPP for each pixel is calculated. Variation in potential NPP can be caused by differences in land use, land cover and physical factors. The variation is reduced by stratification into homogeneous regions (Soriano & Paruelo, 1992; Prince, 2004). The procedure is similar in concept to the use of land classification to determine appropriate uses of land for agriculture or livestock production and for land valuation (FAO, 1976). Within each land capability class (LCC), all pixels are assigned the same potential NPP — the productivity that would have been attained were it not for human factors (Soriano & Paruelo, 1992; Tappan et al., 2004). Without the use of the LCCs, regions having a lower potential production would be confused with degraded areas of high potential. The potential NPP is estimated from the highest NPP found in the LCC to which that pixel belongs. LCCs are derived from climate, soils, land cover and land use, and are independent of actual NPP. In another approach, Residual Trend (RESTREND) analysis (Wessels et al., 2007), the rainfall–NPP relationship is used to obtain the difference between potential and actual NPP using the actual rainfall each year. Both methods model the potential and compare it with the observed NPP but, in RESTREND, the results apply only within the time series of satellite data used (although previous years may influence the starting NPP); LNS, however, should be able to detect preexisting degradation. Methods that have some similarities to LNS have been proposed. Budde et al. (2004) compared the NDVI of a pixel with surrounding pixels using a moving window (31 × 31 km in their Senegal study). The size of the window will affect the result depending on the spatial

1047

patterns of degradation and on whether there are differences in NDVI that are unrelated to degradation. Compared with LNS, this approach will blur abrupt changes in degradation, for example at the boundaries between communal and commercial areas. Another study (Stoms & Hargrove, 2000) used national parks as reference sites to estimate potential NDVI but this is only valid for areas that have the same soils, climate and land use as the park. Lambin and Ehrlich (1997) used a time-series of NDVI to express each year's NDVI relative to the maximum observed for each pixel; while this was intended as a change detection technique, it could be used to monitor active degradation. Boer and Puigdefabregas (2005) estimated potential NPP from NDVI by calculation of the maximum actual evapotranspiration (AET) consistent with the rainfall and potential evapotransiration in each pixel. Unfortunately, as in the case of many more detailed NPP models, the paucity of meteorological stations limits the spatial resolution of the meteorological data to such an extent that key local variations cannot be resolved. While the aim of all these studies was to estimate the potential productivity for comparison with the actual, the use of homogeneous areas (LCC) in LNS clearly has some advantages. LNS is tested here in Zimbabwe (Fig. 1a) where there is a very appropriate, if regrettable in human terms, opportunity to observe land that is indisputably degraded. Degradation was already far advanced in some areas of Zimbabwe in the 1980s (Muller, 1983; Whitlow, 1988), mostly concentrated in areas of subsistence farming, and it has accelerated since that time. These heavily degraded areas have, for most of the 20th century, been occupied exclusively by indigenous peoples (Moyo, 1995). They were established in the colonial period when land was appropriated by Europeans for commercial farming and the indigenous populations were increasingly confined in what became known as “communal”, in contrast with “commercial” (or “general”) and “other” (parks, reserves) lands (Fig.1b, Zimbabwe,1979a). The same tenure system, mostly with similar consequences, is also found in South Africa (Wessels et al., 2008). What started as an inevitable consequence of two divergent forms of land use, became increasingly inequitable as indigenous population densities increased (Fig. 1c) (Roder, 1964; Moyo et al., 2000), ultimately leading to a conspicuous contrast between land degradation in the communal areas and highly productive commercial land. While it is often assumed that communal areas were placed in areas with low land capability, in fact both commercial and communal lands are found in all LCCs (Vincent et al., 1960). Owing to the spatial coherence of these two types of land tenure in Zimbabwe, degradation can be seen even in continental-scale satellite imagery (Fig. 1d). The indisputable difference in degradation between communal and neighboring commercial land that occupy similar environments is used here to test the LNS technique. 2. Methods 2.1. Data Land-cover was obtained from Hansen et al. (2000), precipitation from NOAA (2008), and soils from the SOTER (SOTER, 2002) and the Zimbabwe soil map (Zimbabwe, 1979b). MODIS data at a resolution of 250 × 250 m (6.25 ha) for 2000–2005 were used to estimate NPP. MOD13Q1 16 day normalized difference vegetation index (NDVI) composites of the sums of the NDVI (ΣNDVI) for each southern hemisphere growing season (September–May) were calculated. MODIS has been operational since only 2000, a shorter time series than is available in other data sets, but it was preferred to test LNS owing to its higher spatial resolution and improved radiometric properties. Conclusions reached using MODIS should be applicable to similar sensors. The 30-year average rainfall for Zimbabwe, from gauge data, is 750 mm and, from reanalysis results (GPCP, 2008), for the 10 years from 1997, was 677 mm (s.d. 123 mm). GPCP data were used to

1048


Fig. 1. Maps of Zimbabwe; a – locations of all place names used in text (Map no. 4210 Rev. 1, United Nations, 2004) ; b – administrative districts and land tenure type (white – commercial, black – communal, grey – other); c – Landsat TM mosaic of Zimbabwe with land tenure type boundaries. Note the unusual regional pattern of vegetation in Zimbabwe caused by land cover differences between communal (lighter) and commercial (darker) areas; d – population density in 1982 (Zimbabwe, 1982) with land tenure type boundaries. Note higher population in Communal areas.

calculate these climatological values since NOAA (2008) were not available for 30 years. The average GPCP annual rainfall for 2000–2005 was 679 mm (s.d. 129 mm), consisting of 1 year with a high total (991 mm) and several years below the 30-year average. Thus the average and interannual variability of rainfall in the study period were not unusual. NDVI was converted to NPP using the products of the CASA model (Imhoff et al., 2004). The LNS results were assessed in various comparisons with independent data. These included comparison of the LNS LCC map with the Agro-Ecological Survey map of Natural Regions and Areas in Zimbabwe (Vincent et al., 1960). This survey used soils, rainfall and topography alone, not considering current NPP, and is therefore directly comparable with the LCCs used here. LNS potential NPP was compared with CENTURY model results (Parton et al., 1993; Cramer et al., 1999). Landsat ETM+ bands 7, 4, and 2 false color composites were made from GeoCover Orthorectified Landsat Compressed Mosaics (MDA_Federal,

2000) and were used in visual comparisons with the LNS map, as was the Zimbabwe National Land Degradation Survey (Whitlow, 1988). 2.2. Land capability classification (LCC) The LCCs were defined by stratification of the digital maps of rainfall, soils and land use. The classes were derived using a k-prototypes clustering technique (Huang, 1997; Hargrove & Hoffman, 2004), which is an extension of the k-means method but with the added advantage that it can cluster large data sets consisting of both numeric and categorical variables. The steps were as follows: i. Digital soil and land cover maps were gridded to 250 × 250 m cells, each cell geographically registered with the appropriate MODIS 250 × 250 m pixel.


1049

Fig. 2. k-prototype stratifications of Zimbabwe and potential NPP in each class derived using the Local NPP Scaling method; a — LCC from the Zimbabwe soil map (Zimbabwe, 1979b) and rainfall (NOAA, 2008) (ZSOL–PPT), b — using land cover (Hansen et al., 2000) and rainfall (LULC–PPT). Box in a shows area enlarged in b.

ii. Three LNS analyses were carried out using different combinations of variables: precipitation with land cover (LULC–PPT); SOTER soils map classes with precipitation (SOTR–PPT); and Zimbabwe soils map classes with precipitation (ZSOL–PPT). iii. The number of initial LCC clusters was defined using all permutations of the soils (31classes for the SOTER, and 27 for the Zimbabwe soil map), land cover (7 classes), and four equal ranges of rainfall. The number of rainfall classes was arbitrary, based on an approximate target number of final LCCs. Thus there were 28 initial clusters for LULC–PPT, 124 for SOTR–PPT, and 108 for ZSOL–PPT. iv. k-prototype classification of all grid cells was carried out using Euclidean distances for numeric attributes and the number of mismatches between pixel values for categorical attributes (Huang, 1997). The weighting function in the classification was used to weight the categorical and numeric variables equally. The prototype properties were recalculated after each allocation to clusters and, after all the pixels were allocated, all cells were tested in relation to the updated classification and those

found to be closer to a different prototype were reallocated. This procedure was repeated recursively until no more cells changed clusters or until the iterations reached a pre-set limit (Huang, 1997), whichever was reached first. v. The final k-prototypes clusters were classified using a decision tree (Breiman, 1984) to create definitions of the cluster in terms of the input data. Unreasonable classes were removed by pruning the tree, which led to fusion of some classes.

2.3. Measurement of net primary production (NPP) using satellite data The ΣNDVI for each growing season was used as a surrogate for NPP (Prince, 1991), since there is a near-linear relationship between NPP and ΣNDVI in tropical grassland, cropland and sparse woodland and light use efficiency has been shown not to improve accuracy (Fensholt et al., 2006). In order to estimate NPP in terms of carbon, a simple scaling of annual ΣNDVI into NPP was made by regressing the

Fig. 3. Effectiveness of the three land capability classifications (LCC) as indicated by potential NPP from the Century model (Parton et al., 1993; Cramer et al., 1999). The classes for the three LCC are shown: light grey is the PPT–LULC, dark grey, the PPT–ZSOL; and black, the PPT–SOTR classifications. The size of the symbol indicates the coefficient of variation within each class and the bars indicate ± 1 standard deviation.

1050


ΣNDVI on NPP from the CASA model (Imhoff et al., 2004), thus the NPP used here is a simple linear transform of ΣNDVI.

correlations and also by comparison with the potential NPP from CENTURY (Parton et al., 1993; Cramer et al., 1999).

2.4. Local net primary production scaling (LNS)

2.5.3. Local NPP scaling (LNS) The three LNS maps were compared with independent maps of degraded areas.

The potential and minimum NPP of each LCC class were estimated by the 90th and 5th percentiles of the frequency distribution of ΣNDVI, respectively. The results of parametric and non-parametric estimates of the percentiles were similar and so the parametric frequency distribution was used. The reduction of the actual 5-year interannual mean NPP below the potential NPP was calculated by subtraction of the actual from the potential, and this negative quantity is the LNS metric of degradation. This procedure assumes that there is land in its potential state within each LCC class: if not, the differences between actual and potential NPP and the degree of degradation would be underestimated. The differences between potential and actual productivity of each pixel provided a map of the impact of degradation on productivity — that is the NPP lost as a result of degradation. The percentage of the potential NPP of each pixel can also be used, but each percentage value applies only to the LCC to which it belongs. 2.5. Assessment 2.5.1. Land capability classification (LCC) The LCC was assessed in three ways: first, by estimating the extent to which the LCCs reduced the correlation between the environmental factors that were used in the classification; second, by comparison of the ranking of the LCCs with potential NPP and the rankings using two independent estimates of NPP — rainfall and interannual mean ΣNDVI; and, third, by comparison of the LCCs with the Agro-Ecological Survey map of Natural Regions and Areas in Zimbabwe. 2.5.2. Potential net primary production (NPP) The maps of potential NPP were assessed by comparison with soils and land use (communal or commercial) in order to detect any

i. LNS results in areas that had a wide range of LNS values were compared visually with Landsat images. The high spatial resolution of Landsat (30 m) allowed qualitative interpretation of some aspects of land use and land cover and any relationships with the modeled degradation. ii. Comparison was made with the Zimbabwe National Land Degradation Survey (Whitlow, 1988) which was derived from the density of erosion features shown by aerial photographs. The survey assessed the frequency of erosion features, an aspect of degradation that is independent of observations of NPP. The aerial photographs were mainly acquired in 1980, although some were for as early as 1979 and others as late as 1984. The survey map is gridded at 0.5°× 0.5° and has six classes: no erosion (class 1); 0.1–4% (class 2); then in 4% steps (classes 3–5); and N16% (class 6). Although the survey predated the MODIS satellite data by approximately 20 years, persistence of degraded conditions is a key aspect of desertification (Prince, 2002; Wessels et al., 2004) thus coincidences of severely eroded areas in 1980 and low LNS values in 2000–5 were a test of the skill of the LNS approach in detecting pre-existing degradation. In addition to visual comparison, the means and frequency distributions of LNS for each National Land Degradation Survey class were calculated and the values for each class compared. iii. The user accuracy of prediction of degradation by LNS for each LCC was calculated by logistic regression in order to predict the binomial probability of the identification of degraded or not degraded (Knoke et al., 2002). Two independent measures of degradation were predicted: first, the communal (assumed to

Table 1 Similarities in rainfall between the land capability classes (LCC) derived from land cover and precipitation (LULC–PPT).

The density of the fill in each cell indicates the probability of similarity in mean rainfall between the row and column classes: white, p b 0.001; light gray, p b 0.05; dark grey, p N 0.05; black, self-comparisons. Filled row and column labels indicate classes with multiple similarities with other cells that could have been fused.


be degraded) and commercial (assumed not to be degraded) land uses (Zimbabwe, 1979a) and, second, the National Degradation Survey degradation classes grouped into two ranges (1–2 and 3–6). The logistic regression results were mapped to allow assessment of the spatial variability in predictive ability of LNS. iv. The LNS results were compared with the use of NDVI alone, without the application of any further processing. Logistic regression of ΣNDVI on the two independent degradation metrics (iii) was used to estimate its user accuracy. 3. Results 3.1. Land capability classification (LCC) At the countrywide scale all three LCCs stratifications produced similar maps but with local differences (Fig. 2). SOTR–PPT (not shown) gave similar results to the ZSOL–PPT. LULC–PPT had much smaller polygons than those based on the soils which tend to be more coherent. There were 28 final classes for ZSOL–PPT, 124 for SOTR–PPT and 124 for ZSOL–PPT. The degree to which the three LCCs reduced within-class variation was assessed by ranking the classes according to potential NPP from the CENTURY model (Parton et al., 1993) (Fig. 3) and with rainfall. The differences between the LCCs that used soil (ZSOL–PPT and SOTR– PPT) were smaller than in the case of land cover (LULC–PPT), nevertheless, all three classifications gave satisfactory results. In a significant ranges test of LULC–PPT (Table 1) the majority of classes had statistically significant differences in rainfall. A few classes had multiple similarities with other classes and could have been fused. The three LCCs were compared with the Zimbabwe Natural Regions map (Vincent et al., 1960), an independent classification of land suitability. All three LCCs placed most of the 22 Natural Regions in the same rank order of productivity as the Natural Regions map (Fig. 4). 3.2. Estimation of potential net primary production (NPP) Pixels above the 90th percentile of a LCC class were found in both commercial and communal areas. There was little visual correlation

1051

between potential NPP and land tenure type (communal, commercial), nor with rainfall, but there was a strong correlation of both with actual NPP. Although lower potential areas were more commonly found in communal than commercial areas (Fig. 5), there were sizable areas of both in all potential production classes. Communal areas, however, are generally at lower altitudes than commercial land. The National Land Degradation Survey map (Whitlow, 1988) showed no correspondence of degraded land and natural regions. Similarly the potential NPP maps show no association between communal land use and the climate, soils and terrain slope constraint classes in the Global Agro-Ecological Zone map (Fischer et al., 2000; Plate 28). Thus the association of degradation with communal land and the abrupt changes in vegetation cover across commercial–communal boundaries showed that communal lands are not degraded simply because they have low potential. There were some differences in environmental factors that were not removed by stratification, for example, in some LCCs the highest NPP occurred on locally higher land, presumably associated with higher rainfall or unusable land. As a result the potential NPP was set unrealistically high for other parts of the class with lower rainfall. It might be argued that fertilized croplands and especially irrigation also set the potential too high and so over-estimate degradation, but these management treatments may also be regarded as the potential for cropped land. A comparison of potential NPP estimated by the LNS with NPP from CENTURY (Parton et al., 1993) gave a much higher correlation than with the actual NPP (Table 2). Since the model can only estimate productivity of undisturbed land, that is potential NPP, this confirms that the LNS technique provided reasonable estimates of potential NPP. Among the three LNS analyses, LULC–PPT accounted for the most variation in NPP. 3.3. Local NPP scaling There was general agreement between the LNS maps calculated for SOTR–PPT (Fig. 6) and ZSOL–PPT at both the country and local scales. The LNS map shows some very clear patterns in the location of degradation, similar to those reported by Prince (2004). Land in good condition relative to its potential was found in the commercial areas:

Fig. 4. Potential production (g− 2 year− 1) in the three land capability classifications (LCC) compared with the potential for agricultural production from the Zimbabwe Natural Regions and Areas map (Vincent et al., 1960). The potential of the Natural Regions for agricultural production is indicated by a numeral followed by letter, 1A (maximum) to 5A (minimum). Region XX is omitted since it is unsuitable for any use other than reserves.

1052


Fig. 5. The area of land allocated to commercial and communal use in the five main Natural Regions as a percentage of the total land area occupied by each tenure type. Natural Regions are ranked according to their potential productivity from highest to lowest. Modified from map of Natural Regions and Areas in Vincent et al. (1960).

along the broad SW to NE sweep of the higher land from Plumtree on the Botswana border (see Fig. 1 for place names), to Bulawayo through Gweru and Chivhu and north to Harare; from Gweru northeast through Marondera; from Harare NW through Chinhoy and Karoi; and E–W from Gwanda (Matabeleland South) to Chiredzi (Masvingo). Extensive parks and reserves were also close to potential, for example: the large Hwange reserve S and W of Dete (Matabeleland North); Gonarezhou along the SE border with Mozambique; and below the Zambezi escarpment in the northern part of Mashonaland West between Chirundu and Zumbo, including Mana Pools National Park. In marked contrast, communal land was almost everywhere degraded, often in striking contrast to neighboring commercial land. Particularly severe examples were: in the Save catchment in Manicaland; an area centered on Mutoko (Mashonaland East); along the Shashe river (Matabeleland South); and centered on Gokwe (Midlands). The LNS maps were remarkably similar to the National Degradation Survey country-wide degradation map (Whitlow, 1988) (Fig. 7), despite the difference in methods and the 20-year time difference. Areas mapped as eroded in 1980 almost everywhere had low LNS values. For example: in the Save catchment, around Mutoko (Mashonaland East); along the Shashe (Matabeleland South), east of Chegutu (Mashonaland West) and around Gokwe (Midlands) were all identified in the LNS maps. A detailed comparison of the LNS and National Degradation Survey maps provides some indication of trends in degradation between 1980 and 2000. The mean LNS values for each National Degradation Survey class were calculated and the mid values between the means were used to classify the LNS percentages into six classes (Fig. 8). Increases, decreases or no change in each National Degradation Survey class grid cell were labeled. Because the two maps are based on quite distinct measures of degradation, a comparison depends on the assumption that the absolute values and ranges of the two were unchanged over the 20 years. With this qualification, there were areas of reduced degradation but much larger areas of increase. The degree of agreement of the LNS and the Degradation Survey map was highly significant (p b 0.01 of difference). The LNS values declined with increasing erosion in the Degradation Survey (Fig. 9a) and in land use classes in the order commercial, parks and communal (Fig. 9b). In contrast to the LNS, the annual mean ΣNDVI showed no

such decline across Degradation Survey classes or the land use categories (Fig. 9a,b), and was therefore less able to detect degradation. Several areas of Zimbabwe that were very little affected by erosion in 1980 had very large negative LNS values in 2000 (Fig. 9c). While the LNS and degradation maps measure different properties of the land, the greater area identified by the LNS may also indicate some extension of degradation over the 20-year gap between the aerial photography used by the National Land Degradation Survey and the MODIS data used for LNS. This speculation is supported by the skewed distributions of LNS in the higher erosion classes (Fig. 9c). For each LCC class the accuracy of the LNS designation of degradation was calculated using logistic regressions. Two independent criteria for degraded/not degraded were used, communal/not communal, and the National Degradation Survey classification divided into two ranges, 1–2 and 3–6. The results indicate that the user accuracy of prediction of degradation, judged by comparison with the Survey, was: SOTR–PPT 73%, ZSOL–PPT 73% and LULC–PPT 64% (Fig. 10). For comparison, the accuracy of prediction of degradation in the two Survey class ranges using ΣNDVI values alone, without stratification by the LCC, was 69%.

Table 2 Comparison of LNS with potential NPP from an independent global model that uses climate forcing data alone (Parton et al., 1993; Cramer et al., 1999). NPP

Land capability classification

Correlation with independent measure of potential NPP (r2)

LNS potential NPP

Zimbabwe soil map and precipitation (ZSOL–PPT) classes SOTR soil map and precipitation (SOTR–PPT) classes Land cover and precipitation (LULC–PPT) classes All pixels in Zimbabwe

46.1%

16,815

49.1%

16,502

55.0%

16,967

22.5%

610,514

Remotely sensed (actual) NPP

Degrees of freedom

The correlation between the potential NPP from the global model and potential NPP calculated with LNS was higher for all three LNS analyses than it was with the actual NPP.


Fig. 6. Local NPP Scaling (LNS) of Zimbabwe using the ZSOL soils map and precipitation (ZSOL–PPT) land capability classification. Communal and Commercial area boundaries shown in black. Inset, higher resolution segment SW of Gweru showing communal area degradation (top left) and commercial area degradation (lower right).

Qualitative, visual comparisons of the LNS maps with Landsat (MDA_Federal, 2000) generally showed coincidence of the boundaries of communal lands with low LNS (Fig. 11). The higher spatial resolution of the Landsat data (30 m) showed innumerable examples of stark contrasts in land condition across commercial–communal boundaries, often separated by only a fence line. It was very clear at

1053

Fig. 8. Change in degradation between 1980 Zimbabwe land degradation survey (Whitlow, 1988) and 2000 LNS degradation map.

this fine scale that low LNS values were associated with sparse vegetation in communal land. In addition to identifying the degraded parts of the communal areas, the LNS maps suggested variation in the degrees of degradation within communal areas, and also identified degraded areas in commercial land (Fig. 6). In a vegetation change study in Buhera District (Manicaland) Mambo and Archer (2007) used Landsat data to identify areas that

Fig. 7. Comparison of: a — LNS with SOTR soils map and precipitation (SOTR–PPT); and b — National Land Degradation Survey (Whitlow, 1988). Legend indicates the LNS values in units of loss of NPP and the degradation class in the National Survey.

1054


had become more degraded over the period from 1992 to 2002. Interestingly, these change areas coincide closely with areas where LNS indicated less degradation. Two suggestions are offered that may, individually or together, account for this apparent paradox. First, the Buhera study measured vegetation change, while LNS measures static vegetation condition. Second, it may be that the areas identified as severely degraded by LNS were not susceptible to further degradation, rather only in the areas where degradation was moderate could further decline take place. This latter point is speculative and further study of this important comparison is warranted. 4. Discussion Land capability classification to create classes of uniform productive potential is fundamental to LNS. The same concept is widely used in land evaluation (FAO, 1976; McRae & Burnham, 1981; Wessels et al., 2004), however, there is no global, comprehensive and consistent method for definition of land capability. Rainfall and land cover have

been used for LNS (Prince, 2002) and, since soils are an important determinant of productivity and risk of degradation (Eswaran et al., 1997, 1999), soil type was added here. Residual inhomogeneities in LCC classes may cause errors in estimation of the potential NPP, as was found in some cases in Zimbabwe. Similar problems are likely if classes contain more and less productive soils or highly productive “hot-spots” such as wetlands, riparian features, irrigated or fertilized crops. The effect of such hotspots is that actual productivity is scaled according to the potential set by unrepresentative parts of the class, thus overestimating the degradation. By definition, unrepresentative features are small and the effect is greatly reduced if the individual satellite observations used to calculate NPP are large relative to these features. Potential NPP estimation for LCC classes using the productivity of the upper 90th percentile assumes that each class contains some non-degraded areas. If this is not the case, degradation will be underestimated. The selection of the 90th percentile was arbitrary and a formal test of the effect of changing this would be useful, however, the main purpose in

Fig. 9. LNS values for three different stratifications of Zimbabwe; a — mean of the three LNS analyses (SOTR–PPT, ZSOL–PPT, and LULC–PPT) in each National Degradation Survey class, and mean ΣNDVI for comparison, b — LNS and NDVI values in three land use classes, c — National Degradation Survey and LNS scores of municipal wards (4th level administrative units; 1200 total). The number of municipal wards falling into each of 5 equal ranges of the National Degradation Survey (Whitlow, 1988) and, within each range, the numbers of municipalities in five ranges of ZSOL–PPT LNS values.

S.D. Prince et al. / Remote Sensing of Environment 113 (2009) 1046–1057 Table 3 Effect of degradation on country-wide net primary production (NPP) in Zimbabwe for local NPP scaling applied to three land capability classifications; a — mean loss of NPP in g cm− 2 year− 1) below potential and statistics, b — percentage losses and at three levels (0–1, 1–2, N2 standard deviations). a. Land capability classification

Loss of net primary production compared with potential (g cm− 2 year− 1)

LULC–PPT SOTR–PPT ZSOL–PPT

Mean

Median

Mode

Standard deviation

57.4 58.8 60.9

51.1 50.0 55.0

34.6 33.3 40.8

38.3 40.0 39.2

b. Land capability classification

LULC–PPT SOTR–PPT ZSOL–PPT

At potential NPP (% country)

18.1 16.2 16.1

Below potential NPP (% country)

0–68% (0–1 std) below

68–95% (1–2 std) below

N 95% (2 std) below

31.2 31.8 29.7

29.7 (60.9) 30.2 (62.1) 29.5 (59.2)

21.0 (81.9) 21.8 (83.8) 24.8 (84.0)

Figures in brackets are the cumulative percentages.

the use of the upper frequency range rather than the maximum value, was to minimize the effects of extreme outliers. In the application to Zimbabwe, the mean of the annual sums of per-pixel NDVI were averaged over 5 years. There are other possibilities, for example averaging annual LNS values. Factors other than degradation may also reduce or increase productivity, such as removal of woodland for agriculture or other management that is not incorporated in the derivation of the LCC, or enhancement of productivity by run-on in addition to local rainfall. Careful inspection of the results can mitigate these problems and the LCC can and should be adapted to the specific purpose of the analysis. LNS focuses attention on areas that may contain degradation, but detailed investigation is needed to confirm the diagnosis and assign a cause (Botha, 2000). k-prototypes clustering allowed for a consistent, quantitative, and rigorous classification scheme, using a combination of numeric and categorical data. The computation resources needed, however, were large, mainly because of the iterative reassignment of pixels (Hargrove & Hoffman, 2004). The implementation of the algorithm for this application could be improved by adding functions to control the number of classes produced when multiple data layers are used and to prevent the formation of very small classes.

1055

In LNS, the productive potential of the land cover is the reference for assessment of the degree of degradation. Productive potential, however, can be defined in different ways. To be useful for land management, the potential is relative to the specific land use, not the productivity of the pristine, natural vegetation cover, sometimes called potential vegetation. In applications to degradation the use of potential, natural vegetation, in many cases, would map all agricultural land as degraded. In the case of agriculture the most productive areas are generally fertilized and possibly irrigated, and the selection of these as the estimator of potential NPP in LNS is not necessarily inappropriate in a predominantly agricultural region, it is in fact indicative of the maximum NPP of the LCC. LNS could be adapted for other applications by selection of different values of the potential NPP. For example, in dynamic global vegetation models where disturbance is not accounted for, the potential natural vegetation in the absence of land cover alteration, past or present, is appropriate. Clearly the choice of a reference, potential NPP must match the particular application. The choice in selection of a definition for “potential” might seem to be a flaw, but it can provide flexibility in application of LNS. This study in Zimbabwe found severe degradation in large areas of the country: only approximately 16% was at its potential NPP and over 80% was up to two standard deviations of the LNS frequency distribution (95%) below the potential (Table 3). An advantage of the use of NPP to detect degradation is that lost production can be expressed directly in units of loss of carbon fixation per year. For Zimbabwe the loss in 2000–5 was estimated to be 17.6 Tg Cyr− 1 of the entire national potential NPP (132.5 Tg Cyr− 1) giving an actual NPP of 115.0 Tg Cyr− 1. It is interesting to note that (Imhoff et al., 2004) estimate that 24.5 Tg Cyr− 1 is appropriated for human use in Zimbabwe, approximately the same as is lost by degradation. Since the locations of land with low NPP relative to the potential were unrelated to environmental factors such as rainfall and soils (and land cover in LUCC–PPT), it is clear that the degradation has been caused by human land use, concentrated in communal areas. Nevertheless, the proportion of NPP lost (~13%) is quite small in view of the appearance of the communal lands, but it agrees with Scoones (1992) finding that there was no detectable reduction in livestock production in Zimbabwe's communal areas. This assessment of LNS was made possible by the large areas of indisputable degradation in Zimbabwe and the stark contrast between degraded and non-degraded land. The importance of nondegraded reference sites has been recognized by the Committee for the Review of the Implementation of the United Nations Convention to Combat Desertification (UNCCD) who emphasized the need for “quantifiable and readily verifiable benchmarks” (Convention to Combat Desertification, 2002). Where similar circumstances occur,

Fig. 10. User accuracy of three LNS analyses in prediction of degradation. Degradation was determined by presence of communal land. Accuracies in percentages; a — SOTR–PPT, b — ZSOL–PPT, c — LULC–PPT.

1056


Fig. 11. Visual comparison of Landsat (a) and LNS (b) for Chiwundura, 20 km N of Gweru, Zimbabwe. Communal area and surrounding commercial areas are all in natural region IIIB (Semi-intensive livestock production: with support from small grain crop production). Lighter tones represent brightness for Landsat and degradation in the LNS image. Note overall uniformity of the communal area shown by Landsat (a), but differences in the degree of degradation indicated by LNS (b). Landsat data from GeoCover 2000 true color composite (MDA Federal 2000).

such as in the northern provinces of South Africa (Wessels et al., 2004), similar tests may be undertaken. The establishment of a network of such paired sites throughout global drylands would be an important step towards UNCCD's goals. Acknowledgement This work was partly supported by NASA grant NAG5 9329 and the World Food Programme. References Adeel, Z., Safriel, U., Niemeijer, D., & White, R. (Eds.). (2005). Millennium ecosystem assessment. Ecosystems and human well-being: Desertification synthesis. Washington, D.C.: World Resources Institute. Boer, M. M., & Puigdefabregas, J. (2005). Assessment of dryland condition using spatial anomalies of vegetation index values. International Journal of Remote Sensing, 26, 4045−4065. Boer, M., & Smith, M. S. (2003). A plant functional approach to the prediction of changes in Australian rangeland vegetation under grazing and fire. Journal of Vegetation Science, 14, 333−344. Botha, J. H. (2000). An assessment of land degradation in the Northern Province from satellite remote sensing and community perception. South African Geographical Journal, 82, 70−79. Breiman, L. (1984). Classification and regression trees. Belmont, California: Wadsworth International Group. Budde, M. E., Tappan, G., Rowland, J., Lewis, J., & Tieszen, L. L. (2004). Assessing land cover performance in Senegal, West Africa using 1-km integrated NDVI and local variance analysis. Journal of Arid Environments, 59, 481−498. Convention to Combat Desertification (2002). Recommendations and conclusions of the African regional conference preparatory to the first session of the Committee for the Review of the Implementation of the United Nations Convention to Combat Desertification (UNCCD–CRIC1), Windhoek, Namibia: Secretariat of the Convention to Combat Desertification, 12 pages. Cramer, W., Kicklighter, D. W., Bondeau, A., Moore, B., Churkina, C., Nemry, B., et al. (1999). Comparing global models of terrestrial net primary productivity (NPP): Overview and key results. Global Change Biology, 5, 1−15. Eswaran, H., Almaraz, R., van den Berg, E., & Reich, P. (1997). An assessment of soil resources of Africa in relation to productivity. Geoderma, 77, 1−18. Eswaran, H., & Reich, P. (2003). World soil resources map index. Web page available online on http://www.nrcs.usda.gov/technical/worldsoils/mapindx/#regional: United States Department of Agriculture, Natural Resources Conservation Service. Accessed 3 June 2003. Eswaran, H., Reich, P., & Beinroth, F. (1999). Global desertification tension zones. Sustaining the global farm, Purdue University and the USDA–ARS National Soil Erosion Research Laboratory.

FAO (1976). A framework for land evaluation, Soils Bulletin 32, M-51. Soil Resources Development and Conservation Service, Land and Water Development Division Rome: Food and Agriculture Organization of the United Nations. Fensholt, R., Sandholt, I., Rasmussen, M. S., Stisen, S., & Diouf, A. (2006). Evaluation of satellite based primary production modelling in the semi-arid Sahel. Remote Sensing of Environment, 105, 173−188. Fischer, G., van Velthuizen, H., Nachtergaele, F., & Medow, S. (2000). Global agroecological zones (Global-AEZ). Web page available on-line on http://www.fao.org/ ag/AGL/agll/gaez, Rome & Laxenberg: Food and Agriculture Organization of the United Nations (FAO) and International Institute for Applied Systems Analysis (IASA). Accessed January 12 2007. Gore, C., Katerere, Y., Moyo, S., Mhone, G., Mazambani, D., Ngobese, P., et al. (1992). The case for sustainable development in Zimbabwe: Conceptual problems, conflicts, and contradictions, report prepared for the United Nations Conference on Environment and Development (UNCED), 159pp. Harare: Environmental and Development Activities (ENDA–Zimbabwe) and Regional Network of Environmental Experts (ZERO). GPCP (2008). Global analyses of monthly precipitation derived from satellite and surface measurements. Web page available on-line on http://lwf.ncdc.noaa.gov/oa/wmo/ wdcamet-ncdc.html, Asheville, NC: Global Precipitation Climatology Project , NOAA World Data Center for Meteorology, National Climatic Data Center. Accessed 24 July. Grohs, F. (1994). Economics of soil degradation, erosion and conservation: A case study in Zimbabwe. Kiel: Wissenschaftsverlag Vauk. Hansen, M. C., DeFries, R., Townshend, J. R. G., & Sohlberg, R. (2000). Global land cover classification at 1 km spatial resolution using a classification tree approach. International Journal of Remote Sensing, 21, 1331−1364. Hargrove, W. W., & Hoffman, F. M. (2004). Potential of multivariate quantitative methods for delineation and visualization of ecoregions. Environmental Management, 34, S39−S60. Huang, Z. (1997). Clustering large data sets with mixed numeric and categorical values. First Pacific Asia Knowledge Discovery and Data Mining Conference. Singapore: World Scientific. Imhoff, M. L., Bounoua, L., Ricketts, T., Loucks, C., Harriss, R., & Lawrence, W. T. (2004). Human appropriation of net primary productivity (HANPP). Web page available online on http://sedac.ciesin.columbia.edu/es/hanpp.html: Socioeconomic Data and Applications Center (SEDAC). Accessed 25 July. Knoke, D., Bohrnstedt, G. W., & Mee, A. P. (2002). Statistics for social data analysis, 4th ed : Thomson Wadsworth. Lambin, E. F., & Ehrlich, D. (1997). Land-cover changes in sub-saharan Africa (1982–1991): Application of a change index based on remotely sensed surface temperature and vegetation indices at a continental scale. Remote Sensing of Environment, 61, 181−200. Lepers, E., Lambin, E. F., Janetos, A. C., DeFries, R., Achard, F., Ramankutty, N., et al. (2005). A synthesis of information on rapid land-cover change for the period 1981– 2000. BioScience, 55, 115−124. Mambo, J., & Archer, E. (2007). Assessment of land degradation in the Save catchment of Zimbabwe. Area, 39, 380−391. McRae, S. G., & Burnham, C. P. (1981). Land evaluation. Oxford: Clarendon Press. MDA_Federal (2000). GeoCover. Web page available on-line on http://glcf.umiacs.umd. edu/portal/geocover/edition.shtml: Accessed 19 January 2009 from Global Land Cover Facility, University of Maryland. Moyo, S. (1995). The land question in Zimbabwe. Harare (Zimbabwe): SAPES Trust. Moyo, S., Rutherford, B., & Amanor-Wilks, D. (2000). Land reform & and changing social relations for farm workers in Zimbabwe. Review of African Political Economy, 27,181−202.

S.D. Prince et al. / Remote Sensing of Environment 113 (2009) 1046–1057 Muller, T. (1983). A case for a vegetation survey in a developing country based on Zimbabwe. Bothalia, 14, 721−723. Niemeijer, D., & Mazzucato, V. (2002). Soil degradation in the West African Sahel how serious is it? Environment, 44, 20−31. NOAA (2008). FEWS-NET/MFEWS/AFN Data Archive. RFE1.0 10-Day Rainfall Estimates. Web page available on-line on http://www.cpc.ncep.noaa.gov/products/fews/ data.shtml: NOAA/ National Weather Service National Centers for Environmental Prediction Climate Prediction Center. Accessed 24 July 2008. O'Malley, R., & Wing, K. (2000). Forging a new tool for ecosystem reporting. Environment, 42, 20. Parton, W. J., Scurlock, J. M. O., Ojima, D. S., Gilmanov, T. G., Scholes, R. J., Schimel, D. S., et al. (1993). Observations and modeling of biomass and soil organic matter dynamics for the grassland biome worldwide. Global Biogeochemical Cycles, 7, 785−809. Pickup, G. (1998). Desertification and climate change — The Australian perspective. Clomate Research, 11, 51−63. Prince, S. D. (1991). A model of regional primary production for use with coarseresolution satellite data. International Journal of Remote Sensing, 12, 1313−1330. Prince, S. D. (2002). Spatial and temporal scales of measurement of desertification. In M. Stafford-Smith & J. F. Reynolds (Eds.), Global desertification: Do humans create deserts? (pp. 23−40). Berlin: Dahlem University Press. Prince, S. D. (2004). Mapping desertification in southern Africa. In G. Gutman, A. Janetos, C. O. Justice, E. F. Moran, J. F. Mustard, R. R. Rindfuss, D. Skole, & B. L. TurnerII (Eds.), Land change science: Observing, monitoring, and understanding trajectories of change on the Earth's surface (pp. 163−184). Dordrecht, NL: Kluwer. Prince, S. D., Brown de Colstoun, E., & Kravitz, L. (1998). Evidence from rain use efficiencies does not support extensive Sahelian desertification. Global Change Biology, 4, 359−374. Prince, S. D., Goward, S. N., Goetz, S., & Czajkowski, K. (2000). Inter-annual atmosphere– biosphere variation: Implications for observation and modeling. Journal of Geophysical Research Atmospheres, 105, 20,055−20,063. Prince, S. D., Wessels, K. J., Tucker, C. J., & Nicholson, S. E. (2007). Desertification in the Sahel: A reinterpretation of a reinterpretation. Global Change Biology, 13, 1308−1313. Reynolds, J. F. (2001). Desertification. In S. Levin (Ed.), Encyclopedia of Biodiversity (pp. 61−78). San Diego, CA: Academic Press. Reynolds, J. F., Smith, D. M. S., Lambin, E. F., Turner, B. L., II, Mortimore, M., Batterbury, S. P. J., et al. (2007). Global desertification: Building a science for dryland development. Science, 316, 847−851. Reynolds, J. F., & Stafford Smith, M. (Eds.). (2002). Global desertification: Do humans create deserts? Berlin: Dahlem University Press. Roder, W. (1964). The division of land resources in Southern Rhodesia. Annals of the Association of American Geographers, 54, 41−52. Safriel, U., & Adeel, Z. (2005). Dryland systems. Ecosystems and human well-being: Current state and trends (pp. 948). Washington, D.C.: Island Press.

1057

Scoones, I. (1992). Land degradation and livestock production in Zimbabwe's communal areas. Land Degradation and Rehabilitation, 3, 99−114. Soriano, A., & Paruelo, J. M. (1992). Biozones — Vegetation units defined by functional characters identifiable with the aid of satellite sensor images. Global Ecology and Biogeography Letters, 2, 82−89. SOTER (2002). Soil and terrain database for Southern Africa (SOTRSAF) 1:2.5 million, version 1.0. Rome: FAO Web page available on-line on http://www.fao.org/ag/agl/ agll/soter.stm. Accessed 8 August 2008. Stoms, D. M., & Hargrove, W. W. (2000). Potential NDVI as a baseline for monitoring ecosystem functioning. International Journal of Remote Sensing, 21, 401−407. Tappan, G. G., Sall, M., Wood, E. C., & Cushing, M. (2004). Ecoregions and land cover trends in Senegal. Journal of Arid Environments, 59, 427−462. Thomas, D.S.G., & Middleton, N.J. (Eds.) (1992). World atlas of desertification. London: United Nations Environment Programme, Edward Arnold. UNCED (1993). Agenda 21: Earth Summit — The United Nations programme of action from Rio: Managing fragile ecosystems: Combating desertification and drought. United Nations Conference on Environment and Development (pp. 294). UNEP (Ed.) (1997). World atlas of desertification. London & New York: Arnold & Wiley, on behalf of UNEP. Verstraete, M. M. (1986). Defining desertification: A review. Climatic Change, 9, 5−18. Vincent, V., Thomas, R. G., & Staples, R. R. (1960). An agricultual survey of Southern Rhodesia: Part I agro-ecological survey. Salisbury: Federation of Rhodesia and Nyassaland. Wessels, K. J., Prince, S. D., Frost, P. E., & van Zyl, D. (2004). Assessing the effects of human-induced land degradation in the former homelands of northern South Africa with a 1 km AVHRR NDVI time-series. Remote Sensing of Environment, 91, 47−67. Wessels, K. J., Prince, S. D., Malherbe, J., Small, J., Frost, P. E., & VanZyl, D. (2007b). Can human-induced land degradation be distinguished from the effects of rainfall variability? A case study in South Africa. Journal of Arid Environments, 68, 271−297. Wessels, K. J., Prince, S. D., & Reshef, I. (2008). Mapping land degradation by comparison of vegetation production to spatially derived estimates of potential production. Journal of Arid Environments, 72, 1940−1949. doi:10.1016/j.jaridenv.2008.05.011. Whitlow, R. (1988). Land degradation in Zimbabwe, 62+Appendices. Harare: Department of Natural Resources, Government of Zimbabwe/Department of Geography, University of Zimbabwe. WMO (2005). Climate and land degradation. Geneva: World Meteorological Organization. Zimbabwe (1979a). Land classification. Harare, Zimbabwe: Department of the Surveyor General. Zimbabwe (1979b). Provisional soil map of Zimbabwe–Rhodesia. Edition 2. Salisbury, Zimbabwe Surveyor-General, Chemistry and Soil Research and Specialist Services Section, Pedology and Soil Survey. Zimbabwe (1982). Map of rural population density: August 1982. Based on enumeration areas. Harare: Zimbabwe Surveyor-General, Central Statistical Office.