Remote sensing-derived national land cover land use maps: a comparison for Malawi

56 %
44 %
Information about Remote sensing-derived national land cover land use maps: a comparison...

Published on December 7, 2017

Author: rsmahabir


1. This article was downloaded by: [George Mason University] On: 04 September 2014, At: 13:48 Publisher: Taylor & Francis Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK Geocarto International Publication details, including instructions for authors and subscription information: Remote sensing-derived national land cover land use maps: a comparison for Malawi Barry Haack a , Ron Mahabir a & John Kerkering b a Department of Geography and Geoinformation Science, George Mason University, Fairfax, VA, USA b United States Forest Service, Washington, DC, USA Accepted author version posted online: 08 Aug 2014.Published online: 04 Sep 2014. To cite this article: Barry Haack, Ron Mahabir & John Kerkering (2014): Remote sensing-derived national land cover land use maps: a comparison for Malawi, Geocarto International, DOI: 10.1080/10106049.2014.952355 To link to this article: PLEASE SCROLL DOWN FOR ARTICLE Taylor & Francis makes every effort to ensure the accuracy of all the information (the “Content”) contained in the publications on our platform. However, Taylor & Francis, our agents, and our licensors make no representations or warranties whatsoever as to the accuracy, completeness, or suitability for any purpose of the Content. Any opinions and views expressed in this publication are the opinions and views of the authors, and are not the views of or endorsed by Taylor & Francis. The accuracy of the Content should not be relied upon and should be independently verified with primary sources of information. Taylor and Francis shall not be liable for any losses, actions, claims, proceedings, demands, costs, expenses, damages, and other liabilities whatsoever or howsoever caused arising directly or indirectly in connection with, in relation to or arising out of the use of the Content. This article may be used for research, teaching, and private study purposes. Any substantial or systematic reproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or distribution in any form to anyone is expressly forbidden. Terms & Conditions of access and use can be found at and-conditions

2. Remote sensing-derived national land cover land use maps: a comparison for Malawi Barry Haacka , Ron Mahabira * and John Kerkeringb a Department of Geography and Geoinformation Science, George Mason University, Fairfax, VA, USA; b United States Forest Service, Washington, DC, USA (Received 17 April 2014; final version received 4 August 2014) Reliable land cover land use (LCLU) information, and change over time, is impor- tant for Green House Gas (GHG) reporting for climate change documentation. Four different organizations have independently created LCLU maps from 2010 satellite imagery for Malawi for GHG reporting. This analysis compares the procedures and results for those four activities. Four different classification methods were employed; traditional visual interpretation, segmentation and visual labelling, digital clustering with visual identification and supervised signature extraction with application of a decision rule followed by analyst editing. One effort did not report classification accuracy and the other three had very similar and excellent overall thematic accura- cies ranging from 85 to 89%. However, despite these high thematic accuracies there were very significant differences in results. National percentages for forest ranged from 18.2 to 28.7% and cropland from 40.5 to 53.7%. These significant differences are concerns for both remote-sensing scientists and decision-makers in Malawi. Keywords: land use; land cover; greenhouse gases; Malawi; IPCC 1. Introduction Inadequate information on land cover and land use (LCLU) – both in regards to current LCLU and changes in LCLU over time – is a common problem in economic and envi- ronmental planning. In the absence of accurate LCLU information, policy-makers often fail to make informed decisions. Sound decisions depend on accurate information, yet every country – especially low-income countries – faces severe, competing demands for the financial and human commitments necessary to maintain a LCLU information system equal to its policy-making requirements (Cummings 1977). The frequent inadequacy of LCLU information may be due to a number of reasons: difficulties in accessing some regions because of limited or failed infrastructure or civil and military disturbances; lack of trained personnel, equipment or funds to collect information properly; or rapid changes in the resource base not detectable by traditional data collection methods, such as the high rates of deforestation in many areas of the world caused by increased population pressures (Haack & English 1996). One data source useful in providing current, reliable LCLU and other resource information is remote sensing. Historically, remote sensing was primarily conducting via aerial photography that was combined with ground surveys to obtain LCLU information (De Bruijn 1987; Lutchman 1987; Fonji & Taff 2014). More recently, *Corresponding author. Email: © 2014 Taylor & Francis Geocarto International, 2014 Downloadedby[GeorgeMasonUniversity]at13:4804September2014

3. information obtained by aerial photography has been augmented by satellite-borne sensors. Spaceborne remote sensing has effectively collected data since 1960 with the successful operation of the first of many meteorological satellites. Non-meteorological spaceborne systems have systematically acquired earth surface information since the launch of the first Landsat satellite in 1972, providing over 40 years of archived data. The time series of compatible data provided by these systems are extremely useful in understanding (LCLU) trends and modelling future conditions under different parame- ters. Moreover, the collection of high-resolution photography from Corona and other similar declassified spaceborne systems, during the 1960s, provide a high-quality historical record to further complement such analyses. Currently, LCLU change (LCLUC) is increasingly recognized as both a cause and consequence of climate change. LCLUC affects the biophysics, biogeochemistry and biogeography of the Earth’s surface and the atmosphere, with far-reaching conse- quences to human well-being (Giri et al. 2013). Accurate and timely information on LCLU and LCLUC is needed to understand the impact of climate variations on the structure and functioning of earth’s ecosystems and provision of ecosystem goods and services (Singh 1989; Rindfuss et al. 2004). The link between changes in LCLUC and climate variability has been confirmed by studies using modelled and observed data (Chase et al. 2000; Kalnay & Cai 2003; Ezber et al. 2007; Nunez et al. 2008). Simi- larly, observed and predicted climate change information is needed to understand the distribution and dynamics – both realized and potential – of LCLUC (Giri 2013). LCLUC could act as an indicator of climatic shift and can be monitored from space. For example, land cover is one of the 50 Essential Climate Variables (ECVs) identified by the Committee on Earth Observation Satellites that is technically and eco- nomically feasible for systematic observation. The Global Terrestrial Observing System further identified land cover as one of the five highest priority ECVs along with biomass, glacier and ice caps, soil moisture and permafrost. Similarly, generating land cover is now recognized as an official task of the Group on Earth Observations (Giri 2013). Information on Agriculture, Forest and Other Land Use is essential for the United Nations Framework Convention for Climate Change (UNFCCC) to devise policies and plans for cost-effective ways of combating climate change, either by increasing the removal of greenhouse gases through carbon sinks from the atmosphere or by reducing anthropogenic emissions (Trines & Hohne 2006). Also, the United States Global Change Research Programme has identified LCLUC as one of the key elements of interconnected issues of climate and global change (Rosenqvist et al. 2003). The United Nation’s Reduced Emissions From Deforestation and Forest Degradation programme highlights the need of mapping and monitoring of forest carbon stocks because defores- tation and forest degradation account for up to 20% of anthropogenic carbon emissions and are now included in climate change negotiations (Visseren-Hamakers et al. 2012). When combined with in situ measurements, spaceborne remote sensing provides distinct advantages over only field measures in providing timely and meaningful information for these initiatives (Rindfuss et al. 2004). Mapping LCLU and LCLUC is an important aspect of monitoring Green House Gases (GHG). Many countries have embarked on mapping efforts and a range of other data collection initiatives to report GHG data on a systematic schedule to the UNFCCC. Bilateral and multilateral development institutions often support these report- ing activities in developing countries (Bodansky 2010). Unfortunately, these develop- ment-funded activities are often uncoordinated, sometimes resulting in redundant and 2 B. Haack et al. Downloadedby[GeorgeMasonUniversity]at13:4804September2014

4. potentially contradictory data (Kalirajan et al. 2011). Moreover, many classification systems exist for extracting LCLU information from remotely sensed data. No interna- tionally accepted method exists around which such efforts might be aligned (FAO 2005). In Malawi, four independent, national-scale LCLU-mapping activities were recently completed using four different approaches. Multiple products might appear beneficial, yet, upon closer review, these parallel efforts not only represent inefficient use of limited donor funds, but they also create difficulties for the Government of Malawi; absent of one official LCLU data-set or a harmonized version of these four LCLU maps, the Government of Malawi lacks clarity on LCLU and, thus, may have difficulties making informed policy decisions related to LCLUC. Such decisions are especially important for the Government of Malawi since the country lacks complete and consistent national-scale LCLU base data (Palamuleni et al. 2010). The purpose of this study is to compare the various parameters and end-products from the four, recently completed land cover-mapping efforts in Malawi. Such a com- parison may be useful for the Government of Malawi, the remote-sensing community and, most importantly, for other countries contemplating similar national mapping efforts. This comparison is initiated with a brief review of the geography of Malawi in section 2. In section 3 a brief description of each LCLU product is given while section 4 compares each product in greater detail. Section 5 compares national LCLU statistics for each product. Finally, section 6 concludes this paper. 2. Geography of Malawi Malawi is a landlocked country located in southeast Africa. Prior to gaining indepen- dence from the United Kingdom in 1964, Malawi was known as Nyasaland. Malawi is bordered by Zambia to the northwest, Tanzania to the northeast and Mozambique on the east, south and west (Figure 1). The country is separated from Tanzania and Mozambique by Lake Malawi. Malawi is over 118,000 sq km (45,560 sq mi), roughly the size of the US State of Pennsylvania in size with an estimated population of over 16 million (Angelwicz 2012; Stock 2012). The Great Rift Valley runs through Malawi’s interior from north to south. In the mountainous sections of Malawi surrounding the Rift Valley, plateaus rise generally from 900 to 1200 m above sea level, although some rise close to 2500 m in the north (Cole & de Blij 2009). Malawi’s climate is hot in the low-lying areas in the south of the country and temperate in the northern highlands. The altitude moderates what would be an other- wise equatorial climate. Malawi’s’ peak rainfall is in January in the southern part, while February to March in the northern part depending on the movement of the ITCZ (Nicholson 2000; Cole & de Blij 2009). These weather patterns influence agricultural growing patterns and thus the appearance of agriculture on different image dates. Malawi is one of the world’s least-developed countries. The country has a low life expectancy and a high infant mortality (Hall & Midgley 2004). In 2005, it was esti- mated that around 52% of the population were living below the poverty line, 74.6% of which came from rural poverty (World Bank 2007). The economy is primarily based on agriculture and around 85% of the population lives in rural areas. More than one- third of GDP and 90% of export revenues come from agriculture. The main agricultural products of Malawi include tobacco, sugarcane, cotton, tea, corn, potatoes, sorghum, cattle and goats. The main industries are tobacco, tea and sugar processing, saw- mill products, cement and consumer goods (Place & Otuska 2001; Aryeetey-Attoh et al. 2009). Geocarto International 3 Downloadedby[GeorgeMasonUniversity]at13:4804September2014

5. The Malawian Government depends heavily on outside aid to meet development needs. As much as one-third of the country’s gross national income comes from exter- nal donor contributions with almost 40% of these funds provided by the four interna- tional donors; the United Kingdom, the European Commission, Norway and the World Bank (UNDP 2011). The Government of Malawi faces challenges in building and expanding the economy; improving education, health care, environmental protection; and becoming financially independent. The limited economic conditions in Malawi and the ongoing support from various bilateral and multilateral parties partially explain the four parallel national land cover-mapping efforts as described in the following sections. 3. National LCLU mapping in Malawi Four LCLU, remote sensing-based mapping efforts were recently completed for Malawi. Those efforts were funded by the United Nations Food and Agricultural Figure 1. Location of Malawi. 4 B. Haack et al. Downloadedby[GeorgeMasonUniversity]at13:4804September2014

6. Organization (UNFAO), World Bank, the Japan International Cooperation Agency (JICA) and the United States Agency for International Development (USAID). All the four projects produced maps for multiple years to assess LCLUC. Yet, a common thread through these efforts is that each completed maps based upon 2010 imagery. Those four activities are briefly summarized in the following sections and then compared by individual parameters and output products. 3.1. United Nations Food and Agricultural Organization UNFAO, in collaboration with the FAO Malawi office, supported national mapping institutions to assess land resources and their change over time using tools and method- ologies developed by FAO. This effort employed Landsat Thematic Mapper (TM) and Enhanced Thematic Mapper (ETM) imagery for the years 1990, 2000 and 2010 (FAO 2012). 3.2. World bank The World Bank funded a consortium of organizations to conduct national mapping for Malawi. That consortium was led by LTS International of Scotland and included HYD- ROC Consult, Bunda College of Agriculture in Malawi, The University of Edinburgh and the Centre for Development Management (LTS International et al. 2012). This effort compared national LCLU for the years 1973, 1992 and 2010. This project used an existing LCLU map for 1992, processed Landsat Multispectral Scanner (MSS) data to produce a map for 1973 and imagery from the French Systeme Pour le Observation de la Terre (SPOT) sensor for 2010. These LCLU maps were then used to simulate future LCLU conditions in Malawi for 2030 and 2050. 3.3. Japan International Cooperation Agency This JICA-funded effort was conducted by the Asia Air Survey Company Limited in association with the Department of Forestry within the Ministry of Environment and Climate Change Management of the Government of Malawi. National LCLU maps were created for the years 1990, 2000 and 2010. The 1990 map was a reclassification of an existing map of 26 classes produced by the Department of Forestry with external assistance. The 2000 map was an update of the 1990 map obtained by comparison of Landsat imagery for 1990 and 2000. The 2010 map was produced by processing space- borne imagery from the SPOT sensor and supplemented with other imagery sources (Asia Air Survey Company Ltd 2012). 3.4. United States Agency for International Development USAID funds NASA and other organizations under the SERVIR programme that, among other activities, has supported regional nodes in Nairobi and Kathmandu to encourage the use of remote sensing and associated spatial tools for multiple environmental and developmental activities. The Nairobi node is the Regional Centre of Mapping of Resources for Development (RCMRD) and was funded to map LCLU at three temporal nodes for six countries in Eastern and Southern Africa, including Malawi. The mapping was for the years 1990, 2000 and 2010 and employed Landsat imagery (RCMRD 2012). Geocarto International 5 Downloadedby[GeorgeMasonUniversity]at13:4804September2014

7. 4. Mapping comparisons The following sections document and compare various mapping parameters and results for the four independent efforts. These sections focus on the year 2010 as that is the one constant year across all four activities. The primary factors for comparison are source imagery, classes mapped, Minimum Mapping Unit (MMU), processing proce- dures, thematic accuracies and countrywide LCLU statistics. 4.1. Source imagery Data collected by the many different global satellite sensors currently in operation vary in resolutions and other characteristics, including costs. Currently, there are estimated to be over 60 different operational spaceborne sensors in the civil sector with fine-to- moderate spatial resolutions. There have been two long-lasting operational satellite sys- tems that are in many ways complementary – the United States Landsat satellites and the French SPOT. Since LCLU-mapping efforts – especially those conducted for climate change-related analyses – require data over time, these two systems have advantages over other systems, in that they each have an extensive archive of internally compatible imagery. Since 1972, eight Landsat satellites have been launched, seven successfully, with different primary sensors. The older sensor is the nominally 80-m spatial resolution MSS that collects data in four spectral regions. The more-advanced 30 m spatial resolu- tion TM and the slightly changed ETM collect data in seven primary spectral regions from the visible to the thermal infrared. The newest addition to the Landsat family, Landsat 8 launched in 2013, is an improvement over ETM with its Optical Land Ima- ger sensor. A standard Landsat frame is about 185 km on a side and Landsat platforms have a temporal resolution of 16–18 days (Campbell & Wynne 2011). Landsat has sev- eral advantages over other systems including an older archive, a larger area of coverage per frame and, for the last five years, all Landsat data have been made available at no cost. The six SPOT satellites in operation since 1986 can provide data with varied spatial resolutions in both panchromatic and multispectral modes. SPOT 6 has a 1.5 m pan- chromatic and a 6-m four-band multispectral sensor. A SPOT scene covers about 60 by 60 km and has great flexibility and frequency of temporal acquisition because its sensors may be employed in off-nadir mode (Campbell & Wynne 2011). After an evaluation of imagery options, the FAO effort selected Landsat TM and ETM for all three image epochs, 1990, 2000 and 2010 based upon optimum temporal and spatial coverage, data quality and that the data are free. Portions of 12 Landsat scenes are required to cover Malawi (Figure 2). The World Bank effort by LTS International employed SPOT-5 imagery for the 2010 mapping of Malawi was based upon high geometric resolution, multispectral capabilities, radiometric sensitivity and good positional accuracy among others. These SPOT data have three multispectral bands at 10 m and a panchromatic band at 5 m. The data were pansharpened to 5 m for mapping. Figure 3 is a SPOT mosaic for Malawi created by this effort with major catchments delineated. JICA acquired data from the Advanced Visible and Near-Infrared Radiometer-2 (AVNIR-2) sensor onboard the Japanese Advanced Land Observing Satellite (ALOS). This was supplemented by SPOT and Advanced Spaceborne Thermal Emission and Reflection Radiometer data for 2010. The 2010 ALOS AVNIR-2 imagery has 10-m 6 B. Haack et al. Downloadedby[GeorgeMasonUniversity]at13:4804September2014

8. spatial resolution in four wavelengths and a swath of 70 km. The 1990 map was a reclassification of an existing map and the 2000 map an update of the 1990 map derived from change detection using Landsat TM and ETM imagery for 1990 and 2000, respectively. The SERVIR RCMRD products for all three years were based upon Landsat imagery. The extensive Landsat archive and free data policy are understandable factors in data selection for this activity. 4.2. Classes Determining a remote-sensing classification system and class definitions is one of the least appreciated, most difficult and most important considerations in the appropriate use of remote sensing. Frequently, there is a difference between what classifications or definitions the user or decision-maker may want and what the data can accurately pro- vide. In many cases, it is the difference between land use required by policy-makers and land cover extractable from the imagery. Figure 2. FAO Landsat scenes to cover Malawi (FAO 2012). Geocarto International 7 Downloadedby[GeorgeMasonUniversity]at13:4804September2014

9. In mapping surface features, and especially mapping with remotely sensed data, there is a distinction between LCLU. Land cover is what is physically on the surface such as vegetation, water or buildings. Land use is the function of that cover for human activities (DeFries & Townshend 1999). For example, land cover may be forest but the land use may be recreational – a park, for instance. Remote sensing generally provides land cover information from which the analyst may infer land use. This inference can be done either by augmenting land cover obser- vations with additional data (e.g. socio-economic) or by processing (e.g. contextual Figure 3. World Bank pansharpened SPOT mosaic with major catchment areas delineated (LTS International et al. 2012). 8 B. Haack et al. Downloadedby[GeorgeMasonUniversity]at13:4804September2014

10. information), and not based on spectral classification alone (Alpin 2003). Typically, decision-makers prefer land use data but it is easier to extract land cover information with remote sensing. Most surface mapping using remote sensing extracts a mixture of land use and land cover classes. There are numerous guidelines for selecting a classification for remotely sensed data. For example, FAO (2005) suggests that a classification system should follow a hierarchical framework, accommodating different levels of information that move from broad-level classes and subdivide into successive and more detailed levels. Such a framework was developed by Anderson et al. (1976) and is still widely used today; a four-level hierarchical system for use with different spatial resolutions of remotely sensed data from satellites to low-altitude aircraft, with and without supplementary field activities. Consideration must also be given to variation in rules used to construct class definitions for different classifications systems. Such differences can lead to instances of the same feature with different properties across systems. A class definition for for- est, for example, can differ based on parameters including height, the type of vegeta- tion, per cent coverage, minimum size of forest units and ownership (e.g. state or private entity). Moreover, classification systems must be consistent over time and space to ensure the accurate and methodologically sound comparison of LCLU statistics and LCLUC. Such consistency across LCLU classification systems should also be done for local, regional and national programmes (Di Gregorio & Jansen 2000). Failure to observe such consistency may lead to results that are misleading and incomparable. The use of data gleaned from comparisons of data based on inconsistent classification procedures can result in ill-informed and erroneous high-level decisions. LCLU categories recommended by the UNFCCC Good Practice Guide (Penman et al. 2003) and Guidelines for Agriculture, Land Use and Forestry (Eggleston et al. 2006) were the primary focus of all four mapping efforts in Malawi. These guidelines suggest a two- tier or two-level classification system. The second, more detailed, level is typically 12–15 country-specific classes that can be aggregated into the six consistent classes across the globe for IPCC reporting. Those six classes with definitions are listed in Table 1. For three of the four 2010 products in Malawi, the six broad land cover classes of the IPCC were employed. The FAO effort used a slightly different classification system. The FAO maps were initially created using 24 classes based upon a classification sys- tem created by FAO. Those 24 classes were then combined to eight classes similar, but not equal, to IPCC guidelines. For the purpose of comparison to the other studies, tree plantation, trees and shrubs were combined to a general ‘forest class’; bare soil – a small feature – was renamed ‘other lands’; and water was assumed to be equivalent to the IPCC wetlands class that is a combination of open water and wetlands. Given the small percentages of the plantation (0.8%), shrub (1.1%) and bare soil classes (0.2%), these conversions are reasonable. The aggregation of bare soil is further supported by the IPCC definition of ‘other land’ in Table 1. 4.3. Minimum mapping unit A key parameter in producing any map is the smallest feature to be independently identified. This is called the MMU and is a statement of cartographic generalization. Typically, a MMU is stated in surface area units such as hectares or square kilometres. A MMU is more common in visual interpretation as in digital processing, it is often the pixel size. However, in digital processing, classified pixels are often filtered or smoothed to a larger, more suitable MMU. Geocarto International 9 Downloadedby[GeorgeMasonUniversity]at13:4804September2014

11. The MMU of the smallest polygon labelled in the FAO analysis was 2 ha. The World Bank LTS documentation does not state any MMU so it is assumed to be the SPOT pansharpened 5 m pixel. The JICA reclassification of the existing 1990 national map employed a MMU of 100 ha. The 2010 image-derived map was aggregated to a 25-ha MMU. The RCMRD final classification of Landsat 30-m data was filtered with a 3-by-3 window in a smoothing process which changes the effective MMU to 90 m. Comparison of maps at different MMUs is difficult. Comparison of statistics based upon different MMUs as in this examination is less of a concern. 4.4. Processing procedures There are two basic methods for the extraction of LCLU information from satellite- based remotely sensed data. Those are visual interpretations by analysts of the images from a computer screen and computer-based processing using the actual digital num- bers collected by the satellite sensor and appropriate image analysis software (Franklin & Wulder 2002; Berberoglu & Akin 2009). These two approaches have respective advantages and disadvantages and require different capabilities and infrastructures. Neither method is necessarily more accurate, more scientific or more repeatable. There are subjectivities in both approaches (Richards 2005). Frequently, a combination of these two methods is employed with an initial product created by computer-based analysis and then refined during a quality-control review by an analyst. In both traditional and computer-based LCLU extraction, the data are often radio- metrically and geometrically preprocessed. For visual analysis, qualified interpreters map the established LCLU classes, often with a stated MMU. Classification accuracy and consistency between interpreters is extremely important. This approach, as does digital processing, normally requires field verifications and extensive quality-control effort. In computer-based LCLU extraction, the remote-sensing digital values are directly manipulated by a computer to identify surface classes. This process may be called automatic classification, digital processing or numerical analysis. It typically Table 1. IPCC LCLU definitions. IPCC land cover land use class Definition Forest land Land spanning more than 0.5 hectares with trees higher than 5 metres and a canopy cover of more than 10%, or trees able to reach these thresholds in situ Cropland This category is applicable to cultivatable land and cultivated land, and also to agroforestry areas which are not defined as forest land by the national definition Grassland This category includes pastures and grazing lands. This category includes not only natural lands but also recreation areas Wetlands This category includes areas which are covered by water or wet areas (such as Peat bog) which is not classified as forest land, cropland or grassland Settlements This category covers all developed lands (including transport infrastructure) Other lands This category includes bare lands, rocks and unmanaged areas which are not covered by any of the above categories Source: Asia Air Survey Company Ltd 2012. 10 B. Haack et al. Downloadedby[GeorgeMasonUniversity]at13:4804September2014

12. requires a procedure of spectral signature extraction and then the application of a statistical decision rule for each pixel (Lawrence & Wright 2001; Jensen 2004). Successful digital classification depends on distinctive signatures for the classes of interest from image pixels, and the ability to reliably distinguish these signatures from other spectral response patterns in the imagery (Pal & Mather 2003; Lu & Weng 2007). There are two primary approaches to spectral signature extraction from remotely sensed data: supervised and unsupervised. The primary distinction between these two approaches is the amount of information the operator has about the images. In a super- vised classification, the user generally has some calibration or training data, meaning that the user generally knows the location of some features on the image. In unsuper- vised classification, an algorithm is chosen that will take a remotely sensed data-set and find a pre-specified number of statistical clusters in multispectral space (Al-Tahir et al. 2009). Although these clusters are not always equivalent to actual classes of land cover, this method is used without having prior knowledge of the ground cover in the study site (Tso & Mather 2001). An interesting mixture of classification methods were selected for Malawi and are presented following. An image-segmentation approach for LCLU classification was selected by FAO. This method, also called object-oriented as opposed to pixel-based, divides the image into spatially continuous and spectrally homogeneous regions or objects (Im et al. 2008; Blaschke 2010). The segmentation produces a vector layer of objects that represent regions with similar pixel values with respect to various characteristics or computed properties such as colour, intensity or texture and assumed to be the same basic LCLU (Budreski et al. 2007). This method allowed the generation of a very detailed layer of polygons in a short time, eliminating interpreter subjectivity in bound- ary delineation and increased the cost/benefit efficiency. This process created and classi- fied over 230,000 image objects for Malawi which were visually interpreted by Malawian national experts trained by FAO for this purpose. Figure 4 is an example of the segmentation polygons. The interpretation incorporated ancillary data in the polygon labelling and was followed by field visits to areas of uncertainty. The World Bank-LTS 2010 SPOT-based analysis included geometric correction, pansharpening to 5 m, image enhancement and mosaicing as preprocessing components. Field work collected 100 points for ground truth or calibration prior to final processing. The actual classification was done using visual interpretation of both the 2010 SPOT and the 1973 Landsat mosaics. Figure 5 is a flow chart of the imagery analysis procedures by LTS. The overall procedures for the JICA mapping are presented in Figure 6. As stated earlier, the 1990 map was a reprocessing of an existing map from 26 to 6 classes and the 2000 map was a visual update of the 1990 map by comparison of Landsat imagery for the two dates. As part of that process, Normalized Difference Vegetation Index values were compared for each epoch of Landsat images. The 2010 map was initiated by standard preprocessing methods including rectifica- tion, atmospheric and topographic correction, cloud cover and shadow removal and then mosaicing. The actual classification was done using unsupervised signature extrac- tion and visual interpretation of the output clusters. To assist with the cluster identifica- tion, 424 ground truth points were collected by field work. Initially, 30 clusters were created which were aggregated to 13 and then visually identified as one of the six IPCC GHG reporting classes using interpretation keys and topographic maps as ancillary data. Geocarto International 11 Downloadedby[GeorgeMasonUniversity]at13:4804September2014

13. The RCMRD-processing method was a standard supervised signature extraction and application of a maximum-likelihood decision rule. Figure 7 illustrates the processing flow for RCMRD. As in most LCLU mapping, the Landsat thermal band and, when available, the panchromatic bands were removed prior to analysis. This is in part due to their different geometric resolutions compared to other spectral bands. Multiple types of ancillary sources including Google Earth and elevation data were used in the selec- tion of areas of interest for signature extraction. The classified images were reviewed by analysts in a quality-control process and, when appropriate, visually edited. Figure 8 shows the classified maps of Malawi for the three years of interest as produced by RCMRD. Note that these maps have extended a buffer around the national boundary. Given the scientific research emphasis on digital processing for LCLU extraction, it is interesting that all four of these mapping efforts in Malawi used, in essence, a varia- tion of visual interpretation or at least relied upon visual components for final products. Those methods varied from traditional analyst delineation and labelling of class poly- gons, interpretation of software-determined segments or objects, analyst-assigned clas- ses to software-determined clusters to visual editing of more traditional-supervised signature extraction and application of a decision rule. Figure 4. Example of segmentation in FAO processing. Landsat subscene before and after segmentation for classification (FAO 2012). 12 B. Haack et al. Downloadedby[GeorgeMasonUniversity]at13:4804September2014

14. 4.5. Thematic accuracies Users of spatial data must be very aware of data accuracies. There are several types of map accuracies. Locational accuracy is typically easy to identify and correct by rectifi- cation. More difficult to determine and more important is thematic or class accuracy. The spatial sciences, and especially remote sensing, have a large literature on the importance and procedures for assessing thematic accuracies (Foody 2004; Congalton & Green 2009; Foody 2009). There is general agreement that using an error matrix with class-specific user and producer accuracies as well as overall accuracies is most appropriate for remotely Figure 5. World bank image-processing flow diagram (LTS International et al. 2012). Geocarto International 13 Downloadedby[GeorgeMasonUniversity]at13:4804September2014

15. sensed map products. In addition, many studies include a Kappa statistic as a comple- mentary measure of accuracy. The primary issues in accuracy assessment are not how to report accuracies but how large a sample to have, how to select the sample and what will be the validation or truth data to compare to the classification (Janssen & van der Wel 1994; Stehman & Czapelwski 1998). Some of these issues have been identified by Foody (2002), reviewing land cover accuracy assessment areas that limit the ability to appropriately assess, document and use the accuracy of thematic maps derived from remote sensing. Validation efforts should generally include field visits but increasingly use other remote-sensing sources of finer spatial resolution data including Google Earth (Cha & Park 2007; Potere 2008). The FAO mapping included visits to 140 points in the field supplemented by high spatial resolution imagery. They report an overall accuracy of 89.2%, presumably, for the 2010 map but it is not clearly stated. They also report a class change accuracy of 97.7% but not for which change years. The JICA 2010 LCLU map thematic accuracy was ascertained by a stratified random-sampling design resulting in 937 samples. These samples produced an overall Figure 6. JICA work flow (Asia Air Survey Company Ltd 2012). 14 B. Haack et al. Downloadedby[GeorgeMasonUniversity]at13:4804September2014

16. Figure7.RCMRDimage-processingdiagram(RCMRD2012). Geocarto International 15 Downloadedby[GeorgeMasonUniversity]at13:4804September2014

17. Figure8.ClassifiedmapsofMalawiforthreeyearsproducedbyRCMRD(RCMRD2012). 16 B. Haack et al. Downloadedby[GeorgeMasonUniversity]at13:4804September2014

18. accuracy of 87.6% but with a subdivision of the forest land into four subclasses which likely reduced overall accuracy. Producer’s accuracies ranged from 33.3% for other lands, but with only six sample points, to 100% for Settlements but with only nine points. The user’s accuracies have a low of 65.8% for Miombo Woodlands, a subset of forest, to 100% for Evergreen Forest, also a subset of forest. One of the issues with the JICA prod- uct is that it assumed the areas within demarcated forest reserves were actually forest. RCMRD conducted a rigorous field programme for thematic accuracy assessment augmented by other sources such as Google Earth. For Malawi, the 2010 overall accu- racy for the 14 classes in Scheme II was 79.5% which increased to 85.2% for the six classes of Scheme I (Table 2). The level-1 analysis was based upon 542 sample points. Kappa values were 0.74 and 0.79, respectively. User’s accuracies varied from a low of 50.0% for other lands to 100% for wetlands and producer’s from 12.5% for other lands to 96.5% for settlements. The low results for other lands are of limited concern, as there were few samples. Of the 542 pixels in the validation sample set, only 10 were in the other lands class. The World Bank LTS report did not include any thematic accuracy information. Of the three efforts for which overall thematic accuracies were reported, they varied from 85.2 to 89.2%. These are excellent, as a general goal in the industry is 85% (Anderson et al. 1976), and very consistent. The differences in thematic accuracies are likely due to variations as discussed previously, that is, source imagery, classes, MMU and processing procedures. Further issues are also expected with differences in acquisi- tion dates of images. The spectral composition of an image can vary due to seasonal changes in LCLU and changes due to extreme events such as hurricanes and climatic conditions. Such differences can in turn impact user or computer interpretation of what is on the ground and changes in processing procedures used for extracting LCLU. The acquisition dates of images used for the mapping products in this study were however not available to these authors. Another possible area of contention in reported accuracy values is the number of samples used for validating each land cover class. For example, JICA had only nine points for validating settlements. Due to variation in the materials used for rooftops in Malawi, including, tin, wood, mud, grass and corrugated iron, fail- ure to adequately take into account the wide range of roofing materials used could lead to bias accuracy estimates in LCLU. Furthermore, the accuracy assessment reported in the documents available to the authors of this study is not conclusive, such that it is difficult to judge the authenticity of the results obtained. A more thorough assessment could be made if data on the absolute locations of validation sites becomes available. Table 2. RCMRD-USAID thematic classification Malawi 2010. Class Forest Settlement Cropland Other Wetland Grassland Totals User’s accuracies Forest 110 0 8 0 5 4 127 86.6% Settlement 2 55 6 0 1 0 64 85.9% Cropland 19 1 225 6 0 16 267 84.3% Other 0 1 0 1 0 0 2 50.0% Wetland 0 0 0 0 40 0 40 100.0% Grassland 5 0 5 1 0 31 42 73.8% Totals 136 57 244 8 46 51 542 Producer’s accuracies 80.9% 96.5% 92.2% 12.5% 87.0% 60.8% Overall % 85.3 Geocarto International 17 Downloadedby[GeorgeMasonUniversity]at13:4804September2014

19. 5. Comparison of national statistics Table 3 compares the LCLU statistics for the four efforts while Figure 9 displays these results visually on a column bar chart. As stated previously, the FAO map did not use the six IPCC classes while the other three products did. FAO had eight classes and slightly different class names. Those eight classes were renamed or combined to the six required by IPCC and those changes are not considered significant as the affected clas- ses were very minor in extent. A much greater difficulty in initial comparisons was that three of the efforts used the standard international boundary while the fourth, RCMRD, mapped and included a 10-km buffer around the entire country. In order to make all 2010 map products comparable, the RCMRD provided 2010 map was reduced to the common international boundary employed by the other three projects. Given that the three reported thematic accuracies were very good and very consistent, the final class percentages and areas are surprisingly different. Some of the obvious discrepancies are forest which varied from 18.2 to 28.7% of the country (21,527–33,952 sq km, respectively), and cropland from 40.5 to 53.7% of the country (a difference of about 15,000 sq km). Even the classes that are relatively small had sig- nificant variations, such as grasslands ranging from 2.6 to 9.0% of Malawi. It is also Table 3. Malawi LCLU national mapping comparisons 2010. JICA World bank FAO RCMRD USAID Class Sq km % Sq km % Sq km % Sq km % Forest 24,177 20.4 21,527 18.2 34,865 28.7 33,687 28.6 Cropland 59,415 50.2 63,483 53.7 47,752 40.5 47,959 40.7 Grassland 3180 2.7 3018 2.6 10,601 9.0 9013 7.6 Wetland 30,902 26.1 29,268 24.7 23,691 20.1 26,039 22.1 Settlement 513 0.4 717 0.6 1714 1.4 731 0.6 Other 132 0.1 286 0.2 213 0.2 410 0.3 Totals 118,319 99.9 118,299 100.0 117,846 99.9 117,839 99.9 Figure 9. Malawi LCLU national mapping comparisons 2010. 18 B. Haack et al. Downloadedby[GeorgeMasonUniversity]at13:4804September2014

20. apparent from Figure 9 that there are two pairings of reported results; JICA with World Bank and FAO with RCMRD. These dramatic differences are initially surprising. They also create obvious problems for scientists and decision-makers. A problem for the gov- ernment is should they attempt to harmonize these differences, which would be quite challenging, or how to select one map over the others. Clearly comparing any of these maps to earlier mapping efforts creates uncertainties as to the validity of any results. A review of the landscape characteristics of Malawi and the mapped classes pro- vides some insights as to why there are differences in results across mapping efforts but does not resolve the differences. Malawi has a complex landscape of many small farms or agricultural areas. This is complicated by transitions of natural vegetation from closed canopy forest to shrubland, savanna and eventually, grasses. Assigning classes to these transitions consistently is very difficult. There is also a complex climatic pat- tern of wet and dry months that is not consistent across the country or from year to year (McSweeney et al. 2014). In such a situation, imagery of different months can be mapped quite differently. However, in the case of Malawi, information on dates of acquisition for imagery used was not present in reports made available for public use for assessment of this variable. Finally, the IPCC classes of forest, agriculture, grass- lands and even wetlands are all variations of vegetation and difficult to consistently separate. Figure 10 is an aerial photograph illustrating the complex landscapes of Malawi. 6. Conclusions It is unusual to have access to independent mapping efforts of large areas for compari- son as in this analysis. The four national maps for Malawi all employed spaceborne imagery for the year 2010. They all mapped, with minor but correctable variations, the six LCLU classes for GHG reporting as recommended by IPCC. It was interesting that there were four different processing strategies that all included a significant visual inter- pretation element. Those four methods included natural boundary or polygon delinea- tions and labelling, visual labelling of software segmentations, clustering and assignment of classes and traditional signature extraction and application of a decision rule but followed by analyst editing. Figure 10. Aerial photograph of Malawi illustrating the complex landscape (Krapt 2010). Geocarto International 19 Downloadedby[GeorgeMasonUniversity]at13:4804September2014

21. The three reported overall accuracies were high – around 85 and 89%, and quite consistent. However, the reported national statistics for 2010, either as percentages or square kilometres, are surprisingly quite different. Forest lands varied between 18.2 to 28.7% of the country (a difference of about 14,000 sq km), cropland from 40.5 to 53.7% (almost 16,000 sq km) and grassland from 2.6 to 9.0%. Given the complexity of the landscape of Malawi and the similarities in classes mapped, these differences are understandable to some degree. Malawi is a very rural society with many small agricul- tural plots, often with mixed cropping, and interspersed with other land covers creating many mixed pixels. There are also complex climatic and meteorological patterns caus- ing the landscape to vary considerably as a function of precipitation. This could result in images of different dates being very different in appearance and the ability to map specific features. The primary classes for IPCC reporting – forestland, cropland and grassland – are all variations in vegetation and could easily be confused in spectrally based mapping. Even the high and consistently reported overall thematic mapping accuracies are under- standable as each mapping effort likely employed internally consistent class definitions. The apparent problem is that each mapping effort had different class definitions. There is really no reason to believe that one of these four efforts is more accurate than the other. It is somewhat interesting that there are two pairs of fairly consistent results; JICA with World Bank and FAO with RCMRD. It is reasonable to assume that all four efforts are scientifically valid and internally consistent but employed different definitions. It would be interesting, though not necessarily valid, to have an indepen- dent evaluation of the accuracies of all four maps, but those results would be biased by the definitions of the evaluation team. Variation in results such as these four products for Malawi indicates concerns and difficulties for both the remote sensing and decision-making communities. These four efforts indicate good accuracies in thematic mapping, but very different statistical results. For the Government of Malawi, they may have the difficult task of either har- monizing these results or selecting one product. For the remote-sensing community, these results indicate how difficult accurate mapping of LCLU can be. Furthermore, with the application of more soft approaches to LCLU classification, such as fuzzy membership, as discussed by Foody (1999), it may very well be that there is need for a rethinking of approaches and application towards the discretization of LCLU. The unfortunate, but important, observation is that LCLU mapping via remote sens- ing is not a consistent process. There is a tremendous amount of judgement involved which creates differences in results. As has been noted elsewhere, remote sensing is both an art and a science. Since many decisions are based upon trends, the lesson learned is to use the same approach to map different dates of a landscape and act upon the trends rather than the actual products. References Alpin P. 2003. Comparison of simulated IKONOS and SPOT HRV imagery for classifying urban areas. In: Mesev V, editor. Remotely-sensed cities. New York, NY: Taylor and Francis; p. 25. Al-Tahir R, Richardson T, Mahabir R. 2009. Advancing the use of earth observation systems for the assessment of sustainable development. Assoc Prof Eng Trinidad Tobago. 38:6–15. Anderson JR, Hardy EE, Roach JT, Witmer RE. 1976. Land use and land cover classification system for use with remote sensor data. United States Geological Survey Professional Paper 964; Washington, DC: US Government Printing Office; 41 p. 20 B. Haack et al. Downloadedby[GeorgeMasonUniversity]at13:4804September2014

22. Angelwicz P. 2012. Migration, marital change, and HIV infection in Malawi. Demography. 49:239–265. Aryeetey-Attoh S, McDade BE, Obia GC, Oppong JR. 2009. Geography of Sub-Saharan Africa. 3rd ed. Upper Saddle River, NJ: Pearson Education; 468 p. Asia Air Survey Company Limited and Department of Forestry, Ministry of Environment and Climate Change Management, Republic of Malawi. 2012. Forest resource mapping project, final report for implementation phase, Lilongwe, Malawi; 483 p. Berberoglu S, Akin A. 2009. Assessing different remote sensing techniques to detect land use/ cover changes in the eastern mediterranean. Int J Appl Earth Obs Geoinf. 11:46–53. Blaschke T. 2010. Object-oriented image analysis for remote sensing. ISPRS J Photogrammetry Remote Sens. 65:2–16. Bodansky D. 2010. The Copenhagen climate change conference: a postmortem. Am J Int Law. 104:230–240. Budreski KA, Wynne RH, Browder JO, Campbell JB. 2007. Comparisons of segment and pixel- based nonparametric land cover classification in the Brazilian Amazon using multitemporal Landsat TM/ETM+ imagery. Photogrammetric Eng Remote Sens. 73:813–827. Campbell JB, Wynne RH. 2011. Introduction to remote sensing. New York, NY: Guilford Press; 667 p. Cha S, Park C. 2007. The utilization of google earth images as reference data for the multitempo- ral land cover classification with MODIS data of North Korea. Korean J Remote Sens. 23:483–491. Chase TN, Pielke RA Sr, Kittel TGF, Nemani RR, Running SW. 2000. Simulated impacts of his- torical land cover changes on global climate in northern winter. Clim Dyn. 16:93–105. Cole R, de Blij HJ. 2009. Survey of Subsaharan Africa: a regional geography. New York, NY: Oxford University Press; 768 p. Congalton R, Green K. 2009. Assessing the accuracy of remotely sensed data, principles and practices. 2nd ed. Boca Raton, FL: CRC Press; 183 p. Cummings RW Jr. 1977. Minimum information systems for agricultural development in low- income countries. New York, NY: Agricultural Development Council; 13 p. De Bruijn CA. 1987. Monitoring a large squatter area in Dar es Salaam with sequential aerial photography. ITC J. 87:233–238. DeFries RS, Townshend JRG. 1999. Global land cover characterization from satellite data: from research to operational implementation? Global Ecol Biogeogr. 8:367–379. Di Gregorio A, Jansen LJM. 2000. Land cover classification system LCCS: classification con- cepts and user manual. Rome: FAO Environment and Natural Resources Service, FAO Land and Water Development Division. Eggleston HS, Buendia L, Miwa K, Ngara T, Tanabe K. 2006. IPCC guidelines for national greenhouse gas inventories. Vol 4. Hayama: Agriculture, Forestry and Other Land Use Insti- tute for Global Environmental Strategies; 196 p. Ezber Y, Sen OL, Kindap T, Karaca M. 2007. Climatic effects of urbanization in Istanbul: a sta- tistical and modeling analysis. Int J Climatol. 27:667–679. FAO. 2005. Land cover classification system LCCS: classification concepts and user manual. Di Gregorio A, Jansen LJM, revised. FAO Environment and Natural Resources Service, FAO Land and Water Development Division. FAO. 2012. Atlas of Malawi, land cover and land cover change, 1990–2010. Rome: United Nations Food and Agriculture Organization; 137 p. Fonji SF, Taff GN. 2014. Using satellite data to monitor land-use land-cover change in North- eastern Latvia. Springer Plus. 3:61 p. Foody GM. 1999. The continuum of classification fuzziness in thematic mapping. Photogrammet- ric Eng Remote Sens. 65:443–451. Foody GM. 2002. Status of land cover classification accuracy assessment. Remote Sens Environ. 80:185–201. Foody GM. 2004. Thematic map comparison: evaluating the statistical significance of differences in classification accuracy. Photogrammetric Eng Remote Sens. 70:627–633. Foody GM. 2009. Sample size determination for image classification accuracy assessment and comparison. Int J Remote Sens. 30:5373–5291. Franklin SE, Wulder MA. 2002. Remote sensing methods in medium spatial resolution satellite data land cover classification of large areas. Prog Phys Geogr. 26:173–205. Geocarto International 21 Downloadedby[GeorgeMasonUniversity]at13:4804September2014

23. Giri C. 2013. Technology assessment and review of global, regional, and national mid-resolution land cover datasets in the lower mekong region (Cambodia, Laos, Thailand and Vietnam). Sioux Falls, SD: USGS, EROS Data Center; 19 p. Giri C, Pengra B, Long J, Loveland TR. 2013. Next generation of global land cover characteriza- tion, mapping, and monitoring. Int J Appl Earth Obs Geoinf. 25:30–37. Haack B, English R. 1996. National land cover mapping by remote sensing. World Dev. 24:845– 855. Hall AL, Midgley J. 2004. Social policy for development. London: Sage; 308 p. Im JJ, Jensen JR, Tullis JA. 2008. Object based change detection using correlation image analysis and image segmentation. Int J Remote Sens. 29:399–423. Janssen LF, van der Wel FJM. 1994. Accuracy assessment of satellite derived land cover data: a review. Photogrammetric Eng Remote Sens. 94:419–426. Jensen JR. 2004. Introductory digital image processing. 3rd ed. Upper Saddle River, NJ: Prentice Hall; 526 p. Kalirajan K, Singh K, Thangavelu S, Venkatachalam A, Perera K. 2011. Climate change and poverty reduction: where does official development assistance money go? No. 318. ADBI working paper series. Kalnay E, Cai M. 2003. Impact of urbanization and land-use change on climate. Nature. 423:528–531. Krapt H. 2010. Malawi, aerial photographs after take-off from Chileka airport in Blantyre Wiki- pedia. [cited 2014 Apr 14]. Available from: 23_13-47-49_Malawi_-_Chileka.jpg Lawrence RL, Wright A. 2001. Rule-based classification systems using classification and regres- sion tree analysis. Photogrammetric Eng Remote Sens. 67:1137–1142. LTS International, HYDROC Consult, Bunda College of Agriculture, the University of Edinburgh and the Centre for Development Management. 2012. Mapping land cover and future land cover projections scenario analysis technical annex 1- integrated assessment of land use options in Malawi Submitted to the world bank and government of Malawi by, LTS Interna- tional Ltd Pentlands Science Park, Bush Loan Penicuik, EH26 0PL United Kingdom; 36 p. Lu D, Weng Q. 2007. A survey of image classification methods and techniques for improving classification performance. Int J Remote Sens. 28:823–870. Lutchman HTJ. 1987. Monitoring land subdivisions of the fringe or Paramaribo using aerial pho- tography. ITC J. 87:248–253. McSweeney C, New M, Lizcano G. 2014. UNDP climate change country profiles Malawi. [cited 2014 Mar 3]. Accessed from: Nicholson SE. 2000. The nature of rainfall variability over Africa on time scales of decades to millennia. Global Planet Change. 26:137–158. Nunez MN, Ciapessoni HH, Rolla A, Kalnay E, Cai M. 2008. Impact of land use and precipita- tion changes on surface temperature trends in Argentina. J Geophys Res: Atmos (1984– 2012). 113:1–11. Pal M, Mather PM. 2003. As assessment of decision tree methods for land cover classification. Remote Sens Environ. 86:1137–1142. Palamuleni LG, Annegarn HJ, Landmann T. 2010. Land cover mapping in the Upper Shire River catchment in Malawi using Landsat satellite data. Geocarto Int. 25:503–523. Penman J, Gytarsky M, Hiraishi T, Krug T, Kruger D, Pipatti R, Buendia L, Miwa K, Ngara T, Tanabe K, Wagner F, editors. 2003. Good practice guidance for land use, land-use change and forestry. Hayama: Institute for Global Environmental Strategies; 593 p. Place F, Otsuka K. 2001. Population, tenure, and natural resource management: the case of cus- tomary land area in Malawi. J Environ Econ Manage. 41:13–32. Potere D. 2008. Horizontal positional accuracy of google earth’s high-resolution imagery archive. Sensors. 8:7973–7981. [RCMRD] Regional Centre for Mapping of Resources for Development. 2012. Project implemen- tation guide: Malawi, land cover mapping for green house gas inventories development pro- ject in East and Southern Africa region. Nairobi, Kenya; 85 p. Richards JA. 2005. Analysis of remotely sensed data: the formative decades and the future. IEEE Trans Geosci Remote Sens. 43:1007–1011. Rindfuss RR, Walsh SJ, Turner BL, Fox J, Mishara V. 2004. Developing a science of land change: challenges and methodological issues. Proc Natl Acad Sci. 101:13976–13981. 22 B. Haack et al. Downloadedby[GeorgeMasonUniversity]at13:4804September2014

24. Rosenqvist A, Milne A, Lucas R, Imhoff M, Dobson C. 2003. A review of remote sensing tech- nology in support of the Kyoto protocol. Environ Sci Policy. 6:441–455. Singh A. 1989. Digital change detection techniques using remotely sensed data. Int J Remote Sens. 10:989–1003. Stehman SV, Czaplewski RL. 1998. Design and analysis for thematic map accuracy assessment: fundamental principles. Remote Sens Environ. 64:331–344. Stock R. 2012. Africa South of the Sahara: a geographical interpretation. 3rd ed. New York, NY: Guilford Press; 593 p. Trines E, Hohne N. 2006. Climate change scientific assessment and policy analysis: integrating agriculture, forestry and other land use in future climate regimes. Bilthoven, Netherlands Environmental Assessment Agency, Report No. 500102002; 154 p. Tso B, Mather PM. 2001. Classification methods for remotely sensed data. New York, NY: Tay- lor and Francis; 332 p. UNDP. 2011. Assessment of development results Evaluation of UNDP contribution. Malawi: United Nations Development Programme USA; 105 p. Visseren-Hamakers IJ, Gupta A, Herold M, Peña-Claros M, Vijge MJ. 2012. Will REDD+ work? The need for interdisciplinary research to address key challenges. Curr Opin Environ Sustain- ability. 4:590–596. World Bank. 2007. Malawi – Poverty and vulnerability assessment: investing in our future Wash- ington, DC: World Bank. Report No. 36546; 301 p. Geocarto International 23 Downloadedby[GeorgeMasonUniversity]at13:4804September2014

Add a comment