Photo courtesy of Neil Palmer, CIAT

Maps describing Earth's land cover tell different stories. Which should we believe?

Even a reliable land cover map isn't a perfect representation of the landscapes it portrays. But without some spatially-explicit picture of the distribution and nature of farmland, grazing lands and agroforestry resources, it is difficult to properly gauge local and regional opportunities for agricultural development or to see where specific policies, technologies and land use practices are bringing about visible change. Where, for example, might further expansion of the cropland frontier be feasible? And would such expansion involve significant tradeoffs in terms of the loss of valuable habitat or of other ecosystem services? Such questions are themselves complex but researchers must first choose a land cover map on which to base their analysis. The choice is not as easy as you might think.

Satellite-based, remote sensing technology is the major contemporary basis for land cover assessment using a judicious blend of image analysis and “ground truthing”. The range of land cover data products generated, however, is confusingly large; image sensors are designed to monitor different land surface and vegetation characteristics and different levels of spatial and temporal resolution, all from distinct processing approaches with varying degrees of ground-based validation. Not surprisingly, the accuracy of different land cover classes also varies among and within the various products. In general, estimates of the areal extent of urban areas, water bodies, bare soil and some types of forests are more reliable than that of cropland, grassland and savanna. In areas where fields are small and heterogeneous, as is typical across much of Africa’s farmlands, the ability to reliably detect cropland is particularly problematic. Furthermore, some management activities (e.g., fallow, intercropping and irrigation) makes it even harder to distinguish agricultural from other land cover types.

HarvestChoice has highlighted some of the practical issues faced in evaluating and choosing amongst the most widely-used land datasets applied to land cover assessments in sub-Saharan Africa. To facilitate a comparison across datasets, a harmonized system of land cover classes was first developed via HavestChoice from 13-27 land cover classes associated with five different products (Table 1) . The new classification system comprised just eight land cover classes (Table 2).

Table 1. Land cover classes of GLC2000,GlobCover,
AVHRR_UMD, GLCC and MODIS (click to enlarge)

 

 

 

Table 2. Eight aggregated classes for five land cover products

The reclassified maps of Africa for the five different regional land cover products are shown in Figure 1. The spatial distribution of some of the classes appears similar across maps, while others clearly do not. For example, the first three maps indicate larger extents of cropland and forests and less grassland throughout Africa compared to the last two maps. However, for regions where classes of land cover are dense or clustered, such as the forested Congo basin and the breadbasket of the Ethiopian highlands, the five maps are consistent. Table 3 summarizes the total area of each harmonized land cover class across four Africa regions: Eastern Africa, Western Africa, Northern Africa and Southern Africa. The coverage area of each aggregated land cover class across regions is displayed in Chart 1.

Figure 1. Comparison of five land cover product maps using a common legend

Table 3. Common land cover type area in Africa regions

Chart 1. Comparison of area (ha) from common land cover types in Africa regions (from top to bottom: Eastern Africa, Northern Africa, Southern Africa and Western Africa) Chart 1.png

The reasons underlying discrepancies between maps are manifold, including the variety of classification schema/methods, instrument/sensors, and quality of validation inputs (e.g., ground truthing, local knowledge). Additionally, the reclassification options from the original classes to the aggregated common classes were not ideal. Nevertheless, the differences among data sources are striking for certain land cover classes, especially those related to agriculture (surprised?).

So which land cover product is the best? Which one should I choose? Before we express our views, it might be useful to zoom into a few countries for a closer look. Figure 2 shows country level maps for Kenya, Uganda, Rwanda, Tanzania and Ethiopia generated under a common legend using three of the same land cover products and the nationally-generated Africover products. Again, the difference between products remains striking and in some cases a quick visual assessment, guided by regional knowledge, reveals obvious inaccuracies. Take Uganda, for example. It is well known that the northeastern region of Uganda experiences low rainfall and is generally unsuited for annual cropping systems. Yet the map generated using GLC2000 displays a high density of crop production in that area, for a product that is deemed reliable and is widely used for agricultural and ecosystem studies in Africa (e.g., the Millennium Ecosystem Assessment and the International Assessment of Agriculture Science and Technology for Development).

Figure 2. Land cover maps for selected countries using a common legend

Besides comparing land cover maps against each other, one can judge the accuracy of cropland cover by comparing maps with national agricultural production statistics. Table 4 and Chart 2 show cropland area in eight states in Nigeria derived from both national statistics and GlobCover. The GlobCover maps appear to over-estimate cropped areas in drier ecological zones (Yobe, Kano, and Gombe states) and under-estimate in more humid zones, especially in humid forest zones (Delta and Cross River States). The same patterns also emerge in Ghana (not shown).

Table 4. Comparison of cropland area (ha) in Nigeria

 

 

 

 

 

Chart 2. Comparison of cropland area (ha) in Nigeria

Back to the original question: Which land cover product should I use?! Well, you won’t be surprised if we give you the standard HarvestChoice answer. It all depends. It depends, amongst other things, on which region, agro-ecological zone, and specific production system you’re interested in, and what questions you are trying to answer. Each product has unique advantages and limitations. So it’s up to you to decide. The good news, all the products are publicly available, free of charge, and in a common storage format. But clearly, improving the detection and monitoring of agriculture land is urgently needed since accuracy of this class is relatively lower than that of other land cover classes. As remote sensing technology advances, we can expect to see higher resolutions and increased temporal frequency of land cover data and more land use products in development. Indeed, in just the past few years quite a few high-quality land cover products at relatively large scales have begun tackling the job of accurately describing the surface of the planet.

Citation

HarvestChoice, 2013. "The Tale of Earth’s Land Cover." International Food Policy Research Institute, Washington, DC., and University of Minnesota, St. Paul, MN. Available online at http://harvestchoice.org/node/5245.

Jun 17, 2013