Landsat footprint
Landsat 8 nominal image footprint for the West Africa mapped area

Mapping the land use and land cover of all of West Africa for three periods in time (1975, 2000 and 2013) using many hundreds of Landsat images required careful consideration with regard to a methodology. Mapping land cover over time requires an approach that generates consistently accurate maps over time for reliable change detection. Of two basic mapping approaches — computer automated classification and visual image interpretation — one needed to be chosen.

The most common approach in land use and land cover mapping is automated classification — throwing huge amounts of digital image data into an image classifier. However, earlier experiences with automated classification in Mali, Senegal and Niger produced disappointing results. Important land use and land cover types, such as agriculture in the Sahel, could not be uniquely differentiated from other types based on their spectral reflectance properties. Automated methods of image classification are based on spectral image data and are often plagued by problems of misclassification. Spectral reflectance of land surfaces — and more broadly spectral response patterns — measured by remote sensors may be quantitative but they certainly are not absolute. They may be distinctive but they are not necessarily unique. In reality, there is often extreme variability of spectral reflectance associated with various land cover types (Lillesand and Kiefer, 1994). This variability poses major challenges in mapping and analyzing land cover types based solely on their spectral properties. For these reasons, the project chose visual image interpretation rather than the semi- or fully-automated approaches.

Mapping land use and land cover using visual interpretation is not without  its  own  challenges, but the combination of firsthand knowledge of the landscapes and reliance on multiple dimensions of information inherent in imagery is a powerful approach to producing highly accurate maps. Mapping land cover from satellite images requires special skills and detailed local knowledge about the area of interest — including its physical, biological and human components. Satellite images, much like aerial photographs, contain a detailed record of features on the Earth’s surface at the time of acquisition. Drawing upon training, field experience, geographic knowledge, an acute power of observation and patience, image analysts mapped the land use and land cover using visual interpretation. They relied on the basic elements of image interpretation: shape, size, pattern, color, tone, texture, shadows, geographic context, and association. The time of year when each image was acquired was also an important factor in identifying the land features. The image interpretation process was facilitated through the use of interpretation guidelines, which included written and illustrated definitions of all of the land use and land cover classes.

Steps in making a land use and land cover map using the RLCM Tool; 1) selecting imagery, 2) overlaying a grid of points, 3) interpreting  and  attributing,  and 4) making land use and land cover map.

The visual image interpretation method works well for several important reasons. First, it lends itself well to working with photographs and images from different satellite systems and formats. Second, it allows expert interpreters to integrate local knowledge with the many dimensions of information contained in images. Third, image interpreters can readily account for, and work around, problems related to seasonal differences from image to image, as well as differences in illumination and atmospheric effects. For example, the human interpreter can effectively distinguish real land changes from many of the ephemeral changes on the land such as burn patterns from annual grass fires. Image interpretations were systematically validated with high-resolution satellite imagery and, in many countries, with visits to the field. Fourth, in order to maintain high accuracy and reduce confusion among land cover types, we defined land use and land cover classes that could be consistently identified and mapped from Landsat satellite imagery. Fifth, the requirement of mapping land use and land cover at multiple periods of time necessitated high accuracy in order to confidently characterize the changes from period to period. When done properly, the visual interpretation method provides the high accuracy needed.

In order to check the accuracy of the maps, the analysts used multiple sources of ancillary data, including thousands of aerial photographs taken by the project team, and recent high-resolution satellite images viewable in Google Earth. The Google Earth tool was particularly useful in systematically checking the mapping of land cover from recent Landsat imagery.

The traditional method of visual interpretation is ideal for the reasons given above, but it is also labor- intensive, particularly for such a vast area; mapping millions of square kilometers for three points in time would have been insurmountable using the traditional method. To expedite the interpretation process while still maintaining temporal accuracy, the U.S. Geological Survey (USGS) EROS Center team developed the Rapid Land Cover Mapper (RLCM) tool. The RLCM tool is a Geographic Information System (GIS) vector/raster hybrid approach that lends itself  to both multiple resolutions and time series mapping. Conceptually, the RLCM is based on the traditional dot grid method for calculating area, employed by foresters for over a century (Schumaher and Chapman, 1972).The RLCM tool generates a digital grid of points that overlays an image.

A simplified time series representation of three land use and land cover interpretations using the RLCM Tool. The cubes symbolize the grid points that are placed over the imagery, color coded by land cover class. The vertical lines show the spatial alignment of the points through time.

Using standard photo interpretation techniques, the interpreter identifies the discrete land use and land cover class for each point. The RLCM tool facilitates the selection and assignment of the point’s land cover class. This is accomplished by simultaneous point selection and cascading period classification. Simultaneous selection allows the interpreter to select many points of a common class and assign them to that specific class with one action. Cascading is a method of completing the classification of a first time period for a given area, then pushing that classification information forward or back in time. After copying the attributed points into another time period, they are displayed over images that correspond to the “new” time period. The interpretation process is repeated, and in this case the previously attributed points are reviewed to determine if they should remain unchanged or be edited to reflect a change in the land cover.

Generally, the image analysts began with the most recent period, then worked back in time to the earlier periods. This resulted in the production of multi-period land use and land cover maps and associated statistics that characterize the changing landscapes at national and regional scales.