Counting Crowns in Nepal: Mapping Vegetation and Land Cover with eCognition Software

The small communities living in Nepal’s high mountain regions depend nearly exclusively on local natural resources for their livelihoods, yet they have to exploit those same resources in order to have shelter and carve out livelihoods. It’s a perpetual Catch-22 that has led to livestock overgrazing, overuse of water and loss of trees. In the remote mountainous region of Jumla, for example, around 90 percent of farmers subsist on agriculture production but their landholdings are so small, it’s challenging to harvest sufficient food. So often, families collect firewood, herbs and other ecosystem goods to augment their income and food resources. Compounding the situation, major changes in climate have reduced water availability, increased temperatures and produced a shift in growing seasons—all of which impact agricultural production, and further drive people to draw from the environment for goods. The geographic isolation of Jumla and other remote areas have made it difficult for Nepalese authorities to readily see how inextricably linked agricultural practices and environment deterioration have become. This has presented authorities with their own dilemma: how to develop initiatives to both improve agribusiness and strengthen the resilience of the natural resource environment.

Intent on improving Jumla’s plight, MoAC in 2010 launched the three-year High Mountain Agri-business and Livelihood Improvement (HIMALI) project. Specifically targeting two watersheds, one of which was the Lorpa watershed, HIMALI’s goals were to enhance the communities’ socioecological resilience to climate change and help design effective local watershed management plans to ensure the sustainability of agribusiness in the region. To identify and recommend solutions, project managers needed to better understand Lorpa’s present vegetation and how its land cover—specifically the forests—had changed over time. Given its geographic isolation, acquiring that picture with traditional, physical surveys would not be feasible. In order to both assess and map the area’s tree cover from the past to the present, the project team needed to have geospatial imagery and image analysis technology that would allow them to classify the vegetation, inventory the forest cover down to the tree-crown level and map that over time. The land classification solution also needed to be able to sufficiently handle the complexities of comparing and classifying vegetation within the challenging mountainous environment of steep slopes and dramatic changes in vegetation.

The responsibility of producing the vegetation-change detection maps fell to Kathmandu’s International Centre for Integrated Mountain Development (ICIMOD), a regional knowledge centre serving the eight countries of the Hindu Kush Himalayas. An ICIMOD team acquired one 2006 QuickBird satellite image and one Ikonos image from 2011 for the change-detection land-cover maps. They also obtained a digital elevation model (DEM) from the shuttle radar topography mission (SRTM) for topographic detail as well as vector data such as buildings, roads and contours. As QuickBird and Ikonos have different resolutions, they integrated the ancillary data with each orthorectified satellite scene to create two separate rule sets within their Trimble® eCognition® software. After pre-processing and validating the quality of the raster data, the data-processing team used Esri’s ArcGIS to calculate multiple indexes to help separate vegetation from non-vegetation areas detail that would be integrated into the classification process. They then wrote customized rules and built two eCognition rule sets to distinguish and map 10 land-cover classes. Specific to the forest cover classes, they instructed the software to automatically delineate individual tree crowns into five size categories. Once the rule sets were complete, it only took the software 30 minutes to run the workflows and produce land-cover maps for 2006 and 2011 showing the change in the Lorpa’s vegetation between those two years. To validate the accuracy of eCognition’s automatic tree crown detection, the ICIMOD team chose ten 2.5-acre (1-hectare) sections at random on each pan-sharpened QuickBird and Ikonos image and manually digitized each visible crown in ArcGIS. They then compared their delineations with the software’s and found the 2011 tree crown classification was 99 percent accurate; the accuracy for 2006 was 97 percent.

• By compiling accurate inventories and maps of the Lorpa watershed’s tree canopy, ICIMOD was able to bring the remote view of the region’s environment to Nepalese authorities’ desktops.
• In presenting the vegetation-change detection maps to officials with both the MoAC and the local district forest office, there were audible 'gasps' at the clearly visible deforestation in the watershed area.
• The classification datasets allowed HIMALI project managers to see how the vegetation has changed over five years, enabling them to better understand the region’s roots of historic deforestation and watershed erosion.
• With the knowledge from the maps, authorities can prioritize investment areas and develop community-centric strategies to help the villages organically grow a healthier agribusiness.
• The detail of the maps is being used to help devise forestry-specific management programs for the watershed area, a critical part of successfully improving the agribusiness in the region and the livelihoods of its people.

• The maps showed the Lorpa watershed suffered a reduction of 5,432 trees or a loss of 12-percent tree canopy between 2006 and 2011.
• The accuracy of the 2011 tree crown classification was 99 percent; the accuracy for 2006 was 97 percent.
• The overall accuracy of the land classification map was 93 percent when compared with manual assessment.