This webinar presentation discussed remote sensing analyses carried out within the NASA grant on “Forest, agricultural, and urban transitions in Mainland Southeast Asia: Synthesizing knowledge and developing theory”. This research fulfils part of the grant objectives by mapping the expansion of tree crops for seven selected Landsat footprints in montane mainland Southeast Asia (MMSEA) (see Figure 1). Specifically we sought to map different tree crops (e.g. rubber, eucalyptus, and cashews); ascertain the date/period of change; and identify the land cover that preceded the change.

Figure 1. The expansion of rubber and cashew plantations from 2000 to 2014 in Northeast Cambodia. Yellow polygons represent areas that were granted for tree crop concessions (source: www.opendevelopmentcambodia.net)

Mapping land use and land cover change in MMSEA is no easy task given the heterogeneity of this fragmented landscape and the speed at which it is changing. We needed an approach that could identify subtle changes in the spectral reflectance of remotely sensed images, as well as cope with the cloud cover and data gaps that limit the information available in the Landsat time series. In the context of mapping urban expansion, researchers faced similar issues and found that a supervised classification of a dense time stack of Landsat data without masking clouds or data gaps was able to deal with these issues.

We considered this approach suitable for our classification problem and performed a supervised classification using the support vector machines (SVM) classifier and a dense time stack of Landsat data (3-6 images per year from 2000 to 2014). The classification scheme consists of stable land cover classes (e.g. water, forests) and the classes that showed change (e.g. rotational agriculture, conversion to tree crops (including information on the date and type of change)). We generated random and strategic points for each Landsat footprint and labelled them using field photos, Google Earth, and the Landsat data as reference. We then split the points into classification training (75%) and verification (25%) areas and used the training points to parameterize the SVM classifier by performing a two-dimensional grid search with internal validation to find the most suitable parameter settings. Each Landsat footprint was classified individually using the best parameter setting.

Figure 1 shows our classification for a region in Northeast Cambodia where rubber and cashew plantations occur. While evergreen and deciduous forests still prevail, large forest areas have been converted to rubber plantations. Comparably few areas were converted between 2000 and 2009 and the majority of plantations were established after 2009. The areas around Banlung and to the south of Stung Treng show smaller and more fragmented plantations than those in the center of Figure 1. Most likely these smaller plantations are owned by smallholders while the large patches represent concession areas. A map of tree-crop concession boundaries (yellow polygons in Figure 1) supports this assumption, but also highlights disagreements between mapped concession boundaries and our classification. Most of the mapped concessions have only been partially converted and there are large scale plantations (most likely concessions), that are not mapped. It will be interesting to look further into this – e.g. what triggered the establishment of smallholder plantations or why some concessions were not (fully) converted.

An assessment of the accuracy of seven Landsat footprints showed an overall accuracy of almost 89% with only a few classes showing user and producer accuracies below 75%. This suggests that the supervised classification of a dense time stack of Landsat data provides a good approach for mapping the expansion of tree crops. We found two ways to minimize misclassifications. First, forest clearances are usually more significant than the spectral reflectance of tree crops or other land use types that follow the clearance. Any land use activity that involves the clearance of forest should therefore be included as a separate class. Second, data gaps and clouds can lead to misclassifications, especially when only a few and clustered training points are available. In such cases the classifier is trained to detect a specific sequence of data gaps that occurs at the location of the clustered training points instead of the actual LCLUC. With an even distribution of classification training and verification points for most of the classes this issue can be reduced.

Within this NASA grant the Landsat classifications are used for two tasks. They support qualitative research that focuses on documenting teleconnections in rapidly changing places to produce an integrated understanding of LCLUC. The Landsat classifications are also used to derive training areas for the classification of the broader MMSEA region using MODIS time series data. Read more about the project