This application compares changes between aggregated 2011 National Land Cover Database land cover
categories with similarly aggregated land cover categories from The Clark Labs 2050 Conterminous US Land
Cover Prediction. It also provides a few summary statistics about possible changes in developed, forest
and agricultural land cover. Look for the soon to be released Clark Labs American Land Change Explorer
application, which provides exhaustive analysis and summaries of potential transitions from each of the NLCD
categories to each of the projected 2050 categories.
Introduction
The Clark Labs’ conterminous US land cover prediction for 2050 was produced as part of the development of the
Land Change Explorer – a web application to illustrate the potential of predictive land change modeling and to
introduce potential users to the Land Change Modeler – a cloud-based software service for land change modeling
to be offered in the ArcGIS Marketplace.
Procedure
The prediction is based on an empirical modeling of the relationship between land cover change from 2001 to 2011
and a series of explanatory variables. The land cover data were at a 30 meter resolution from the National Land
Cover Database (NLCD). The explanatory variables1 were:
- Elevation
- Slope
- Proximity to primary roads
- Proximity to secondary roads
- Proximity to local roads
- Proximity to high intensity development
- Proximity to open water
- Proximity to cropland (used only for transitions to cropland)
- Protected areas
- County subdivisions or counties/incorporated places (depending on the state)2
The modeling procedure used is a newly developed algorithm suitable for distributed computing in a cloud
computing environment3. Briefly, the procedure is based of kernel density estimation of the normalized likelihood
of change associated with varying levels of each independent variable. These estimates are then aggregated by
means of a locally-weighted average where the weights are based on the degree of conviction each variable has
about the outcome at that specific pixel. Testing has shown it to be comparable in skill to a multi-layer
perceptron neural network with the added advantages of rapid calculation and capability of being distributed
across multiple computer nodes.
Because the drivers of change can vary over space, modeling was done separately for each state. All transitions
that met or exceed 2 km2 in area (at the state level) were modeled independently. Within a single state, as many
as 128 individual transitions might occur. In total, over the 48 conterminous states, 3330 transitions were modeled.
The modeling process initially establishes the potential to transition. This potential is expressed as a
continuous value from 0 to 1 at each pixel for each transition. The procedure then uses the Markovian assumption
that the rate of transition experienced in the historical period (2001-2011 in this case) will continue into the
future. A competitive greedy selection process then allocates the projected change4.
Validation
In the training process for each transition, 50% of historical instances of change and 50% of an equal-sized
sample of pixels eligible to change, but which did not (e.g., persistence), were reserved for model validation.
The median accuracy over all 3330 transitions was 80% with 79% of change validation pixels being correctly
predicted and 83% of persistence pixels being correctly predicted. Thus the models, on average, are quite
balanced in their ability to predict change and persistence.
The accuracy associated with more specific transitions varied. A key objective was to be able to monitor and
project anthropogenic changes and thus the explanatory variables chosen were focused on such drivers.
Consequently, the median accuracy of natural to developed transitions (such as deciduous forest to low intensity
development) was 92%. Again, accuracy was evenly balanced (93% for change / 91% for persistence).
Accuracy for transitions between developed categories was lower at 77% (80% change and 75% persistence). A
large part of this is because of the inconsistent manner in which roads are classified in the NLCD system. Roads
are classified as one of the developed categories (high, medium, low and open development) based on the amount
of impervious surface detected within pixels. Alignment of image pixels can cause this to vary resulting in roads
that frequently switch classes between the years mapped.
Natural transitions, such as forest to shrub, had the lowest overall accuracy at 74%. This was expected because
many drivers cannot be predicted with the variables used. An example would be forest fires caused by lightning.
This is also reflected in the fact that accuracy for predicting change was only 71% while that for predicting
persistence was 78%.
Finally, in states with significant cropland development, natural to cropland transitions were modeled with a
79% overall median accuracy. Accuracies for change and persistence were 78% and 81% respectively.
Disclaimer
Note that there are many highly plausible future outcomes and the specific scenario presented is only one of
these (albeit judged to be the most plausible). Also note that each state is modeled separately (on the assumption
that drivers of change many differ between states). As a consequence, there may be some mismatches at the
boundaries between states. Generally, these are only evident for states that have large quantities of natural to
natural transitions (e.g., with forest plantation crop cycles or frequent fire) where the accuracy is lower.
Also note that the protected areas layer does not include all protected areas. Some local conservation land may
be missing. Finally, note that the modeling is based on the assumption that rate of change experienced within
the historical period (2001-2011) will persist into the future.
1
Elevation data were from the National Elevation Database while slope was derived from those data. All
roads data were acquired from the Census Bureau TIGER line files for 2014. Earlier road data would have
been preferred, but errors in earlier TIGER line files were deemed to be unacceptable. Country
subdivisions, counties and incorporated places were also acquired from the Census Bureau. Protected areas
came from the Protected Areas Database of the USGS National Gap Analysis Program. All proximity layers
were derived by Clark Labs.
2
In some states, planning jurisdiction is controlled by county subdivisions (such as in New England), while
in others, planning is governed by a combination of counties and incorporated places (such as in many of
the western states).
3
Eastman, J.R., Crema, S.C., and Rush, H.R., (forthcoming) A Weighted Normalized Likelihood Procedure for
Empirical Land Change Modeling.
4
Greedy selection assumes that the specific pixels that will change are those that are ranked the highest.
Conflicts are resolved by assigning them to the transition with the highest marginal transition potential.