Research

REDD

Clark Labs plays a key role in REDD (Reducing Emissions from Deforestation and Forest Degradation) initiatives through training, advising and software development. Robust modeling tools such as Land Change Modeler and GEOMOD give practitioners the ability to address the complexities inherent in REDD projects. Fully integrated into Land Change Modeler is support for REDD forest project planning aimed at Reducing Emissions from Deforestation and Forest Degradation. The REDD tool facilitates the estimation of baseline emissions from various carbon pools and allows the calculation of deferred emissions and carbon credits.

Also included in TerrSet for REDD modeling is the GeOSIRIS application. Where the REDD tool in Land Change Modeler is mainly meant to be project specific, the GeOSIRIS application is a national-level REDD planning tool for quantifying deforestation, carbon emissions, agricultural revenue, and carbon payments.

The TerrSet software brings a complete land analysis toolkit, compatible with international requirements, for mapping historical baselines and modeling future scenarios, to the REDD community. Its integrated and comprehensive features allow you to process your satellite data for the land cover mapping component as well as manage and conserve forest carbon, biodiversity, and related ecosystem services. The participation of local stakeholders is also possible since TerrSet includes a unique suite of spatial decision support tools.

Machine Learning and Neural Networks

Clark Labs pioneered the introduction of integrated neural networks and has become the leader in the development of the first ever machine learning procedures in a GIS and Image Processing system. Neural networks and related machine learning approaches are so important since they do not depend upon restrictive assumptions about the underlying character of class distributions and are capable of learning complex patterns with limited data. TerrSet is the premier system for integrated neural network and machine learning solutions with the introduction of:

• An advanced Multi-Layer Perceptron (MLP) neural network classifier with the first ever automatic mode supervised training, progressive learning rate adjustment and hidden layer mapping with linear output option. These options allow MLP to be run without significant user input, unless desired.
• A Self-Organizing Feature Map (SOM) neural network which uses a two-dimensional neuron topology with both supervised and unsupervised output options.
• A Fuzzy ARTMAP neural network with both supervised and unsupervised output options.
• A Radial Basis Function neural network classifier.
• An integrated Decision Tree machine learning classifier based on the ID3/C4.5 algorithm.
• The first implementation of the K-Means unsupervised classification procedure as a machine learning algorithm with dynamic feedback and direct training intervention.

Soft Classifiers

TerrSet includes the most extensive set of soft classifiers in the industry. Soft classifiers express the degree of support for each of a set of potential land cover classes at each pixel location. Thus, rather than a single map of most likely class membership, a set of images (one for each class) is produced expressing the degree of support. Soft classifiers can be used for a variety of purposes including uncertainty management (i.e., Why is the classifier having difficulty classifying this pixel?) and sub-pixel classification (i.e., What are the proportions of cover types mixed into this pixel?) Specific innovations developed by Clark Labs include:

• First-of-its-kind, innovative solutions to the band limitations of linear spectral unmixing. Normally, the number of parent classes (end members) in sub-pixel classification is limited to the number of input bands. In TerrSet, this limitation is removed through a logical pairing of soft classifier approaches.
• The first introduction of soft classifiers based on Bayesian, Mahalanobis Typicality, Dempster-Shafer Belief and Plausibility, and Fuzzy Set membership metrics.

Multi-Criteria / Multi-Objective Decision Making

In 1993, Clark Labs introduced the first instance of Multi-Criteria and Multi-Objective decision making tools in GIS. Years later, Clark Labs is still the industry leader, responsible for:

• The first implementation of the Ordered-Weighted Average for multi-criteria evaluation that allows one to balance the relative amount of tradeoff between criteria with decision risk in balancing discordant information.
• The first implementation of the MOLA heuristic for multi-objective land allocation which now included modeling of compactness and contiguity.
• The first GIS software implementation of Saatys Analytical Hierarchy Process (AHP).

Uncertainty Management

The first great horizon for GIS was conquering complexity. Computers and software have done that exceptionally well. The next great horizon is the conquest of uncertainty. Clark Labs has taken a pioneering role in this area with the following selective developments in TerrSet:

• The first-ever implementation of a Dempster-Shafer evidence aggregation procedure in GIS.
• The first soft reclassification procedure (PCLASS) that allows one to map the probability of a location being above or below a threshold (such as sea level rise).
•  first ever module to generate normal and rectilinear distributions for uncertainty analysis such as Monte Carlo.
• The only implementation of spatial prior probabilities for Maximum Likelihood classification.

Spatial Processes

Clark Labs has always been recognized as a pioneer in the analysis and modeling of spatial processes. Specific innovations include:

• The only anisotropic cost distance analysis procedure in the industry. This is a procedure that recognizes that costs vary according to how you are moving through a cell. For example, moving uphill acts as a friction whereas moving downhill acts as an accelerant.
• In support of anisotropic cost analysis, Clark Labs introduced innovative procedures for force vector analysis including the ability to determine resultant vectors.