1. Geospatial statistics and geocomputation
Led the proposition and development of the Bayesian “Markov chain random field” (MCRF) spatial statistical approach (see “Markov chain geostatistics”), including the MCRF theory, specific MCRF models and simulation/cosimulation algorithms, and transiogram concepts and joint modeling methods, during last twenty years. The MCRF theory was based on three basic ideas - spatial sequential Bayesian updating, spatial conditional independence assumption, and the transiogram spatial correlation metric. The approach is mainly for simulating categorical or discrete spatial variables. The development of this approach initially aimed to solve a series of scientific issues I encountered (e.g., defects of a coupled Markov chain model that I discovered in 1999 in my postdoc research) in multidimensional Markov chain modeling in order to develop a predictive soil mapping method, and later aimed to extend Markov chain from a one-dimensional stochastic simulation model to a multidimensional generalized geospatial statistical approach for simulating categorical spatial variables (e.g., various landscapes). The MCRF approach is in accordance with conventional Markov chain theory, Bayes' theorem, conventional geostatistics, and pioneer studies (before 2004) in Markov chain spatial modeling. The generalized MCRF model is a spatial Markov chain locally-conditioned through sequential Bayesian updating on nearest data in different directions. Such a locally-conditioned Markov chain is not a traditional Markov chain anymore. Essentially, it is consistent with a Bayesian network over spatial data. Therefore, the MCRF model represents a fundamental Bayesian spatial model at the neighborhood nearest data level, and the MCRF approach represents a Bayesian geospatial statistical approach. It is the result of a decades-long exploration in Markov chain spatial modeling and geospatial statistics since early 1990s. Even though encountering many difficulties, we still made our effort and developed the MCRF model into some practical methods for real applications.
In the MCRF approach: the spatial sequential Bayesian updating on nearest data was first used by Li (2007a) in spatial statistics (for deriving the MCRF model); the spatial conditional independence assumption of nearest data was first suggested and used by Li (2007a) in spatial statistics (for simplifying the multiple-point likelihood terms in the MCRF model) (conditional independence assumption was thought to be wrong in geostatistics previously); as to the transiogram spatial metric, it was systematically described and analyzed with the proposition of some practical joint modeling methods in a few articles of ours (e.g., Li 2007b) based on the variogram concept, Markov chain theory, related pioneer studies in transition probability-lag diagrams and some new ideas, because it is indispensable in Markov chain geostatistics. In addition, the defects (mainly small class underestimation and layer/parcel inclination in simulated realizations) of the coupled Markov chain model were initially discovered in 1999 in a research that tried to use the model, much earlier than the time the MCRF model was proposed, and they were proven facts rather than something that can be denied by tricks. The misunderstandings on the MCRF approach were scientifically irrational.
In addition, we also used other geospatial statistical methods, such as kriging and GWR (geographically weighted regression), and nonspatial statistical methods in our researches.
2. Resource and environmental informatics
In recent years, our researches in resource, environemntal and ecological informatics were mainly focused on following topics: (1) Land use/cover spatial variation and their time changes. (2) Spatial distributions, uncertainty, and ecological/health risks of pollutants (mainly in soil). (3) Predictive mapping of soil properties (soil organic matter and nutrients) and classes (types, texture, and structure). We conducted these studies using a variety of spatial statistical methods and other methods.
3. GIS and remote sensing of environment
Since 2002, we have been involved in studies in Internet GIS, remotely sensed image processing (e.g., image restoration, land cover/use classification and postclassification), and land use/cover study. Our remote sensing application studies were mainly focused on quantitative processing and analyses of spatial data.
4. Geography and spatial variation of soil
Very early researches involved these areas, such as 1D Markov chain modeling of soil layer vertical change, spatial variations of soil layer thickness and emerging depth and their effects on field water capacity, stochastic simulation of field water and solute transport along soil profiles, and the physical-chemical properties, amelioration, genesis and classification of black clay soils in China. These researches were important topics in soil science at that time in China.