Search

This research uses Markov random fields to model stochastic interactions among classes over space and time, which allows a global Bayesian optimization of the classification result. A maximum likelihood supervised classification method is used to derive initial classification result purely based on the spectral information. Both the spatial and temporal contextual information are derived and represented from modeling the image pixel class labels in predefined spatial and temporal neighborhoods. In this study, we improve the spatial context by introducing the concept of distance decay and dealing with missing data by making the spatial context locally adaptive. For modeling the temporal context, we explore and evaluate a new method called “cubic spatio-temporal neighbor” besides commonly adopted transition probability matrices (TPMs) by considering the temporal correlation length and phonological pattern. The optimization of MRF model parameters (weights) for combining spectral, spatial and temporal information in the classification is an important problem. The possible occurrence of strongly undesirable negative parameters which are neglected in the literature is dealt with in this paper by using the scaled difference in coefficient estimates (DFBETAS) which is normally used in the field of Statistics. We have implemented the method using C# programming language and successfully applied it to the urban land cover classification with Landsat image time series over a decade and to the snow surface melting detection in Antarctic ice sheet with daily satellite observations. Our application examples demonstrate that the proposed algorithms are effective and efficient in modeling contextual information and optimizing parameters, and the incorporation of spatio-temporal information with our MRFs method results in significant improvement as large as 8.9% in classification in comparison with the use of spectral information alone.

Early word learning takes place across different contexts. For example, in a single day a child may hear the noun “cup” in a wide range of places (e.g., in the car, at the playground, in the stroller) and from a wide range of speakers (e.g., mother, father, sibling). Understanding the role of spatial and speaker context in word learning is important because context affects learning and memory (Godden & Baddeley, 1975; Hayne, MacDonald, & Barr, 1997; Rovee-Collier & Default 1991; Smith 1982; Smith, Glenberg, & Bjork, 1978 and Vlach & Sandhofer, 2011). By examining the role of context using two methodologies to examine multiple contexts, this dissertation suggests context is an integral component of the basic characteristics of word learning. In Paper 1, I experimentally asked what role visual attention plays in an infant’s category generalization in a new context. Forty-eight English monolingual infants, ages 16-20 months (25 males, Mage = 17.42 months, SDage= 1.46 months), were presented with eight novel noun generalization categories. During the learning phase infants were presented with five category exemplars in either: (1) all of the same background context, (2) all varied background contexts, or (3) a combination of same and varied background contexts. All infants’ category generalization performance was tested in a never before seen context. Results suggest that visual attention during learning is associated with category generalization abilities in a new context only for infants whose learning took place in a combination of same and varied background contexts. In Paper 2, I used naturalistic methods to examine what contexts children were in when they (a) were exposed to nouns and (b) produced nouns. Eight families, whom were selected from a larger study conducted by the UCLA Sloan Center on Everyday Lives of Families (CELF: Ochs & Kremer-Sadlik, 2013), participated in the current study. All families had a focal child who was between the ages of 1.5-4.5 years old. Families were filed over multiple days as they carried out everyday-family activities. Video recordings were coded for the spatial and speaker contexts in which the focal child’s language input and output took place. The results describe the spatial and speaker contexts in which children are exposed to and produce nouns. Further, results suggest a positive association between the number of spatial and spatial contexts in which children were exposed to words and the children’s rate of production of those words. By examining how contextual factors affect word learning, this dissertation elucidates some of the learning mechanisms children use to learn language. Moreover, by demonstrating the effect of context on language, this dissertation will validate the need to include contextual factors in word learning research. Together, the results of these studies add to a small but growing body of research suggesting the environmental factors in children’s early learning environments that may support early language learning.