An indicator of forest dynamics using a shifting landscape mosaic

Abstract

The composition of a landscape is a fundamental indicator in land-cover pattern assessments. The objective of this paper was to evaluate a landscape composition indicator called ‘landscape mosaic’ as a framework for interpreting land-cover dynamics over a 9-year period in a 360,000 km² study area in the southern United States. The indicator classified a land parcel into one of 19 possible landscape mosaic classes according to the proportions of natural, developed, and agriculture land-cover types in a surrounding 4.41-ha neighborhood. Using land-cover maps from remote sensing, the landscape mosaics were calculated for each 0.09-ha pixel in the study area in 1996 and 2005. Mosaic transition matrices estimated from the pixel change data were then used to develop two Markov chain models. A “landscape mosaic” model was a temporal model of the shifting landscape mosaic, based on the probability of landscape mosaic change for all pixels. A “forest security” model was the same, except that the Markov states were defined by both the landscape mosaic and the land-cover of each pixel, which allowed interpreting forest land-cover dynamics in the context of a shifting landscape mosaic. In the forest security model, the overall percentage of forest decreased from 33% in 2005 to 17% at steady-state, and there was little change in the relative distribution of existing forest area among landscape mosaic classes. In contrast, the landscape mosaic steady-state was reached later, and indicated that a maximum of 10% of total area was available for forest. The implication was that forest security depended ultimately on the dynamics of the landscape mosaics that contained forest, not on forest dynamics within those landscape mosaics.

Introduction

The composition of a landscape is a fundamental indicator of patterns within that landscape (O’Neill et al., 1988, Li and Reynolds, 1994), if only because few other pattern indicators are interpretable independent of landscape composition (Gardner et al., 1987, Gustafson, 1998). Pattern indicators for large-area ecological assessments from land-cover maps must include composition indicators that are applicable to all land-cover types and interpretable with respect to many societal and ecological concerns like biodiversity, urban sprawl, and water quality. Establishing a foundation for using those indicators requires testing them with neutral models, examining their statistical properties, and learning how to interpret them (Turner et al., 2001). The objective of this study was to lay part of the foundation for a landscape composition indicator called ‘landscape mosaic’ through analysis and interpretation of land-cover changes from 1996 to 2005 in a study area along the south coast of the United States.

The landscape mosaic indicator comes from the ‘landscape pattern type’ indicator (Wickham and Norton, 1994) which has been adapted for large-area pattern assessments using land-cover maps derived from remote sensing (Riitters and Wickham, 1995, Jones et al., 1997, Riitters et al., 2000). The indicator classifies a land-cover pixel according to the land-cover composition in a fixed-area neighborhood surrounding that pixel, and a map of landscape mosaics may be constructed by classifying every pixel on the land-cover map. A map of landscape mosaics can help to visualize ‘interface zones’ (e.g., the ‘forest–urban interface’) and other spatial gradients of land-cover composition across a region (e.g., Riitters et al., 2000), but more work is needed to develop the indicator as a framework for interpreting land-cover dynamics.

To evaluate the landscape mosaic as a framework for interpreting pixel-level forest dynamics, we defined forest security as the likelihood that a pixel of forest remained as forest over time, which depended on the landscape mosaic that contains a forest pixel. In a dynamic landscape, land-cover change at the pixel level also changes the mosaic at the landscape level, so that the security of a particular pixel of forest can change over time even if that pixel remains as forest. One example is urban sprawl at the forest–urban interface. Forest clearing to build a house is likely within the forest–urban interface because of infrastructure such as roads and water service. Forest removal (a pixel-level change) changes the forest–urban interface (a landscape-level change) by altering its composition and ultimately shifting its location. Additional forest clearing in the same neighborhood may become more or less likely in the future, and the likelihood of forest clearing in new neighborhoods may change as the forest–urban interface itself moves. If forest security depends on forest location relative to a forest–urban interface, then forest security changes when the forest–urban interface moves.

The modeling problem in our study was to link forest change at the pixel level with mosaic change at the landscape level. Gustafson (1998) noted that the conceptual model of a shifting landscape mosaic (Bormann and Likens, 1979) has been useful for describing landscape dynamics, but that the temporal component of that conceptual model needed more development. In contrast, Markov chain models have a long history in temporal analysis of landscape change, but most implementations are not spatial models (see reviews by Baker, 1989, Brown et al., 2004). For Markov models, the most common spatial approach has been to estimate pixel-level transition probabilities based on the spatial attributes or context of each pixel, such that the probabilities of transitions among Markov states can change over time (Brown et al., 2004). An alternate approach to incorporate spatial information was used by Flamm and Turner (1994) in a ‘patch’ transition model. In that study, the transition probabilities were held constant, and the Markov state of a patch was defined by enumerating its spatial attributes such as soil type, land ownership, and vegetation cover type. We used a similar approach at the pixel level in this study, whereby the Markov state of a pixel was defined both by its land-cover and by the landscape mosaic that contained it. We considered two implementations of that basic model, including a temporal model of the shifting landscape mosaic for all land-cover types, and a model that distinguished between forest and nonforest pixels. The first model was used to characterize the long-term or steady-state distribution of all pixels among landscape mosaics, and the second model allowed interpreting forest dynamics and steady-state distributions of forest in the context of a shifting landscape mosaic.