Mapping the Future: Spatial Models of Decadal-Scale Landscape Change in Northern Australia

Abstract

Repeated sequences of digitised and geo-referenced historical aerial photography provide a powerful means of understanding landscape change over time scales beyond conventional ecological monitoring, especially in remote areas. We use this method to develop spatially explicit models of vegetation change at five sites in the Australian monsoon tropics, and assess the goodness-of-fit of the model to observed change. Models were derived by (i) converting the landscape into cells using a grid-overlay method and classifying vegetation as either closed forest (CF) or savanna (SV); (ii) examining repeat sequences of aerial photographs (taken in 1947, 1972 and 1997) to determine the temporal sequence of vegetation change in each cell (i.e., stasis versus change to different vegetation class); (iii) using thematic layers to derive cell-based measures of elevation, slope angle and drainage distance; (iv) counting the number of like-habitat points in the area surrounding each cell, with the fractal dimension of this area also calculated; and (v) using a bootstrapped generalised linear modelling approach (which controls for potential spatial autocorrelation) to derive a cell-based probability map of change based on the landscape features from (iii) and (iv) fitted to (ii) (an example is given in Figure 1). (Figure Presented) Overall, the closed forest expanded by an average of 42 % total coverage over the fifty-year period, although the rate of change varied across sites. The dynamics of the closed forest-savanna system were predicted with reasonable reliability based on a comparison of observed and predicted CF coverage and the ratio of cells correctly versus incorrectly predicted to change from CF to SV, or vice versa. Mean per cent discrepancy between observed and predicted coverage of CF among sites was only 1.03 % (range: 0 to 2.3 %), and the overall spatial agreement between observed and predicted maps was 88.9 % (range: 78.9 to 94.1 %). Model fit based on the average of two 25-year couplets (1947-1972 and 1972-1997) was consistently superior to those based on a single 50-year couplet (1947-1997). Single-couplet fits had a mean deviance from the observed CF coverage of 6.2 % (range: 0.4 to 23.2 %), compared to only 0.9 % (range: 0.2 to 2.1 %) for the average of two couplets. Spatially explicit fits were also superior for two vs. one time couplet (87.1 vs. 85.0 %, respectively). CF expansion occurred most frequently in fire-protected sites along forest edges and regression in the more fire-prone areas. Possible drivers for this expansion may include changed fire regimes associated with the cessation of traditional Aboriginal fire management or the 'fertilizer' effect caused by the continued increase in global atmospheric CO2 over the course of the 20th century. This effect may be changing the competitive balance between C3 trees of the CF the largely C4 tropical grasses of the SV. Our results also demonstrate quantitatively the benefit of the additional information contained in repeat sequences of imagery over 'snapshots' taken only at the beginning and end of an observation period.