The Role of Landscape Metrics and Spatial Processes in Performance Evaluation of GEOMOD (Case Study: Neka River Basin)

Performance evaluation is crucial for land cover change modeling. The main objective of this study is performance evaluation of GEOMOD using landscape metrics and spatial processes in landscape transformation for modeling change patterns of forest cover in Neka River Basin, north of Iran. Therefore, the land cover maps for the years 1984, 2001, and 2010 were used as the observed land cover maps. Suitability map from forest to non-forest was produced using weighted linear combination algorithm. Fuzzy membership functions and analytical hierarchy process were used to standardization and weight of criteria, respectively. We calculated the indices including class area, patch density, edge density, mean fractal dimension index, interspersion and juxtaposition index, effective mesh size, and mean related circumscribing circle using Fragstats and spatial processes using decision tree algorithm in Land Change Modeler. The relative error obtained by comparison of observed map versus simulated map for patch density, related circumscribing circle, and for effective mesh size metrics was the highest. The model was able to predict shape complexity, fragmentation, compactness and spatial heterogeneity, and area of forest class with high consistency. Landscape transformation analysis determined attrition according to the decrease in patch density and area of forest. Besides, the model predicted the same spatial process. The results of this research showed that this method can produce comprehensive information with high performance from uncertainty of result accuracy.
Extended Abstract
1-Introduction
Land cover changes are recognized as a major driver of global ecosystem changes as well as key factor in global climate change (Tian et al., 2011). In order to analyze and predict these changes, researchers have designed different types of models. There are several sources of uncertainty in simulation models that can be categorized into three categories including data, model, and process of future change (Pontius & Neeti, 2010). Several models have been developed to predict land cover changes such as CLUE-S (Verburg et al., 1999), DINAMICA (Soares-Filho et al., 2002), Land Change Modeler (Eastman, 2006), CA-Markov (Eastman, 2006), and GEOMOD (Pontius et al., 2001). GEOMOD is a model based on Geographic information system that can simulate the location of deforestation zone using bio-geographical and socio-economic characteristics as well as spatial data of forest in different periods (Echeverria et al., 2007). As an advantage of GEOMOD,  compared with complex models, it does not require large amounts of data for calibration and validation (Echeverria et al., 2007). Landscape metrics can lead to an increase in interpretation and better evaluation of the results of land cover change models. The main objectives of this study are (1) simulation of Hyrcanian Forest changes in Neka River Basin using GEOMOD; (2) performance evaluation of the GEOMOD model using landscape metrics and spatial processes in landscape transformation.
2-Materials and Methods
Neka River Basin is located in the east of Mazandaran province, north of Iran. The district lies between 53° 17′ 30″ to 54° 44′ 22″ E and 36° 27′ 46″ to 36° 41′ 8″ N. GEOMOD is used to predict dynamics of Hyrcanian Forest for the years 2001, 2010 and 2022 (because the observed land cover maps for the years 2001 and 2010 were available to compare with simulated land cover maps). In this study, a multi-criteria evaluation procedure was used to generate the transition suitability map from forest to non-forest. Multi-criteria evaluation consists of three main procedures: Boolean intersection, Weighted Linear Combination (WLC), and Ordered Weighted Average (OWA). In the present study, WLC was employed to combine factors and constraints. Factors were standardized in spatial information system using fuzzy membership functions. Six factors including distance from residential area, distance from agricultural land, distance from rangeland, distance from road, elevation and slope were employed. Analytical Hierarchy Process was used to weight the criteria using pairwise comparison (Moeinaddini et al., 2010). Seven landscape metrics were selected for accuracy assessment of model using Fragstats software (McGarigal et al., 2002). The relative error was calculated to quantify the difference between landscape metrics derived from the observed and the simulated layers (Sakieh and Salmanmahiny, 2016). Analysis of spatial processes was also calculated to evaluate the model performance using decision tree algorithm in Land Change Modeler (Bogaert et al., 2004).
3-Results and Discussion
Forest showed a decrease of 3000 ha (4.1 % of forest area, and 1.6 % study area) from 1984 to 2001. This class decreased from 69169 ha to 67198 ha (2.8 % of forest area, and 1.1 % study area) between 2001–2010 period. According to the modeling results, a decrease 4225 ha was revealed in this class of land cover.
The area under forest showed a decreasing trend from 2001 to 2010, and the model showed a good consistency between the forest areas of reference and simulated maps with a relative error value of zero. Observed maps depicted a decreasing trend for patch density during 1984–2010 (from 1.5 to 1.3). GEOMOD could also predict a decreasing trend for this index. Difference between the real data and modeling effort in 2010 was more than 2001, and the model was not able to simulate the patch density with high accuracy. Model was not able to predict the values of patches per unit area, but it is well predicted area and distribution of largest patches according to good agreement of the forest area for 2001 and 2010. Edge density decreased in ground truth layer during 1984–2010 (from 14.7 to 12.9 m/ha). The model results also demonstrated a decreasing trend with relative error values of < 1 in both years for this index. Decrease in edge density of simulated and reference maps suggest the reduction of spatial heterogeneity and conversion of forest patch to non-forest class (Munsi et al., 2010). Mean fractal dimension index indicates the complexity of the patch shape. According to the results of this metric, forest in the study area showed simpler in shape during 23 years. GEOMOD also predicts less complexity in shape but the difference between ground truth data and simulation in 2001 (relative error of 0.4 %) was more than 2010 (relative error of 0.1 %). Reference data for mean related circumscribing circle index showed decreasing trend (from 0.33 to 0.30) during 1984–2010. In terms of this index, the model generates good agreement between reference and observed maps in 2001, and 2010 with relative error values of 5 %, and 0.4 %. Spatial process was attrition in the ground truth data due to the reduction in area, and number of patch per unit area during 1984–2001, and 2001–2010. Similarly, the simulation outputs represent attrition for two periods.
4-Conclusion
Landscape metrics lead to increase and improve performance evaluation of land cover models. Type and number of indices used in various studies are different and various metrics are recognized effective and useful in assessing the characteristics of models. In the present study, seven metrics were used to evaluate the model performance in simulating of forest pattern. These metrics indicated high potential for evaluating simulation success based on description of shape, density, aggregation, interspersion and juxtaposition, and proximity of patches. Besides, the results of this study showed that spatial change processes are also useful tools for evaluating the performance model.