Modeling Past Vegetation Change Through Remote Sensing and G.I.S: A Comparison of Neural Networks and Logistic Regression Methods

Past change in remnant vegetation patches was modeled using remotely-sensed MSS and TM imagery and G.I.S. The images covering 27 years from 1973 to 2000 were used in detecting change in vegetation through post-classification comparison and modeling it through neural networks and logistic regression methods. Physical factors, image-based layers and landscape metrics provided the 19 predictor variables used in the modeling. The area of study is the catchment of the Boorowa River in New South Wales, Australia, around 110 kms northwest of Canberra. Modeling decrease in vegetation patches over the past 27 years through neural networks and logistic regression methods proved successful. Relative operating characteristic (ROC), and a modified version of multi-resolution goodness of fit (MGF) tests together with visual comparison were used to assess success of the modeling approaches. Also, the relative effect of the 19 predictor variables were evaluated through ROC and MGF methods using 19 reduced-variable models and the full model. Overall, the neural networks method performed slightly better than the logistic regression. A surrogate layer for agricultural intensity (MAXNDVI) and the ratio of MSS band 4 to NDVI (MSS4toNDVI) were found the most important predictor variables of change in vegetation. This was also the case with the logistic regression method where additionally the slope and Para (perimeter to area ratio of the patches) parameters were demonstrated to be the other important predictor variables. The neural network method was found to be more sensitive to inclusion or exclusion of the variables as compared to the logistic regression method. The increased number of variables also created a somewhat dispersed pattern of modeled vegetation as opposed to the naturally more clumped vegetation pattern.