Incorporating neighborhood scale effects into land loss modeling using semivariograms

Scale effects are pervasive in geospatial modeling and affect the reliability of analysis results. This paper examines the neighborhood scale effects on the performance of land loss models in coastal Louisiana where Lower Mississippi River Basin is located. The study incorporates both natural and human variables and their corresponding neighborhood scale variables into land loss modeling. Semivariogram analysis was used to determine each explanatory variable’s neighborhood size at which the semivariance between sample points begins to level off. A new ‘neighborhood’ variable for each of those variables detected with a neighborhood size was created using the focal statistics to represent the neighborhood scale effects. Two land loss stepwise regression models, one without neighborhood variables and the other with neighborhood variables, were developed to test if incorporating neighborhood scale effects could improve the land loss model performance. Results show that the model’s overall accuracy improved significantly from 65.43 to 74.43% after including the neighborhood variables. Six neighborhood variables, in addition to 14 original variables, were selected as significant predictors of land loss probability. The six neighborhood variables include distance to the coastline, land fragmentation, oil and gas well density, percent of water area, pipeline density, and percent of the vacant house. The analysis shows that including variables representing the scale effects are critical for better performance in land loss modeling. Study findings add new insights into the complex land loss mechanism and help derive more accurate land loss predictions to inform coastal restoration and management decision-making.