Date of Award

2004

Level of Access

Campus-Only Thesis

Degree Name

Master of Science (MS)

Department

Forest Resources

Advisor

Steven A. Sader

Second Committee Member

Robert G. Wagner

Third Committee Member

Alan S. White

Abstract

Monitoring the composition of regeneration stands is an important goal of forest management in Maine. Critical investment decisions concerning silvicultural treatment need to be made early in stand development as these treatments (or no treatment) will largely define the future composition and thus the value of the products from the stand. A methodology was developed and tested using regression modeling for combining high spatial and moderate spatial resolution satellite imagery to determine forest type and density in a northern Maine study area. First, high spatial resolution (4-meter) binary softwood and hardwood maps were created using multi-temporal IKONOS imagery. Then, regression relationships (63 different models) for the individual binary maps and Landsat ETM+ variables (multispectral values and vegetation indices), including several dates/seasons of imagery (August, September, October, May), were tested. Employing the Akaike's information-theoretic approach, the best regression models from the a prion set were selected. Comparison of the Akaike's information criterion (AIC) provides a straightforward and objective method for examining and determining acceptable regression models. A model incorporating variables from four dates of Landsat ETM+ imagery produced the best results (lowest AIC value) in predicting both softwood and hardwood cover. However, the two-date combination of October and Mav variables provided an acceptable alternative model as the adjusted R2 value was 0.(->4 compared to 0.65 for the best 4-date model. The best single-date model produced inferior results with a sharp drop in AIC values and adjusted R2, compared to the multi-date models. Spatiallv-explicit maps of softwood and hardwood forest cover density were generated. These maps were compared to the forest landowner's geographic information system (GIS) stand cover maps and agreement assessments were performed. The best model produced a map with an overall agreement of 79%. The producer's accuracies and user's accuracies were 94% and 82%, respectively, for softwood and 76% and 85% for hardwood The best spatially-explicit forest type and density map was examined for selected regeneration areas and areas that had been pre-commercially treated. Comparison of stand composition before and after treatment of these sites is one example of how the combined map products could be utilized as a tool for foresters in planning and evaluating silvicultural treatments. Although the models produced good results for mapping softwood and hardwood cover and density in this studv, future work is needed to investigate models with additional variables (different vegetation indices, image texture information, site attributes) to improve regression relationships. To evaluate model performance and cost-efficiency in producing the forest cover maps, it is recommended that future work address alternative sources of imagery (e.g., scanned aerial photos) to create the high resolution maps. The Landsat-derived forest cover map provides density information (percent cover) for every 28-meter pixel (approximately 1/4 acre), which is more detailed than traditional maps that the land managers maintain in their GIS. Given the low cost per area of Landsat imagery, the regression method applied to spring and fall Landsat imagery for predicting softwood and hardwood densities may be a potentially cost-effective tool for examining spatiallv-explicit, regeneration stand conditions over large forest ownerships in Maine.

Share