Generating High-Quality Landsat Time-Series and Its Applications in Forest Studies

Forest covers about 40 percent of the Earth's total land surface and is of tremendous ecological and economic value. Spatially explicit knowledge of forest composition and biophysical attributes is very important for monitoring forest development and for informing decisions for sustainable development. Landsat images have been widely used to map forest composition and estimate biophysical attributes but the accuracy is not yet satisfactory. This dissertation seeks to develop new approaches to produce high-quality seasonal Landsat time-series and then classify detailed forest types and model forest aboveground biomass (AGB). First, a new method for removing thick clouds was developed based on a modified neighborhood similar pixel interpolator (NSPI) approach. Second, a new Geostatistical Neighborhood Similar Pixel Interpolator (GNSPI) was developed for gap-filling. After cloud removal and gap filling processes, we can obtain high-quality Landsat time-series data. Third, a hierarchical classification method was proposed to get detailed forest types from dense Landsat time-series to improve forest mapping accuracy. This method integrates a feature selection and an iterative training-sample-adding procedure into a hierarchical classification framework. The proposed method has been tested in Vinton County of southeastern Ohio. The accuracy of these forest types reaches 90%. Last, NDVI time-series derived from six Landsat images across different seasons was used to estimate AGB in southeast Ohio by empirical modeling approaches. Results clearly show that NDVI in the fall season has a stronger correlation to AGB than that in the peak season, and using seasonal NDVI time-series can obtain more accurate AGB estimations and less saturation than using a single NDVI. This study demonstrates the value of multi-seasonal Landsat images for improving forest studies.