These early results were obtained using one year’s set of FIA field data, DISTANCE: EUCLIDEAN WEIGHTING FUNCTION: NO WEIGHTS. NUMBER OF PLOTS: 696 NUMBER OF BANDS: 34 NUMBER OF NEIGHBORS: 1 VOLUME CROSS-VALIDATION RMSE (ALL)= m3, % VOLUME ALL MIN AND MAX =0 and VOLUME ALL MEAN= VOLUME STANDARD DEVIATION (ALL)= VOLUME BIAS (ALL)= Class\Ref >160 User’s acc > Prod.acc VOLUME CROSS-VALIDATION RMSE (LOWLAND)= m3, % NUMBER OF PLOTS (LOWLAND)= 164 VOLUME LOWLAND MIN AND MAX= 0 and VOLUME LOWLAND MEAN= VOLUME STANDARD DEVIATION (LOWLAND)= VOLUME BIAS (LOWLAND)= Class\Ref >160 User’s acc > Prod.acc VOLUME CROSS-VALIDATION RMSE (UPLAND)= m3, % NUMBER OF PLOTS (UPLAND)= 532 VOLUME UPLAND MIN AND MAX= 0 and VOLUME UPLAND MEAN= VOLUME STANDARD DEVIATION (UPLAND)= VOLUME BIAS (UPLAND)= Class\Ref >160 User’s acc > Prod.acc Wall-to-wall extension of Forest Inventory and Analysis: K-nearest neighbor estimation and classification 3. Early results concerning accuracy, imagery and data Reija Haapanen, Alan Ek, Marvin Bauer, Kali Sawaya -- Department of Forest Resources, University of Minnesota The new FIA 4- subplot data has some special features due to the close proximity of subplots; in forest cover type classification, the nearest neighbors tend to be in the same cluster. When use of neighboring subplots is prohibited, the accuracies fall to 30% Of the satellite imagery tested (March, April, May and July combined), the first three dates tend to carry enough information for cover type classification Upland-lowland stratification improves forest cover type classification Example of program outputs showing volume accuracies only 696 subplots for our image. The volume accuracies are quite low partly due to the small number of plots. However, by using simple optimization of the field plot locations, the errors drop considerably. Field measurements Large-area statistics & error estimations Preprocessed image & field data combined Preprocessed field data Statistics for small area (e.g., county) Thematic forest maps Image analysis: estimation & classification Preprocessed image data Original satellite image Preprocessing New FIA 4-subplot field cluster design Spectral band 1 of Landsat 7 ETM+ satellite image A georeferenced Landsat 7 ETM+ image, clouds excluded. The black dots represent FIA field plot clusters The Minnesota land use/land cover classification Cover type classification Post-processing ‘kNN’ 2. Flow of the kNN process and materials used The kNN estimation and classification method is a simple, but very powerful way to extend a wide range of field data to landscapes The method assigns a pixel the field data of the most similar pixel, for which actual field plot data exists. The similarity is defined in terms of the feature space (e.g., Euclidean distance between satellite image pixels) kNN retains the full range of variability inherent in the sample Forest/ non-forest stratification Research funding provided by: The USDA Forest Service Forest Inventory and Analysis program (FIA) has been conducting state-by-state and ultimately nationwide forest inventories for decades. Yet these field plot based inventories have not been able to produce precise county and local estimates and useful operational maps. Additionally, traditional satellite-based forest classifications have been unable to match detailed forest type identification with ground based survey definitions to provide for interpolation and extrapolation of FIA data. Precise classification has been limited to general or aggregate classes of little use for improving inventory precision and providing truly useful operational forest maps. The k-nearest neighbors approach (kNN), adapted from the Finnish Forest Research Institute, offers a means of applying satellite and GIS data so as to impute forest cover type, timber volume, and other FIA data from field plots surrounding large or small areas on the basis of the spectral characterization of neighboring pixels. To the extent that such post-stratification can be successfully applied, the method offers agencies and industry a) greater precision at survey unit to local levels of estimation, and b) detailed inventory attributes within type polygons over large areas. In fact, our researchers can map virtually any FIA plot attribute. The method produces estimates and maps according to the actual inventory classifications and definitions rather than an abstract set that must later be reinterpreted. 1. Rationale Photo courtesy of Dave Hansen