Import of kNN: Difference between revisions
Jump to navigation
Jump to search
No edit summary |
No edit summary |
||
| (One intermediate revision by one other user not shown) | |||
| Line 1: | Line 1: | ||
[[Category:User's Guides|PlanStart|PlanStart]] | |||
[[Category:PlanStart]] | |||
[[Category:Database]] | |||
'''The ''k''NN-Sweden data''' consist of pixelwise estimates of forest states. A complete coverage is possible by using remotely sensed information, usually satellite image data. Sensors onboard, e.g., a SPOT satellite obtain spectral signatures in different wavelength bands (usually IR, R, G, and B) in each pixel, with a ground resolution approx. corresponding to 25 by 25 meters. Geo-referenced, previously inventoried NFI-plots will serve as reference data. In a feature space, defined by the satellite image, the ''k'' Nearest Neighbours (usually with ''k'' = 10) are found among the reference plots. Estimations of volume, age, tree species composition, etc. are then derived as weighted mean values (e.g. weighted inversely proportional to the distance in the feature space). | '''The ''k''NN-Sweden data''' consist of pixelwise estimates of forest states. A complete coverage is possible by using remotely sensed information, usually satellite image data. Sensors onboard, e.g., a SPOT satellite obtain spectral signatures in different wavelength bands (usually IR, R, G, and B) in each pixel, with a ground resolution approx. corresponding to 25 by 25 meters. Geo-referenced, previously inventoried NFI-plots will serve as reference data. In a feature space, defined by the satellite image, the ''k'' Nearest Neighbours (usually with ''k'' = 10) are found among the reference plots. Estimations of volume, age, tree species composition, etc. are then derived as weighted mean values (e.g. weighted inversely proportional to the distance in the feature space). | ||
However, such mean values are not sufficient for the long-term prognoses and the advanced forest management simulations to be analysed by the system. Single tree-data is here needed, preferably as a sample of plots with trees (or seedlings or saplings). Instead of using the pixelwise mean values, the nearest neighbour is imputed, i.e. ''k'' = 1. This specific reference plot is identified in the NFI database (by "�?R-TRAKT-P�?LSLAG") where the actual inventory could have been performed years ago (hence, an outlier-reduction of the reference database is first done, excluding plots with assumed major changes in forest state). A suitable reference database is firstly defined by the same region- and county-code as the analysis area, possibly also a range in altitude (e.g. 100 - 300 meters a.s.l.). Moreover, instead of imputing a plot to each pixel of the current stand, the imputation could be performed in a sample of corresponding pixels (randomly or systematically sampled). A sample of, e.g., 10 pixels will imply the imputation of 10 plots per stand. | However, such mean values are not sufficient for the long-term prognoses and the advanced forest management simulations to be analysed by the system. Single tree-data is here needed, preferably as a sample of plots with trees (or seedlings or saplings). Instead of using the pixelwise mean values, the nearest neighbour is imputed, i.e. ''k'' = 1. This specific reference plot is identified in the NFI database (by "�?R-TRAKT-P�?LSLAG") where the actual inventory could have been performed years ago (hence, an outlier-reduction of the reference database is first done, excluding plots with assumed major changes in forest state). A suitable reference database is firstly defined by the same region- and county-code as the analysis area, possibly also a range in altitude (e.g. 100 - 300 meters a.s.l.). Moreover, instead of imputing a plot to each pixel of the current stand, the imputation could be performed in a sample of corresponding pixels (randomly or systematically sampled). A sample of, e.g., 10 pixels will imply the imputation of 10 plots per stand. | ||
By ''k''NN imputation the forest state of a stand is described by sample plots, seemingly field inventoried in the current stand. Some things give the hoax away though; the within-stand variation regarding volume, age, etc. might be either | By ''k''NN imputation the forest state of a stand is described by sample plots, seemingly field inventoried in the current stand. Some things give the hoax away though; the within-stand variation regarding volume, age, etc. might be either unrealistically low (if the imputation is done "with replacement", i.e. one reference plot can represent a stand more than once) or unrealistically high (e.g., Maturity class K1 and S1 should not occur in the same stand, such stand should in real life be split into two stands). | ||