Epsilon Archive for Student Projects

Abstract

A lot of research has focused on which laser metrics and which k-Nearest Neighbour (kNN) distances give the most accurate estimations. Most studies suggest that the kNN distance k-Most Similar Neighbour (kMSN) is the most accurate for estimating forest variables from local training data. However, little research has focused on how to best acquire training data for estimating forest variables. The aim of this study was (1) to evaluate if a faster and simpler inventory method can be used to estimate a similar accuracy for diameter distribution as unbiased sample plot inventory, by using Airborne Laser Scanning (ALS) data and kMSN imputation. (2) To reduce the sample size in order to find the threshold for minimum training data size without reducing the accuracy. Three different sampling methods were compared; circular plot inventory Sample 1 (S1) and two transect inventories Sample 2 and Sample 3 (S2 and S3). S1 use plots with 5 m radius. Both S2 and S3 were located between the S1 circular plots, using approximately a 2 m sampling width. For S2 a transect start between two circular plots i and i+1, and ends between two circular plots i+1 and i+2. For S3 a new transect was created approximately every 10th tree, which on average created shorter transects. The field data from the three inventory methods was linked with ALS data to create three sets of reference data, with ALS data being extracted from the centre of each plot and transect. The reference data was then used for a kMSN imputation on a set of validation plots. Twenty five plots with 40 m radius were used to validate the kMSN estimates and all trees on the validation plots were callipered. The three sets of imputed trees were compared to the measured variables (ground truth) on the validation plots to assess the accuracy. The imputed mean diameter had a relative Root Mean Square Error (RMSE) of 15.4%, 16.4%, and 17.4% for S1, S2, and S3 respectively. The absolute error index means were 84.4, 91.9, and 82.2 for S1, S2, and S3 respectively. The relative error index means were 0.38, 0.33, and 0.33 for S1, S2, and S3 respectively. The results showed that training data from transect inventory can be used to estimate diameter distributions (both absolute and relative) with similar accuracy as training data from circular plot inventory. The transect stem density need to be measured with higher accuracy to get reliable estimates of forest variables when using training data from transect inventory. The results also showed that a fairly small set of training data (100-150 plots) can be used without reducing the accuracy much. An attempt at imputing stands was made but the estimates were not very accurate. K was set to one and the training data consisted of 90 micro stands, it would be preferable to use a higher k and a larger set of training data.