Browsing by Subject "neulaskato"

The purpose of this study was to investigate the usability of aerial images and Landsat TM in estimating Scots pine defoliation. Estimation methods tested were unsupervised classification, maximum likelihood method, mixed model and linear regression model. Image features for needle loss detection were selected with stepwise linear regression and mixed model technique. As a part of the study the relationship between needle loss and leaf area index (LAI) was examined. The relationship between image features, needle loss and leaf area index was also examined. Numerical aerial images and Landsat TM satellite images were used. Textural features were calculated from aerial images and spectral vegetation indices from the satellite image. The study site was located in Ilomantsi, Finland. 71 field sample plots were measured and located with GPS. Field plots were circular plots. Trees with diameter brest height (dbh) over 13,9 cm were measured from 13 meter radius and trees with dbh 5,0 - 13,8 cm were measured from 7 meter radius. Needle loss of all pines was estimated. Needle loss for the plot was calculated as an average weighted by tree height. Four different class combinations were tested in classification. Plots were divided in 2, 3, 4 and 9 classes depending on their needle loss. Different image feature combinations and classification methods were tested. Classification was done by cross validation. Classification results were compared with original classes. The reliability of the classification was tested using accuracy matrix and kappa value. A mixed model was also used for aerial image features. The best image feature combination with all classification methods was the aerial image feature combination selected with stepwise selection method. Both spectral and textural features were included in the stepwise selection result. Classification accuracy varied between 38,0 % (9 classes) and 88,7 % (2 classes). The best explanatory variable for needle loss was the aerial image NIR channel maximum radiation (r2=0,69). However, unsupervised and supervised classification might have produced too positive results because of correlation in the data. Mixed model technique was used to select the variables for the linear model. Mixed model was used to reduce the effects of the correlation. The model classification accuracy varied between 35,2 % (9classes) and 87,3 % (2classes). According to mixed model selection result no textural features were significant predictors for needle loss. Classification results with Landsat image features were slightly poorer than with the best aerial image feature set (accuracy between 25,4 % and 88,7 %). The relationship between needle loss and LAI was poor (r2=0,27). Needle loss and LAI also correlated with different image features. LAI correlated slightly better with textural features than needle loss. Spectral vegetation indices calculated from Landsat TM correlated moderately with both needle loss and LAI. Indices VI3 (r2=0,56), MIR/NIR (r2=0,51) and RSR (r2=0,44) had the strongest connection to needle loss. Spectral vegetation indices could be a potential way for large area needle loss detection.