Vegetation Survey using Images obtained during Field Sampling with Digital Cameras

Abstract

Current vegetation surveys are mainly conducted through the visual observation of coverage and species. This method has the disadvantage that it is not quantitative or reproducible. The aim of this study was to develop an objective method to survey vegetation using ground images obtained in the field. The data used in this paper were obtained in August 2012 at the Tottori Sand Dunes. The data were divided into training data and test data. First, we selected training data from our dataset and applied texture analysis. Then, classification models were obtained by using the texture characters as variables. Variables used in the models were selected as follows: 1) variables highly correlated with other variables were excluded based on the Pearson product-moment correlation coefficient, 2) multinomial logistic regression analysis was conducted to obtain the best combination of variables for classification, and 3) the accuracy of the models was estimated by linear discriminant analysis using the best combination of variables. Estimation was performed based on three categories: plant and soil classification, class classification (i.e., distinction between monocotyledon and dicotyledon) and species classification. A comparison of the RGB combined model and the NIR single model was also conducted. In the plant and soil classification, the accuracy of the RGB model and the NIR model were 96% and 87%, respectively. While the soil recall rate of the NIR model was low (lower than 50%), the RGB model showed high rates in all categories. In the class classification, the accuracy of the RGB model and the NIR model were 86% and 74%, respectively. Although the NIR model showed low rates in some categories (i.e., the recall rate of monocotyledons was 28%), the RGB model showed a high rate in all categories (higher than 70%), which was similar to the plant and soil classification. In the species classification, both the RGB model and the NIR model showed low accuracy (RGB: 55%, NIR: 39%). The accuracy differed largely among species (0 - 95%) and many of the misclassifications occurred within the same class. These results demonstrate the potential of coverage estimation and plant classification using this method.