A method for automatic creation of a vegetation map using high-resolution aerial photographs of unmanned aerial vehicles
Woodland overgrowth caused by the reduced flood disturbance and succession resets are critical technical issues to ensure effective river management in Japan. Woodland overgrowth can increase the risk of flooding and decrease the bio-diversity. Japanese river managers must control the woodland overgrowth and maintain the condition of vegetation. To maintain the woodland in a good condition, Japanese river managers must monitor the state of vegetation dynamics. Currently, Japanese river management uses vegetation maps to monitor the state of vegetation dynamics. Although vegetation maps are effective in monitoring the state of vegetation dynamics, creating vegetation maps will incur high cost. Unmanned aerial vehicles(UAVs) exhibit strong potential to monitor the status of vegetation conditions. As an example of strong potential, we consider the potential of high-resolution aerial photographs. High-resolution aerial photographs provide detailed information about the vegetation growth conditions such as surface solid material and vegetation invasion on the river terrace. The high-resolution aerial photographs aid in understanding the vegetation dynamics in a limited area. If we are attempting to monitor the condition of vegetation dynamics across a large area, analysis of the images will enhance the quantitative monitoring of large spatial heterogeneous areas that are located within the river environment.Based on this background, we conducted image analysis using the high-resolution aerial photographs and applied decision tree analysis to the image analysis results (objects) after which we verified whether it is possible to create an automatic vegetation map. First, we apply object-based segmentation methods, which group together individual image pixels to objects with common characteristics using the image spectrum information (RGB information and brightness) and shapes. Second, we applied the machine learning methods (such as decision tree analysis) to the objects along with cross validation methods.We classified the objects into training data (supervised data) and validation data. After applying machine learning based on the training data, the accuracies of the machine learning methods were evaluated.From the results, we successfully created some approximate automatic vegetation maps. The automatic vegetation map represents the main part (dominating part) of the landscape (water area, gravel bed condition, annual vegetation community) and monitors the investigation and initial succession of the woodlands. The precision of the automatically created vegetation map is approximately 65%, and 35% of the errors are observed to be concentrated among minor objects. These results are produced by minor segmentation, and we hypnotize that reconsidering the initial step will improve the precision.