A Preliminary Study on Tree-Top Detection and Deep Learning Classification Using Drone Image Mosaics of Japanese Mixed Forests
| Autor: | Ferran Roure, Sarah Kentsch, Ha Trang Nguyen, Koma Moritake, Maximo Larry Lopez Caceres, Daniel Serrano, Yago Diez |
|---|---|
| Rok vydání: | 2020 |
| Předmět: |
010504 meteorology & atmospheric sciences
Process (engineering) business.industry Computer science Deep learning 0211 other engineering and technologies Usability 02 engineering and technology TOPS Machine learning computer.software_genre 01 natural sciences Field (computer science) Tree (data structure) Data acquisition Artificial intelligence Cluster analysis business computer 021101 geological & geomatics engineering 0105 earth and related environmental sciences |
| Zdroj: | Lecture Notes in Computer Science ISBN: 9783030661243 ICPRAM (Revised Selected Papers) |
| DOI: | 10.1007/978-3-030-66125-0_5 |
| Popis: | Tree counting and classification tasks in forestry are often addressed by costly, in terms of labour and money, field surveys carried on manually by forestry experts. Consequently, computer vision techniques have been used to automatically detect tree tops and classify them in terms of species or plant health status. The success of the algorithms are highly dependent on the data, and most significantly in its quantity and in the number of challenges it presents. In this work we used Unmanned Aerial Vehicles to acquired extremely challenging data from natural Japanese mixed forests. In a first step, six common clustering algorithms were used for tree top detection. Furthermore, we also assessed the usability of five different deep learning architectures to classify tree tops corresponding to trees in different degrees of affectation from a parasite infestation. Data covering an area of 40 ha are used in extensive experiments resulting in a detection accuracy of over 80% with high location accuracy and up to 90% with lower accuracy. Classification results produced by our algorithms reached error rates as low as 0.096 for classification. Data acquisition and runtime considerations show that this techniques is useful to process real forest data. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|