Aerial Semantic Mapping for Precision Agriculture using Multispectral Imagery

Abstract

Nowadays constant technological evolution cover several necessities and daily tasks in our society. In particular, drones usage, given its wide vision to capture the terrain surface images, allows to collect large amounts of information with high efficiency, performance and accuracy. This master dissertation’s main purpose is the analysis, classification and respective mapping of different terrain types and characteristics, using multispectral imagery. Solar radiation flow reflected on the surface is captured by the used multispectral camera’s different lenses (RedEdge-M, created by Micasense). Each one of these five lenses is able to capture different colour spectrums (i.e. Blue, Green, Red, Near-Infrared and RedEdge). It is possible to analyse the various spectrum indices from the collected imagery, according to the fusion of different combinations between coloured bands (e.g. NDVI, ENDVI, RDVI. . . ). This project engages a ROS (Robot Operating System) framework development, capable of correcting different captured imagery and, hence, calculating the implemented spectral indices. Several parametrizations of terrain analysis were carried throughout the project, and this information was represented in semantic maps by layers (e.g. vegetation, water, soil, rocks). The obtained experimental results were validated in the scope of several projects incorporated in PDR2020, with success rates between 70% and 90%. This framework can have multiple technical applications, not only in Precision Agriculture, but also in vehicles autonomous navigation and multi-robot cooperation.