Weed Mapping Using Very High Resolution Satellite Imagery and Fully Convolutional Neural Network

Abstract

An introduced pasture grass (Andropogon gayanus - Gamba grass) is spreading through the savannah of northern Australia, with detrimental ecosystem consequences that include increased fire severity. In order to monitor the spread and impact of Gamba grass, a scalable solution for mapping this invasive weed over large areas is required. Recent developments in convolutional neural networks designed for semantic segmentation have proven useful for distinguishing vegetation in an automated manner. We construct training data for supervised learning from an airborne LiDAR-derived point cloud using existing techniques and tune the hyper-parameters of a ResUNet-a to produce a viable solution for detecting Gamba grass in very high resolution satellite imagery.