Weed detection, classification and mapping using drones.

Many of the safety risks in the natural resource industry are drawn from having staff traversing difficult ground. This risk is exacerbated when these staff have a task to perform such as filling data sheets, that require them to shift their attention from their surroundings. Interest in aerial means of capturing data to negate this risk and other challenges is increasing.

We have developed a number of methods to map weeds with varying degrees of automation in the analysis process. Capturing extremely high resolution data with drones is simple, cheap and easy. The challenge is converting that data into actionable management intelligence.

We utilize multiple spectral bands to help differentiate between landscape structures and teach software the signatures of distinct landscape structures, in a process to automate weed identification over large areas. This process of differentiation is straight forward for those structures that are clearly different, blackberry in grasslands next to a water body for example (See Figure #1). The difficulty lies in structures of similar spectral reflectance such as two different species of Eucalyptus (Figure #2). From the air, discerning differences in species that look similar with confidence, becomes very difficult.

Figure #1: Regions of interest polygons for spectral classification.

Figure #2: Regions of interest polygons for spectral classification over Eucalyptus.

A solution to this challenge is to review data manually. Someone with experience in the ecological vegetation can identify target species. For large projects, this may be time consuming. However, relative to current ground-based techniques, it is a safer and time effective process.

Whilst the current semi-automated classification methods struggle with mis-identification, the addition of circularity coefficients and object detection templates to spectral signature recognition software, may improve the accuracy significantly. However, current methods are sufficient for efficient, spatial mapping of vegetation in select circumstances.

If you have had experience mapping weeds or target vegetation species with drones, please let us know. We would enjoy hearing your story and sharing what we have learnt in an attempt to raise the quality of service in this space.