Economic Potential of RoboWeedMaps - Use of Deep Learning for Production of Weed Maps and Herbicide Application Maps
1P. Rydahl, 1O. Boejer, 2N. Jensen, 3B. Hartmann, 4R. Jorgensen, 5M. Soerensen, 3P. Andersen, 2L. Paz, 2M. B. Nielsen
1. IPM Consult ApS, Denmark
2. I-GIS A/S, Denmark
3. Datalogisk A/S, Denmark
4. AgroIntelli A/S, Denmark
5. Danfoil A/S, Denmark
In Denmark, a new IPM ‘product chain’ has been constructed, which starts with systematic photographing of fields and ends up with field- or site-specific herbicide application.
A special high-speed camera, mounted on an ATV took sufficiently good pictures of small weed plants, while driving up to 50 km/h. Pictures were uploaded to the RoboWeedMaps online platform, where appointed internal- and external persons with agro-botanical experience executed ‘virtual field inspection’ (VFI) to determine weed species and classes of weed size.
These determinations served 2 purposes:
- input to a 4th IT generation and fully field validated decision support system, named IPMwise, which 1) evaluates needs for control and 2) identifies cost-sorted options for herbicide application, which are returned to the farmer max. 24 hours later
- Training of ‘deep learning’ (DL) to enable a gradual transition to automatic weed discriminations
In 2017-2021, around 400 pictures/ha were taken in 84 Danish fields grown with cereals and maize, and 76 trainings of DL were executed. With logistic planning, around 150 ha/day were photographed. This out-compete photographing of small weed plants by drones. When using a picture size of 5 mB, this ended up in 300 gB/day, which is ‘big data’, requiring special set-ups for both data transmission and -processing.
Currently, the following object in pictures can be discriminated by DL with sufficient agronomical robustness:
- irrelevant objects such as soil, stones, dead plant material, etc.
- crop plants, yet: cereals, maize and oilseed rape
- weed plants, yet: monocots and dicots plus dicots furtherly discriminated as Cirsium arvense
The results from VFIs in the 84 fields were also entered in IPMwise, which returned cost-sorted options for control. By use of the farmer’s planned/executed whole-field treatments as references, a theoretical economic potential of 57-73% reduction of herbicide input was found in different cereal crops, equal to averagely 33 Euro/ha. In maize, additional photographing is required to estimate potentials.
On top of herbicide savings arising from field-specific herbicide application as quantified above, an additional potential arises from a spatially more precise weed control (site-specific control). Until now this has been examined for 1 dicot species, Cirsium arvense.
This species was identified in 13 of the 84 fields, where additionally 88% reduction from site-specific herbicide application was measured. The RWM platform will deliver weed maps, while IPMwise integrated with an online Farm management system (FMS) ‘Naesgaard Mark’, delivers recommendations for whole-field- or site-specific treatments, as preferred by the farmer.
The RWM and IPMwise systems have both demonstrated generic qualities suitable for scaling.
