More often than not, when an old house in Perth is demolished in winter, a few weeks later a spectacular germination of blue lupins (aka sandplain lupins) will emerge from where the house once stood. These weeds have a longer seed bank life than Keith Richards has birthdays! This is not a weed that we can beat by depleting the seedbank, Peter Newman* writes.
Blue lupins are very closely related to the domesticated narrow-leaf lupin (or white lupin), so whichever herbicide is survived by white lupins is also survived by blue lupins.
There are no herbicide options for the control of blue lupins in white lupins, and blue lupins are incredibly competitive where just a few plants per square metre can be enough to warrant re-seeding the crop.
A team of researchers from AHRI and the University of Western Australia, led by Monica Danilevicz have recently taken the approach of using weed labelling and deep learning algorithms to see if they can distinguish between the two lupin species in images captured by UAV or ground-based cameras. The aim is to detect and spray just the blue lupins in crop.
Both blue lupins and white lupins look similar to one another with palmate leaves, but the blue lupin has “fatter” leaves. This is not like trying to detect wild radish in wheat, we are talking about differentiating very similar plants from one another.
They had a win! The researchers found that they could find 80% of the blue lupins in a crop of white lupins, with blue lupins occupying just 2% of the field. They found the best results when high-resolution images were captured early in the crop’s development stages.
This research is not for the faint-hearted. It involved sitting for hours drawing boxes around blue lupins in over 9000 images. Perhaps if they label another 9000 images, they’ll get to 90% accuracy!
Narrow-leaf lupins (Lupinus angustifolius) are an important legume crop, but are a very small crop globally, hence the development of new herbicide options is limited. They are well suited to acidic, sandplain soils that many farmers in the world would liken to farming on the beach.
Blue lupins (Lupinus cosentinii) were established on sandplain soils in Western Australia in the early 1900’s as they are a legume that grows well on these soils. Their seed shatters at maturity and the grain has high alkaloid levels. They are also known as bitter lupins.
The blue lupin farming system involved grazing blue lupins with sheep, where the sheep mostly ate weeds and avoided the blue lupins due to their bitter taste. The seed shattered and provided sheep feed over summer, and a wheat crop often followed in the following winter.
The success of the growth of this valuable legume on sandy soils led to the domestication of the lupin to develop the narrow leaf lupin which is now a successful crop on these sandy soils.
_medium.jpg)
Blue lupins are very competitive plants that outgrow their domesticated relatives. They mostly emerge with opening winter rains; thus the main control measure is to wait for a germination of blue lupins which are knocked down prior to seeding. This leads to yield penalties as a result of this delayed seeding.
The herbicide tolerance of white lupins is almost identical to that of blue lupins so there are no selective herbicide options.
Therefore, the main future options for control of blue lupins appear to be to breed a white lupin with tolerance to a herbicide such as the Group 2 herbicides, and/or develop an algorithm that will enable growers to detect and spray just the blue lupins with non-selective herbicides. This would clearly leave dead patches in the crop, but if undertaken early in the crop’s development, the white lupin crop would be able to fill these spaces.
The researchers found that they could detect 80.3% of blue lupins using the algorithm they developed. This accuracy dropped off for images captured at a 20m height due to the reduced image resolution. This is a great result, and the accuracy of the algorithm may likely be improved by feeding more labelled images into the deep learning model.
Previous studies have shown 90% accuracy of detecting wild radish in canola, and another study found 93% accuracy detecting Italian ryegrass in wheat.
In these images, the blue lupins only occupied 2% of the ground area. This is considered to be a low infestation.
._photo_by_grdc_medium.jpg)
Both lupin species present heliotropism, which tilts the leaves toward the sun to increase solar absorption. The study showed that heliotropism may make it easier to identify sandplain lupins. This study aimed to collect images close to noon each time.
We are on the verge of very exciting developments in weed control using AI and deep learning models to help us detect and spray weeds in crops. This will reduce herbicide use, give us options to control herbicide-resistant weeds, as well as potentially lead to the development of weed control solutions where currently there are none, such as controlling blue lupins in white lupins.
This important research brings us another step closer to this technology becoming mainstream. We won’t beat blue lupins by targeting the seed bank; investing in technology such as plant breeding or weed detection is what we need to solve this problem, and we are well on the way with both.
*Peter Newman is western extension agronomist with WeedSmart.
