Fusarium head
blight (FHB) is a serious disease that affects wheat and barley production.
South Dakota alone suffers $20 million in losses annually from FHB. The early
detection of FHB disease will improve the efficiency, accuracy, and capability
of FHB resistance screening in wheat, durum, and barley breeding. Such diseases
can be automatically or semi-automatically diagnosed using supervised machine
learning algorithms and image processing techniques. Several methods have been
proposed to plant disease detection. Among these methods, Convolutional Neural Networks
(CNNs) are an effective method to automatically identify these diseases.
Recently, different approaches have been proposed to boost CNNs’ performance.
Most of these methods, however, suffer from a lack of a mechanism that allows
them to integrate global and local contextual information adaptively to extract
hidden patterns inside the input images. In this regard, we propose the
multi-scale attention U-Net. Our main goal is to improve U-net by introducing
an attention mechanism at the network bottleneck. In the proposed method, to adaptively
emphasize the importance of both spatial and channel dimensions we propose to
include the dual attention mechanism. To this end, using the global information
of each channel we learn the scaling coefficient to improve the object learning
process. In addition, by learning the
self-attention map we impose the spatial attention map on the feature space to
adaptively emphasize the important regions. We apply the suggested dual
attention in a multi-scale fashion to encourage a multi-scale learning process.
To train the proposed network, we captured 12,000 images in an SDSU wheat field
in Volga, SD. In addition to covering healthy and diseased plants, these images
include different stages of the disease. A group of plant disease experts
annotated the prepared images to be used for training the deep model. The model
was trained to separate the diseased areas from the healthy areas by receiving
the input images using the annotated images. For the first time to the best of
our knowledge, we determined the stage of disease based on the segmentation
results of the model. As a result, we could detect the diseased areas of the
plants, as well as determine the disease level of infection. In the test phase,
an automatic plant scanning robot was developed and used to evaluate the
performance of the proposed model. The experimental results demonstrate that
our approach accurately estimates crop contamination.
ACKNOWLEDGEMENT AND DISCLAIMER
Confidential- please do not share the content of the
submission publicly before the 2022 National
FHB Forum. This material is based upon work
supported by the U.S. Department of Agriculture, under Agreement,
59-0206-2-143, 59-0206-2-153, and FY22-SP-002.
This is a cooperative project with the U.S. Wheat & Barley Scab Initiative.
Any opinions, findings, conclusions, or recommendations expressed in this
publication are those of the author(s) and do not necessarily reflect the view
of the U.S. Department of Agriculture.