Page 247 - Handbook of Deep Learning in Biomedical Engineering Techniques and Applications

238 Chapter 8 A review on plant diseases recognition through deep learning

6.3 Visualization techniques
Visualization method improves the transparency of the learning
process by allowing the user to visualize what is happening in
the network. Occlusion visualization technique is used on images
to observe the changes in the probability that fall on the appro-
priate class [50]. This paper states that the background of the im-
age may sometimes be considered as the actual image; thus, there
may be some chances to miss the symptoms. Brahimi et al. stated
that the occlusion technique was expensive and takes vast time,
which leads them to use saliency maps based on gradient values
to calculate the importance of pixels in the corresponding node
[51]. Here only the positive gradients are only considered to prop-
agate as the activation functions to acquire more accurate visual-
izations. Sladojevic et al. proposed 13 types of plant diseases by
using CafeNet CNN architecture along with filters, which yielded
96.30% of CA, which is better than some traditional methods [52].
T-distributed stochastic neighbor embedding (t-SNE) is used to
visualize the characteristics of their connected layer to assess
the distance among the defined classes [53]. Mohanty et al. imple-
mented the publically available PlantVillage data set to identify
the plant disease by using AlexNet and GoogLeNet CNN architec-
tures [54]. The efficiency is evaluated employing precision (P),
recall (R), and F1 score. Segmentation and edge maps techniques
are implemented to identify the olive plant disease [55]. Four
types of cucumber diseases were detected in Ref. [56], and the
accuracy was compared with SVMs, Random Forest, and AlexNet
models. Detectors are used in some deep learning models that
predict the percentage of disease in the plant. The detectors are
SSD, RFCN, and Faster-RCNN used along with ZFNet, GoogLeNet,
AlexNet, and VGG [57]. ResNet-50, MobileNet-V1, and Inception-
V2 CNN models along with SSD and Faster-RCNN detectors are
used to detect banana leaf disease [58]. Through heat maps, the
plant disease images are given as input for various combinations
of CNN, and the expected output was the probability of occur-
rence of the disease [59]. Diseases in the soya plant are classified
and detected using LeNet [60]. The wheat plant diseases along
with their features are made visualized using VGG-FCN and
VGG-CNN models [61]. Fusarium wilt in radish is detected with
the VGG-CNN model and K-means clustering method to show
the marks of diseases [62].

6.4 Hyperspectral imaging with deep learning
models

Several imaging techniques are available to detect the plant
disease in an early stage. Some of the available techniques are