Professional Documents
Culture Documents
Shin Et Al - 2021 - A Deep Learning Approach For RGB Image-Based Powdery Mildew Disease Detection
Shin Et Al - 2021 - A Deep Learning Approach For RGB Image-Based Powdery Mildew Disease Detection
Original papers
A R T I C L E I N F O A B S T R A C T
Keywords: In this study, Deep Learning (DL) was used to detect powdery mildew (PM), persistent fungal disease in
Disease detection strawberries to reduce the amount of unnecessary fungicide use, and the need for field scouts. This study opti
Powdery mildew mised and evaluated several well-established learners, including AlexNet, SqueezeNet, GoogLeNet, ResNet-50,
Deep learning
SqueezeNet-MOD1, and SqueezeNet-MOD2. Data augmentation was carried out from among 1450 healthy and
Convolutional neural network
Artificial intelligence
infected leaf images to prevent overfitting and to consider the various shapes and direction of the leaves in the
field. A total of eight clockwise rotations (0◦ ; the original data, 45◦ , 90◦ , 135◦ , 180◦ , 225◦ , 270◦ , and 315◦ ) was
performed to generate 11,600 data points. Overall, the six DL algorithms that were used in this study showed on
average of >92% in classification accuracy (CA). ResNet-50 gave the highest CA of 98.11% in classifying the
healthy and infected leaves; however, considering the computation time, AlexNet had the fastest processing time,
at 40.73 s, to process 2320 images with a CA of 95.59%.When considering the memory requirements for
hardware deployment, SqueezeNet-MOD2 would be recommended for PM detection on strawberry leaves with a
CA of 92.61%.
* Corresponding author at: Department of Engineering, Faculty of Agriculture, Dalhousie University, Truro, NS B2N5E3, Canada.
E-mail address: YoungChang@Dal.Ca (Y.K. Chang).
https://doi.org/10.1016/j.compag.2021.106042
Received 17 June 2020; Received in revised form 11 October 2020; Accepted 2 February 2021
Available online 27 February 2021
0168-1699/© 2021 Elsevier B.V. All rights reserved.
J. Shin et al. Computers and Electronics in Agriculture 183 (2021) 106042
major component of the third wave of the modern agricultural revolu detecting disease in tomatoes and yielded similar accuracies (Durmuş
tion (Robert, 2002). PA was initially used for the targeted allocation of et al., 2017); however, for mobile applications with light memory re
fertilisers to suit different soil conditions (Rosa et al., 2000). Since then, quirements, SqueezeNet was found to be more appropriate. GoogLeNet
PA has been developed for the automated guidance of agricultural ve (ILSVRC 2014 winner), also known as Inception, has also been used for
hicles and tools, autonomous machines and processes, research on disease detection in cassava. For example, Ramcharan et al. (2017)
farms, and the automated management of agricultural production sys determined that Inception V3 was more effective than previous versions,
tems (Zhang et al., 2002). Collecting data using sensors that are and it was aligned with detecting diseases in cassava (Szegedy et al.,
mounted on machines (i.e., unmanned ground vehicles [UGV], un 2016). However, they reported some limitations of Inception related to
manned aerial vehicles [UAV], satellites, and airplanes) is non- memory requirements and computational complexity. In studies
destructive and applicable over large geographical areas. The analysis comparing AlexNet with GoogLeNet, the differences were variable; for
of ground- and aerial-based imagery are often critical components in PA example, Mohanty et al. (2016) showed that GoogLeNet was more
(Liaghat and Balasundram, 2010), especially given the importance of effective than AlexNet in detecting an array of 26 diseases found on 14
image identification and classification techniques in the practice. different crop species whereas the accuracy rates for the two algorithms
Machine learning (ML), which includes deep learning (DL), can be were similar when used to detect nine diseases common to tomatoes.
used to detect pathogens, pests, nutrient deficiencies, and other abiotic Lastly, Fuentes et al. (2017) compared different versions of the ResNet
stressors (Teke et al., 2013). In the case of ML without the use of DL, algorithm (ILSVRC 2015 winner) for detecting tomato diseases, where it
hereinafter referred to as the “non-DL”, several feature extraction was determined that ResNet-50 outperformed ResNeXt-50 (ILSVRC
methods (e.g., histogram of oriented gradients [HOG], speeded up 2016 winner).
robust features [SURF], and gray level co-occurrence matrix [GLCM]) Despite the application and success of different CNNs for detecting
have been evaluated and compared (Shin et al., 2020). The best crop diseases, there is no existing study that demonstrates the applica
extraction method can subsequently be used as an input for the learning bility for detecting PM on strawberry leaves. Developing a DL technique
algorithm. In comparison to non-DL approaches, DL is more effective in that could rapidly assess PM status in strawberry fields would help
handling additional complexity and hierarchical structure in the data inform producers when targeted applications of fungicides are required
(Qi et al., 2017). The key aspect of DL is that it identifies and extracts the and reduce disease-scouting, labour requirements. Hence, the objectives
best features during the learning procedure (LeCun et al., 2015). In the of this paper are to differentiate between PM infected and uninfected
case of DL, the processing step of DL is more simplified than non-DL strawberry leaves by (1) comparing the performance between super
because the optimal feature method is applied to the learning algo vised non-DLs and CNNs; (2) modifying the architectures of four CNNs
rithm automatically (Alom et al., 2018). Furthermore, testing time of the (AlexNet, SqueezeNet, GoogLeNet, and ResNet-50) and developing
DL, which is processed with the data first seen in the algorithm, is SqueezeNet-MOD1 and SqueezeNet-MOD2; and (3) evaluating their
typically faster than that of non-DL approaches (Chen et al., 2014). Fast performance with respect to CA, computation time, and memory
processing time is critical for developing algorithms that are going to be requirements.
applied to hardware (e.g., field-programmable gate array [FPGA] or
mobile application) and it could provide producers with results in a 2. Materials and methods
minimal amount of time to facilitate rapid management decisions to be
taken. 2.1. Image datasets
DL can be categorised into three types according to the method and
purpose of the study. The first type is unsupervised learning, which does Healthy and infected strawberry leaves were collected from Bala
not require data labels, and is optimal for identifying representative more Farm Ltd. (45◦ 24′ 35.4′′ N, 63◦ 34′ 26.3′′ W) and Millen Farms Ltd.
features with little data. Two examples of unsupervised learning are (45◦ 23′ 57.6′′ N, 63◦ 33′ 31.1′′ W) in the summers of 2018 and 2019 under
deep belief network and deep autoencoder (Vincent et al., 2010). The the supervision of disease specialists. The varieties of strawberry were
second type is a recurrent neural network, which fits well with Ruby June and Albion. Over the plant growth stages, the formation of
sequential data processing, such as protein or polymer sequences. The asexual conidia occurs when petioles develop, and mycelium begins to
last type is a convolutional neural network (CNN), which has generated spread. Mycelium first appears on leaves and petioles and iteventually
excellent results in image recognition and natural language processing spreads to the fruits. Data consisted of images acquired during the post-
(Deng and Platt, 2014). Among the three types of DL, CNN has been used germination phase and prior to the fruit-bearing stage. After collecting
in various disciplines, along with advanced computer vision (CV), and is the leaves, they were placed in coolers with icepacks in the field and
the most popular for detecting crop diseases due to the high classifica being preserved in an icebox while transportation. Images were taken
tion accuracy (CA) in image recognition (Szegedy et al., 2015, 2016). right after arriving in the lab with a person holding a digital single-lens
The ImageNet Large Scale Visual Recognition Challenge (ILSVRC) is reflex (DSLR) camera: EOS 1300D (Canon Inc., Tokyo, Japan), with
an annual competition aimed at fostering the development of object 3456 × 5184 pixels (Raw CR2 format). The camera had a 1/15 exposure
detection and image classification algorithm using the annotated time with a shutter priority option and ISO 100 with a 46 mm focal
ImageNet dataset (Deng et al., 2009). Outcomes from the ILSVRC length under 5,500 K colour temperature illumination—mimicking
include milestone algorithms and techniques in the field of CV and DL. direct sunlight. The total number of images collected from 2018 to 2019
In this study, four established, state-of-the-art CNNs (AlexNet, Squee consisted of 677 healthy and 773 infected leaves for a total of 1450
zeNet, GoogLeNet, and ResNet-50), along with two developed by the leaves.
authors, SqueezeNet-MOD1 and SqueezeNet-MOD2, were evaluated and The computations were implemented in MATLAB R2019a (The
compared for PM detection on strawberry leaves. MathWorks Inc., Natick, MA, U.S.A.) using an Intel® Core™ i7-8700
The four selected CNNs, AlexNet, SqueezeNet, GoogLeNet, and CPU @ 3.20 GHz with 48.0 GB RAM and 64-bit Windows 10 oper
ResNet-50, were found to be commonly used in agriculture application ating system. All images were first cropped to 908 × 908 pixels, then
(Kamilaris and Prenafeta-Boldú, 2018; Jiang et al., 2019). For example, resized to 227 × 227 pixels or 224 × 224 pixels depending on the
AlexNet (ILSVRC 2012 winner) and SqueezeNet were compared for requirement of the architecture to ensure that the relative size of the
2
J. Shin et al. Computers and Electronics in Agriculture 183 (2021) 106042
leaves on all images was consistent and to reduce the computational Table 1
demand. The required input size in AlexNet, SqueezeNet, SqueezeNet- Depth, size, parameters, and image input size per each deep learning
MOD1, and SqueezeNet-MOD2 was 227 × 227 pixels whereas GoogLe architectures.
Net and ResNet-50 required an input size of 224 × 224 pixels. Images Network Depth Size Parameters (millions) Image input size
were adjusted to ensure that the relative portions of PM were uniform AlexNet 8 227 MB 61.0 227-by-227
across all images, to reduce the computational complexity, and to SqueezeNet 18 4.6 MB 1.2 227-by-227
improve the efficiency of the processing in terms of computational time. GoogLeNet 22 27 MB 7.0 224-by-224
Data augmentation in DL is essential for models to be generalised so ResNet-50 50 96 MB 25.6 224-by-224
that it could be applied more easily to real field situations (Taylor and
Nitschke, 2017). In addition, Rasti et al. (2019) also identified that at
The learning rate is one of the most important hyperparameters to
least 10,000 observations are required to optimise the overall perfor
tune when training CNNs as it tells the optimiser how much weight to
mance of DL algorithms and minimise issues related to model over
move in the opposite direction from the mini-batch since DL models are
fitting. Cawley and Talbot (2010) demonstrated that the main
primarily trained by stochastic gradient descent optimisers. Adjusting
prevention of overfitting is derived from the model selection. They
the learning rate can modify the behaviour the model and yield different
tested the overfitting problem by using leave-one-out cross-validation
results. For example, if the learning rate is low, the optimisation time is
and investigated the effects of overfitting in model selection. Hence,
increased; however, the training process is able to be stabilised.
data augmentation was performed on the whole dataset to improve the
Conversely, if the learning rate is high, the optimisation time is
robustness of the architectures and increase the number of observations
decreased; however, the model may not converge on an optimal solution
(Liu et al., 2018). There are two aspects of data augmentation:
and in some cases, it may diverge (Sutskever et al., 2013). The number of
geometrical transformation (e.g., rotating, flipping, cropping, and
epochs is also a crucial hyperparameter that defines the number of times
resizing the images; Fadaee et al., 2017) and intensity transformation (e.
that the learning algorithm will process through the entire training
g., noise, colour change, brightness enhancement; Fuentes et al., 2017).
dataset. An epoch can be comprised of one or more batches and a mini-
Among them, a rotation technique was applied in order to represent the
batch can be determined according to the size of data. Usually, a batch
different shapes and directions of the leaves in the field. Hence, data
size of 32 is a good starting point while batches of 62, 128, and 256 can
augmentation was carried out using a clockwise rotation along the
be applied as well (Keskar et al., 2016); however, identifying the opti
0◦ (the original images), 45◦ , 90◦ , 135◦ , 180◦ , 225◦ , 270◦ , and 315◦
mised batch size within this given range is generally suggested.
angles. Therefore, the original 1450 observations were increased by
eight times with 5416 observations representing healthy leaves and
2.3.1. AlexNet
6184 observations representing infected leaves for a total of 11,600
AlexNet consists of five convolution layers, three fully connected
observations.
layers, and the softmax function. In the first convolutional layer, 96
kernels (i.e., filters) with a size of 11 × 11 × 3 can be calculated with a
2.2. Comparisons of non-DLs and CNNs stride of four on 227 × 227 input images. Among the five convolution
layers, the first and second convolution layers are followed by the
The performances with respect to CA in non-DL and CNN techniques overlapping max pooling; however, the remaining convolutional layers
were compared with the original dataset of 1450 images. In our recent are connected directly to each other. The last convolutional layers are
study, Shin et al. (2020) compared the classification accuracies between followed by overlapping max pooling before connecting to the two fully
three feature extraction techniques (HOG, SURF, and GLCM) and two connected layers. The second fully connected layers are linked to the
supervised non-DLs (artificial neural network [ANN] and support vector softmax classifiers with two class labels. The architecture of AlexNet in
machines [SVM]). Results of their study showed that the combination of this study was changed from 1 × 1 × 1000 into 1 × 1 × 2 in the last,
SURF and ANN, and GLCM and SVM were optimal to detect PM on fully-connected layer that performs the binary classification (i.e.,
strawberry leaves (Shin et al., 2020). healthy and infected; Fig. 1).
3
J. Shin et al. Computers and Electronics in Agriculture 183 (2021) 106042
Input
Conv [a] MXP [b] Conv 5×5, Conv 3×3,
11×11, pad=2 MXP 3×3, pad=1
3×3,
stride=4 Layer2: stride=2 Layer3:
stride=2
Layer1: 96 256 384
227×227
Input
[a] 1×1, 64 1×1, 128
Conv 7×7, MXP[b] 3×3, 64 3×3, 128
64, stride 2 3×3, 1×1, 256 1×1, 512
stride 2
Conv 1 Conv 2×3 Conv 3×4
224×224
4
J. Shin et al. Computers and Electronics in Agriculture 183 (2021) 106042
Fire 2
[b] [b]
Fire 2
[b] Bypassing
Fire 2
5
J. Shin et al. Computers and Electronics in Agriculture 183 (2021) 106042
Table 3
Performance of six CNN algorithms using the augmented dataset (11,600 images).
AlexNet(%) SqueezeNet (%) GoogLeNet (%) ResNet-50 (%) SqueezeNet-MOD1 (%) SqueezeNet-MOD2 (%)
CA 95.59 ± 0.49[a] b[b] 95.80 ± 0.72b 96.36 ± 1.16b 98.11 ± 0.28 a 96.38 ± 0.57b 92.61 ± 0.84c
Precision 96.14 ± 1.93 96.90 ± 1.20 96.94 ± 2.12 98.46 ± 0.37 97.00 ± 1.21 92.96 ± 1.21
Sensitivity/recall 95.64 ± 1.83 95.29 ± 1.64 96.31 ± 2.64 97.99 ± 0.42 95.84 ± 1.77 93.48 ± 2.49
Specificity 95.54 ± 2.34 96.49 ± 1.41 96.74 ± 2.61 98.25 ± 0.37 96.30 ± 2.30 92.52 ± 1.96
F1-score 95.86 ± 0.44 96.07 ± 0.73 96.56 ± 1.18 98.23 ± 0.26 96.29 ± 0.26 94.10 ± 1.18
6
J. Shin et al. Computers and Electronics in Agriculture 183 (2021) 106042
Table 5 Table 6
Computation times of six CNN algorithms using the augmented dataset (11,600 Analysis of variance of computation time of six CNN algorithms using the
images). Training time was calculated with 9280 images and testing time with augmented dataset (11,600 images).
2320 images. Source DF[a] SS[b] MS[c] F-value p-value
Algorithms Training time (s)[a] Testing time (s)[b]
Algorithms 5 62,771 12554.2 127.64 <0.001
AlexNet 3451.88 ± 40.22[c] 40.73 ± 0.33c[d] Error 24 2361 98.4
SqueezeNet 3555.40 ± 261.15 73.40 ± 3.81b Total 29 65,131
GoogLeNet 9365.25 ± 41.61 87.13 ± 2.38b [a]
Degrees of freedom.
ResNet-50 21738.29 ± 420.16 178.20 ± 8.51 a [b]
SqueezeNet-MOD1 5141.80 ± 27.31 68.55 ± 3.99b Sum of squares.
[c]
SqueezeNet-MOD2 3832.87 ± 95.46 45.70 ± 0.85c Mean square.
* The bold font shows the short computation time among the values which has
no significant difference.
[a]
Training time (seconds) for processing with 9280 data. demands were further reduced. Both modifications of SqueezeNet had
[b]
Testing time (seconds) for processing with 2320 data. the smallest memory requirements in comparison to the other CNN al
[c]
The standard deviation of five repetitions. gorithms presented in this study. Here, SqueezeNet-MOD1 could be
[d]
Means that do not share a letter are significantly different. selected when high CA is required and SqueezeNet-MOD2 could be
selected for when training and testing times need to be reduced and/or
when memory requirements need to be minimised.
It was determined to be a significant difference (p < .001) in CA
between the CNN algorithms (Table 4); therefore, an MMC using 4. Conclusions
Tukey’s test was carried out. Here, it was determined that there were no
significant differences among all the different algorithms with the This study began with a comparison of non-DLs and CNNs with an
exception of ResNet-50. ResNet-50 showed a significantly higher CA of image dataset consisting of PM infected and uninfected strawberry
(98.11%), which aligns with our expectations based on similar results leaves (1450 images). As we expected, CNNs performed better than non-
reported by Bianco et al. (2018) (Table 3). However, the magnitude of DLs when distinguishing between infected and uninfected leaves and
the difference was small. when the dataset was increased, and parameters to the function were
added. The overall goal of our study was to develop a deployable
hardware system to detect PM for a strawberry field to provide recom
3.3. Computation time mendations for fungicide applications. Our testing was expanded to use
an image-rotated, data augmentation technique, leading to a total of
Computation time is also a crucial parameter that needs to be 11,600 data points. ResNet-50 had statistically higher CA (98.11%) but
considered in order to facilitate real-time processing. Even though the the CA among the rest of the algorithms was not significantly different.
popularity of DL has increased, and computation ability has been When considering potential hardware memory requirements,
improving, the amount of time required to train the algorithms is a SqueezeNet-MOD2 had the lowest requirements. In terms of testing
critical consideration (Justus et al., 2018). The training and testing times time, ResNet-50 required the most time with 178.20 s and AlexNet the
shown in Table 5 were calculated as the training time per epoch and the least time with 40.73 s to process 2320 images. The testing time was
number of epochs that should be performed to reach the desired level of significantly different among the CNN algorithms, with AlexNet and
accuracy. Training time was the time required to train the algorithms SqueezeNet-MOD2 having the shortest times.
using 9280 images (i.e., 80% of the total augmented dataset) and test The experimental results showed that the CNN techniques are a
time was the time required to test the algorithm using 2330 images (i.e., promising tool for development of a field deployable strategy to detect
20% of the total augmented dataset). PM in the strawberry leaves. The ideal management scenario would be
In Section 3.2., ResNet-50 was recommended based on having the to prevent the disease at early onset; however, if the imagery is acquired
highest CA; however, it also had the longest computation time compared during later stages of the diseases, preventative practices may be unre
to all other algorithms (Table 5). Significant differences in testing time alistic, and the disease may need to be treated with chemical applica
(p < .001) between the CNN algorithms were observed (Table 6). tions. Using the proposed DL approach, it would be possible to install
AlexNet and SqueezeNet-MOD2 had the shortest computation time these algorithms on agrochemical spraying hardware platform—an area
(Table 5); however, the optimal model should also consider CA and of future research. Although this research was carried out in a laboratory
memory requirements. When considering computation time, AlexNet setting where the lighting conditions and leaf orientation were kept
performed the fastest with an average of 3451.88 s when training the constant, we fully recognise that the proposed techniques warrants
algorithm and an average of 40.73 s when testing the algorithm future research with regards to their transferability to field conditions,
(Table 5). Both training time and testing time in AlexNet excelled rela where there are additional challenges related to irregular lighting con
tive to all DL algorithms in our study. ditions, leaf orientation, and overlapping of leaves.
Based on computational time, the findings from this study were Specifically, study could be extended to develop a fully automated
consistent with the findings of Bianco et al. (2018). Shafiee et al. (2017) hardware (i.e., FPGA or mobile application) in order to help producers
indicated that SqueezeNet is good for deployment on mobile hardware who are struggling with PM disease. Future work will investigate the
applications (Table 1). The results of the processing time with ResNet-50 integration of the CNN algorithms with hardware and develop a disease
are consistent with our expectation that it would require the longest management platform, that would be easy for producers to use by
computational time due to the model complexity. In general, DL might providing accurate and fast results.
require much longer training times compared to non-DL; however, this
is potentially offset by the shorter testing time required (Busseti et al., Funding
2012).
Overall, based on CA, computation time, and memory requirements This research was funded by Natural Science and Engineering
for a hardware implementation, SqueezeNet would be recommended. Research Council of Canada (NSERC) Discovery Grants Program
When considering memory (Table 1), SqueezeNet required the least (RGPIN-2017-05815).
memory and by removing one or three of the Fire modules for
SqueezeNet-MOD1 and SqueezeNet-MOD2, respectively, the memory
7
J. Shin et al. Computers and Electronics in Agriculture 183 (2021) 106042
CRediT authorship contribution statement Iandola, F.N., Han, S., Moskewicz, M.W., Ashraf, K., Dally, W.J., Keutzer, K., 2016.
SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and< 0.5 MB model
size. arXiv preprint arXiv:1602.07360.
Jaemyung Shin: Conceptualization, Methodology, Data curation, Jacob, D., David, D.R., Sztjenberg, A., Elad, Y., 2008. Conditions for development of
Writing - original draft, Investigation, Software. Young K. Chang: powdery mildew of tomato caused by Oidium neolycopersici. Phytopathology 98
Conceptualization, Methodology, Data curation, Writing - review & (3), 270–281.
Jiang, B., He, J., Yang, S., Fu, H., Li, T., Song, H., He, D., 2019. Fusion of machine vision
editing, Software, Resources, Supervision. Brandon Heung: Conceptu technology and AlexNet-CNNs deep learning network for the detection of
alization, Writing - review & editing. Tri Nguyen-Quang: Writing - postharvest apple pesticide residues. Artif. Intell. Agric. 1, 1–8.
review & editing, Supervision. Gordon W. Price: Writing - review & Justus, D., Brennan, J., Bonner, S., & McGough, A.S., 2018, December. Predicting the
Computational Cost of Deep Learning Models. In 2018 IEEE International Conference
editing, Supervision. Ahmad Al-Mallahi: Writing - review & editing, on Big Data (Big Data), IEEE, pp. 3873–3882.
Supervision. Kalia, A., Gosal, S.K., 2011. Effect of pesticide application on soil microorganisms. Arch.
Agron. Soil Sci. 57 (6), 569–596.
Kamilaris, A., Prenafeta-Boldú, F.X., 2018. Deep learning in agriculture: A survey.
Declaration of Competing Interest Comput. Electron. Agric. 147, 70–90.
Keskar, N. S., Mudigere, D., Nocedal, J., Smelyanskiy, M., & Tang, P.T.P., 2016. On large-
batch training for deep learning: Generalization gap and sharp minima. arXiv
The authors declare that they have no known competing financial preprint arXiv:1609.04836.
interests or personal relationships that could have appeared to influence Kobayashi, T., Kanda, E., Kitada, K., Ishiguro, K., Torigoe, Y., 2001. Detection of rice
panicle blast with multispectral radiometer and the potential of using airborne
the work reported in this paper. multispectral scanners. Phytopathology 91 (3), 316–323.
LeCun, Y., Bengio, Y., Hinton, G., 2015. Deep learning. Nature 521 (7553), 436.
Acknowledgements Liaghat, S., Balasundram, S.K., 2010. A review: The role of remote sensing in precision
agriculture. Am. J. Agric. Biolog. Sci. 5 (1), 50–55.
Liu, B., 2017. Sustainable strawberry production and management including control of
This work was also supported by Nova Scotia Research and Inno strawberry powdery mildew.
vation Graduate Scholarship Program and Dalhousie Entrance/In-course Liu, B., Zhang, Y., He, D., Li, Y., 2018. Identification of apple leaf diseases based on deep
convolutional neural networks. Symmetry 10 (1), 11.
Scholarship Programs. The authors would like to thank Millen farm and McGrath, M.T., 2001. Fungicide resistance in cucurbit powdery mildew: experiences and
Balamore farm to providing field access for image collection and challenges. Plant Dis. 85 (3), 236–245.
experiment. Mohanty, S.P., Hughes, D.P., Salathé, M., 2016. Using deep learning for image-based
plant disease detection. Front. Plant Sci. 7, 1419.
Statistics Canada, 2020. Table 32-10-0169-01 Number of farm operators by sex, age and
References paid non-farm work, historical data. Retrived from https://doi.org/10.25318/32100
16901-eng.
Adam, L., Somerville, S.C., 1996. Genetic characterization of five powdery mildew Qi, C. R., Yi, L., Su, H., & Guibas, L.J., 2017. Pointnet++: Deep hierarchical feature
disease resistance loci in Arabidopsis thaliana. Plant J. 9 (3), 341–356. learning on point sets in a metric space. In: Advances in neural information
Alom, M. Z., Taha, T. M., Yakopcic, C., Westberg, S., Sidike, P., Nasrin, M. S., ... & Asari, processing systems, pp. 5099–5108.
V. K. (2018). The history began from alexnet: A comprehensive survey on deep Ramcharan, A., Baranowski, K., McCloskey, P., Ahmed, B., Legg, J., Hughes, D.P., 2017.
learning approaches. arXiv preprint arXiv:1803.01164. Deep learning for image-based cassava disease detection. Front. Plant Sci. 8, 1852.
Asad, M. H., & Bais, A., 2019. Weed Detection in Canola Fields Using Maximum Rasti, P., Ahmad, A., Samiei, S., Belin, E., Rousseau, D., 2019. Supervised Image
Likelihood Classification and Deep Convolutional Neural Network. Information Classification by Scattering Transform with Application to Weed Detection in
Processing in Agriculture. Culture Crops of High Density. Remote Sens. 11 (3), 249.
Bianco, S., Cadene, R., Celona, L., Napoletano, P., 2018. Benchmark analysis of Robert, P.C., 2002. Precision agriculture: a challenge for crop nutrition management. In
representative deep neural network architectures. IEEE Access 6, 64270–64277. Progress in Plant Nutrition: Plenary Lectures of the XIV International Plant Nutrition
Bressem, K.K., Adams, L., Erxleben, C., Hamm, B., Niehues, S., & Vahldiek, J., 2020. Colloquium. Springer, Dordrecht, pp. 143–149.
Comparing Different Deep Learning Architectures for Classification of Chest Rosa, U.A., Upadhyaya, S. K., Koller, M., Josiah, M., & Pettygrove, S., 2000. Precision
Radiographs. arXiv preprint arXiv:2002.08991. farming in a tomato production system. In: Proceedings of the 5th International
Busseti, E., Osband, I., Wong, S., 2012. Deep learning for time series modeling. Technical Conference on Precision Agriculture, Bloomington, Minnesota, USA, 16-19 July,
report. Stanford University, pp. 1–5. 2000. American Society of Agronomy, pp. 1–15.
Cawley, G.C., Talbot, N.L., 2010. On over-fitting in model selection and subsequent Shafiee, M. J., Li, F., Chwyl, B., & Wong, A., 2017. SquishedNets: Squishing SqueezeNet
selection bias in performance evaluation. J. Mach. Learn. Res. 11, 2079–2107. further for edge device scenarios via deep evolutionary synthesis. arXiv preprint
Chen, Y., Lin, Z., Zhao, X., Wang, G., Gu, Y., 2014. Deep learning-based classification of arXiv:1711.07459.
hyperspectral data. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 7 (6), Shin, J., Chang, Y.K., Heung, B., Nguyen-Quang, T., Price, G.W., Al-Mallahi, A., 2020.
2094–2107. Effect of directional augmentation using supervised machine learning technologies:
Cruz, A., Ampatzidis, Y., Pierro, R., Materazzi, A., Panattoni, A., De Bellis, L., Luvisi, A., A case study of strawberry powdery mildew detection. Biosyst. Eng. 194, 49–60.
2019. Detection of grapevine yellows symptoms in Vitis vinifera L. with artificial Sutskever, I., Martens, J., Dahl, G., Hinton, G., 2013. February). On the importance of
intelligence. Comput. Electron. Agric. 157, 63–76. initialization and momentum in deep learning. In: International conference on
Deng, J., Dong, W., Socher, R., Li, L. J., Li, K., & Fei-Fei, L., 2009, June. Imagenet: A machine learning, pp. 1139–1147.
large-scale hierarchical image database. In: 2009 IEEE conference on computer Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., Rabinovich, A., 2015.
vision and pattern recognition, IEEE pp. 248–255. Going deeper with convolutions. In: In Proceedings of the IEEE conference on
Deng, L., Platt, J.C., 2014. Ensemble deep learning for speech recognition. Fifteenth computer vision and pattern recognition, pp. 1–9.
Annual Conference of the International Speech Communication Association. Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., Wojna, Z., 2016. Rethinking the
Durmuş, H., Güneş, E. O., & Kırcı, M., 2017, August. Disease detection on the leaves of inception architecture for computer vision. In: In Proceedings of the IEEE conference
the tomato plants by using deep learning. In 2017 6th International Conference on on computer vision and pattern recognition, pp. 2818–2826.
Agro-Geoinformatics, IEEE, pp. 1–5. Taylor, L., & Nitschke, G., 2017. Improving deep learning using generic data
Fadaee, M., Bisazza, A., & Monz, C., 2017. Data augmentation for low-resource neural augmentation. arXiv preprint arXiv:1708.06020.
machine translation. arXiv preprint arXiv:1705.00440. Teke, M., Deveci, H. S., Haliloğlu, O., Gürbüz, S. Z., & Sakarya, U., 2013, June. A short
Ebrahim, M., Alsmirat, M., & Al-Ayyoub, M., 2018, April. Performance study of survey of hyperspectral remote sensing applications in agriculture. In 2013 6th
augmentation techniques for HEp2 CNN classification. In 2018 9th International International Conference on Recent Advances in Space Technologies (RAST), IEEE,
Conference on Information and Communication Systems (ICICS), IEEE, pp. 163–168. pp. 171–176.
Frank, A.L., McKnight, R., Kirkhorn, S.R., Gunderson, P., 2004. Issues of agricultural Vincent, P., Larochelle, H., Lajoie, I., Bengio, Y., Manzagol, P.A., 2010. Stacked denoising
safety and health. Annu. Rev. Public Health 25, 225–245. autoencoders: Learning useful representations in a deep network with a local
Fuentes, A., Yoon, S., Kim, S., Park, D., 2017. A robust deep-learning-based detector for denoising criterion. J. Mach. Learn. Res. 11 (Dec), 3371–3408.
real-time tomato plant diseases and pests recognition. Sensors 17 (9), 2022. Wang, X., Peng, Y., Lu, L., Lu, Z., Bagheri, M., Summers, R.M., 2017. Chestx-ray8:
He, K., Zhang, X., Ren, S., Sun, J., 2016. Deep residual learning for image recognition. In: Hospital-scale chest x-ray database and benchmarks on weakly-supervised
Proceedings of the IEEE conference on computer vision and pattern recognition, classification and localization of common thorax diseases. In: Proceedings of the
pp. 770–778. IEEE conference on computer vision and pattern recognition, pp. 2097–2106.
Hossin, M., Sulaiman, M.N., 2015. A review on evaluation metrics for data classification Zhang, N., Wang, M., Wang, N., 2002. Precision agriculture—a worldwide overview.
evaluations. Int. J. Data Min. Knowle. Manage. Process 5 (2), 1. Comput. Electron. Agric. 36 (2–3), 113–132.