Abstract.

Sugarcane, being a crucial economic crop, is susceptible to leaf

diseases, resulting in reduced yield and compromised quality.
Traditional manual detection methods suffer from inherent
limitations in accuracy and speed. This paper proposes a new
ResNet-18-RSE-MRCCA model for diagnosing sugarcane leaf
diseases, which incorporates various attention mechanisms, such as
the Squeeze-and-Excitation Module, Multi-Head Self Attention
Module, and Criss-Cross Attention Module. The model automatically
extracts and learns the characteristics of sugarcane pests and
diseases from images, thereby improving the accuracy and
efficiency of classification. Trained initially on the PlantVillage
dataset, the model is then applied to the proprietary sugarcane leaf
disease dataset SLDD, comprising 8600 images with five disease
types against complex backgrounds. Experimental results
demonstrate significant improvements, with a 7.42% increase in
accuracy, 7.51% increase in recall, and 7% increase in F1 score
compared to baseline models. Comparative evaluations against six
classical CNN frameworks underscore the superior efficacy of the
proposed ResNet-18-RSE-MRCCA model. The application of multi-
level attention mechanisms in image classification tasks can
significantly enhance the performance of models, making them
more precise, with stronger generalization capabilities, and more
comprehensive information. Additionally, it can accelerate the
training speed of models, leading to greater advancements in image
classification tasks. This study marks the first application of the
ResNet-18-RSE-MRCCA model for diagnosing sugarcane leaf
diseases against complex backgrounds, offering promising prospects
for advancing plant disease diagnostic technologies.

Keywords: sugarcane disease; CCA attention; support vector machine; multi-head self-

attention ;convolutional neural network

1-Introduction

Sugarcane is the most important cash crop in the sugar industry. Diseases and
pests affecting sugarcane leaves can severely impact the growth cycle and yield of
sugarcane. It has been observed that the occurrence of diseases at different growth

1
stages of sugarcane can lead to a 24.9% reduction in yield, with a disease incidence
rate of 51%, and the most critical sugar content of sugarcane can decrease by
0.56%Error: Reference source not found. Traditional identification of sugarcane leaf
diseases and pests mainly relies on manual observation by workers, a method that is
not only prone to inaccuracies due to subjective experience but also inefficient.
Factors such as incorrect use of pesticides by farmers and inadequate subjective
understanding of various diseases can lead to a reduction in sugarcane yieldError:
Reference source not found. Therefore, achieving intelligent diagnosis of sugarcane
leaf diseases and pests is of great significance for improving both the yield and quality
of sugarcane. Most of the time, the classification of plant diseases and pests is based
on visual observation, where characteristics such as color, texture, and size of
diseased leaf parts are used to determine the type of diseaseError: Reference source
not found. However, this method of disease identification based on human visual
judgment not only consumes considerable time but also results in wastage of
manpower resources and economic losses.
With the rapid development of computer vision technology, significant progress
has been made in the identification and detection of sugarcane leaf diseases and pests.
Yigit et al.Error: Reference source not found used support vector machines to classify
sugarcane leaf disease states and healthy states, achieving an accuracy of 92% by
analyzing the different colors and texture features of the leaves. Basavaiah et al.Error:
Reference source not found proposed the use of decision tree classifiers, which
outperformed random forests by 4% in terms of accuracy due to their excellent
performance under small sample conditions. Ratnasari et al.Error: Reference source
not foundproposed an algorithm that achieved an accuracy of 80% by collecting RGB
images of diseased sugarcane leaves against a simple background and extracting
features using support vector machine algorithms, with feature matrices employing
gray-level co-occurrence matrices. To address the parameter auto-update algorithm in
support vector machines, Zhang et al.Error: Reference source not found proposed the
use of genetic algorithms to automatically optimize the kernel functions in support
vector machines, resulting in improved accuracy but increased time costs. Traditional
machine learning-based diagnosis of sugarcane leaf diseases and pests has drawbacks
such as high data annotation costs, limited model generalization capabilities, and
inability to handle dynamic changes. Additionally, Hossain et al.Error: Reference
source not found proposed the use of the K-nearest neighbor algorithm to classify
plant leaf diseases by extracting different detailed features from the diseased leaf
parts, such as color depth and texture fineness, achieving an accuracy of 96.2%.
However, the K-nearest neighbor algorithm may cause imbalance between training
time and accuracy. Therefore, Chanda et al.Error: Reference source not found
proposed optimizing the parameters of neural networks through the weights of
particles in particle swarm optimization to achieve automatic iterative updates,
achieving an accuracy of 96% in simple backgrounds. This approach allows for
automatic updating of parameters in neural networks. Initially, Lecun et al.Error:
Reference source not found first proposed the application of convolutional neural
networks in real-life situations in 1998. In recent years, various neural networks have
2
emerged. Sethy et al.Error: Reference source not found combined traditional support
vector machine algorithms with convolutional neural network algorithms, achieving
an accuracy of 98% in rice disease identification. Grinblat et al.Error: Reference
source not found also demonstrated that convolutional neural networks outperformed
traditional machine learning algorithms in plant recognition accuracy. Ghosal et
al.Error: Reference source not foundfound that the classification accuracy of similar
images, such as bacterial leaf diseases, was relatively low. Therefore, T. Daniya et
al.Error: Reference source not found proposed preprocessing RGB images using
regions of interest (ROI), mixing images of diseased and healthy leaves as input to the
network, initially classifying them into healthy and unhealthy categories using a two-
level classification method, and then subdividing them into three specific diseases.
This method achieved an accuracy of 93.04% and an F1-score of 91.42%. Sethy et
al.Error: Reference source not found classified 13 plant disease sites using support
vector machines and an improved neural network architecture, achieving an F1-score
of 98%. To improve accuracy, Militante et al.Error: Reference source not found
continuously stacked convolutional layers and pooling layers in neural network
models, achieving a maximum accuracy of 95.40% in classifying sugarcane disease
leaves under simple backgrounds. To obtain more accurate parameters, researchers
such as Ramachandran et al.Error: Reference source not found began using transfer
learning to identify different types of diseases and pests. Pereira et al.Error: Reference
source not found aimed to increase the depth of the network by increasing the number
of ReLU layers on the basis of adding convolutional layers. However, in practical
industrial production, more layers in the network do not necessarily result in better
performance. A greater number of network layers leads to increased computational
costs and waste of computing power, making it relatively difficult to apply. Tang et
al.Error: Reference source not found began using lightweight networks such as
ShuffleNet to detect grape diseases and pests, achieving an accuracy of 95.28%. To
explore more lightweight networks suitable for mobile applications, Howard et
al.Error: Reference source not found successfully applied MobileNet to mobile
devices. However, in actual field detection processes, due to complex backgrounds,
the accuracy of identifying diseases in seven types of cassava leaves was relatively
low. In recent years, different attention mechanisms have emerged. Hu et al.Error:
Reference source not found proposed the SqueezeNet attention module, which mainly
involves channel attention, learning different parameters in different channels. Li et
al.Error: Reference source not found introduced the SE module into convolutional
neural networks to enhance feature learning between different channels. The
TransformerError: Reference source not found module was initially used in the field
of natural language processing (NLP), where the introduction of multi-head attention
mechanisms made large models more sensitive to sequences. Since Transformer
models are often complex, Han et al.Error: Reference source not found first used
encoders from the field of natural language processing in convolutional neural
networks. Sun et al.Error: Reference source not found proposed improving the
relative position encoding in the encoder of the Transformer, ultimately achieving an
accuracy of 98% on public data. Due to the large size of Transformer models, Ma et
3
al.Error: Reference source not found proposed the necessity of using lightweight
networks. ShuffleNet achieved both model lightweighting and performance of
hyperparameters. Carion et al.Error: Reference source not foundwere the first to use
transformers for training in the field of object detection, achieving better results in
accuracy compared to traditional classical object detection frameworks.
Currently, the identification of sugarcane leaf diseases and pests based on RGB
images still faces some challenges: (1) Acquiring and annotating image data of
sugarcane leaf diseases and pests is a time-consuming and tedious task. The number
of publicly available datasets is limited, and the backgrounds are relatively simple,
which restricts the training and performance improvement of deep learning models;
(2) Sugarcane planting environments are diverse, and factors such as light,
temperature, and humidity may affect the appearance of leaves, thereby increasing the
difficulty of image recognition.
This paper presents a novel approach for detecting sugarcane leaf diseases and
pests in complex environments by leveraging a privately created dataset, SLDD, using
data augmentation based on Generative Adversarial Networks (GAN). The SLDD
dataset comprises 8600 images encompassing five different types of diseases, thereby
facilitating improved training effectiveness and generalization capabilities of the
model. Addressing the challenges encountered in detecting sugarcane leaf diseases
and pests in complex environments, the paper proposes a new model named ResNet-
18-RSE-MRCCA, which is built upon the lightweight ResNet-18 network
architecture. This model incorporates various attention mechanisms, including the SE
channel attention module, MHSA multi-head attention module, and CCA cross-
attention module, to effectively capture and learn disease and pest features from
images. As a result, the proposed model autonomously extracts and learns features
related to diseases and pests from images, thereby enhancing the accuracy and
efficiency of classification tasks.