Smart Sugarcane Health Management: Combining Transformer Diagnostics and Adaptive Spraying Strategies

3.1 System overview

The proposed workflow of the proposed system which combines image-based disease detection, severity estimation, climate metadata and selective spraying recommendations through optimization techniques, is represented in Figure 1.

Figure 1. Overview of proposed system

3.2.1 Sugarcane image dataset

The dataset used in this work consist of images from five Mendeley Data repositories [23-27], which ensures the variety in leaf appearance, disease stages, lighting, and backgrounds. The dataset comprises 11 categories (Banded Chlorosis, Mosaic, Ring Spot, Viral Disease, Grassy Shoot, Pokkah Boeng, Rust, Yellow Leaf, Red Rot, Sett Rot), including 10 sugarcane diseases and healthy leaves, totaling 21,715 images. Images were manually reviewed to eliminate duplicates and low-quality samples, and disease labels were verified against published descriptors. The selected dataset provides a representative sample of sugarcane leaves in natural environments, making them suitable for training a robust multimodal classification-severity model. Sample sugarcane leaf images used for this work is shown in Figure 2.

Figure 2. Sample sugarcane images used for this work

3.2.2 Climatic dataset

Important variables like wind speed, temperature, weather, and humidity are included in the climate data used in this study, which was obtained from a public dataset [28]. It offers a representative sample of environmental patterns in important sugarcane-growing regions over a wide geographic and temporal range. The dataset was chosen because it closely reflects the conditions and information that farmers use when deciding how to apply pesticides. This makes it suitable for building a system that also considers environmental factors, which are essential for giving accurate and practical spraying recommendations.

3.3 Data pre-processing

A standardized pre-processing pipeline was used ensuring reliable and uniform quality across the disease datasets and the climatic inputs. The images of sugarcane leaves were resized to $224 \times 224$, the intensity was normalized, and the lighting variations were improved using CLAHE [29].

Table 1. Number of images per class in training, testing and validation sets

To improve variability and reduce overfitting, data augmentation methods such as rotation, flipping, scaling, and color jittering were employed. Temperature, humidity, wind speed, and disease severity were all synchronized hourly for analysis after the Kaggle climatic dataset was cleaned and aligned by interpolating missing values and eliminating unnecessary data [30]. Min-Max normalization was applied to continuous climatic variables to ensure model compatibility [31]. The image dataset was split into training (70%), testing (15%), and validation (15%) sets, while maintaining class proportions to avoid bias [32] (Table 1). Controlled augmentation and stratified sampling helped balance strategies and ensure stable learning across sugarcane disease classes.

3.4 Disease severity estimation model

3.4.1 Model architecture overview

The proposed disease assessment framework takes advantage of a dual-head CNN-Transformer architecture [33, 34] to effectively capture fine-grained lesion patterns and broader structural cues from the leaves of sugarcane. While the CNN backbone captures local texture variation such as the necrotic spot and chlorotic patch. The Transformer model captures the global relationships across the leaf surface. This combination increases the model's ability in recognizing early symptoms and distinguishing visually very similar diseases. Figure 3 gives the system-level block diagram showing the process for disease severity estimation.

3.4.2 CNN backbone

To detect the various types of lesions on sugarcane leaves and their colors, the CNN serves as a primary feature extraction method. The images have been resized to a 224×224×3-pixel resolution to ensure that all input images will have equal image quality. The CNN structure consists of five blocks of convolutional layers, which follow a standard sequencing of Conv2D, Batch Normalization, ReLU and MaxPooling [35]. The backbone uses ResNet skip connections between layers to further help with the improvement of the gradient flow within the network. The network also uses both $3 \times 3$ and $5 \times 5$ convolutional kernels at different scales, making it easier to identify lesions of varying sizes and shapes. The features collected at this local spatial level are enhanced, using a global context through the Transformer block module of the architecture. The parameters used in the CNN are given in Table 2.

The CNN architecture combined with the transformer module, models long-range spatial relationships across leaf surfaces, offering a wider understanding on lesions. Unlike conventional CNNs that capture lesions locally, the transformer addresses lesions distributed over larger areas or with overlapping characteristics. It segments lesions into patches and uses patch embeddings to create individual representations. Comprising four stacked encoder layers with multi-head self-attention (MHA) and feedforward sub-layers, the Transformer learns global interactions within lesions, enabling identification of infection stages and enhancing contextual understanding beyond local feature extraction.

The CNN model generates a feature map with multiple layers of depth, which is then divided into patches that do not overlap, and the patches are converted into tokens. After the tokenizing of the patches, each of them is processed through 4 layers of multi-head attention for global context modelling. The improved feature representation is then fused with the CNN feature presentation and passed into both the softmax classification output head and the regression output head that is used to determine the severity of the disease.

The Transformer uses the scaled dot-product attention mechanism [36], computed as given in Eq. (1)

$ Attention (Q, K, V)=\operatorname{softmax}\left(\frac{Q K}{\sqrt{d_k}}\right) V$         (1)

where, $Q, K$, and $V$ represent the query, key, and value matrices generated from the patch embeddings, and $d_k$ denotes the key dimensionality.

The multi-head attention (MHA) mechanism extends this by performing several attention operations in parallel as given in Eq. (2) allowing the network to simultaneously focus on different lesion features such as color variation, margin sharpness, and necrotic spread. Transformer parameters used in this work is shown in Table 3.

$M H A(X)= Concat\left(h_1, \cdots \cdots, h_H\right) W_o$           (2)

Table 3. Transformer parameters

3.4.4 Feature fusion and dual-head output

A framework using both CNN and transformer features can improve recognition of lesions and provide an estimation of lesion discrimination. A lightweight cross attention system allows the global token from the transformer to focus the important CNN features [37], which allows subtle signs of the early stages of infection to be highlighted. The fused image is then passed through two heads. The classification head contains a softmax classifying 11 categories of diseases while reducing overfitting through dropout; the regression head produces continuous values from 0 to 1 to indicating the severity of an affected leaf. The dual head structure allows the model to assess both disease diagnosis and severity, helping improve performance through two-way connectivity with agronomy in pesticide application decisions.

3.4.5 Model validation

Five-fold cross-validation was used for model validation on 21,637 images, guaranteeing balanced representation for precise performance metrics. Various architectures such as the ResNet50, DenseNet201, EfficientNet-B7, and ViT-B/16 was trained under consistent conditions and the performance was evaluated. The F1-scores showed clear and consistent improvements, confirming that the model’s gains are reliable.

3.4.6 SLIC super-pixel severity estimation

For estimating disease severity, we employed SLIC super-pixel segmentation using 500 super-pixels and a compactness of 20 [12]. The purpose of this module is to generate a lightweight and interpretable measure of disease intensity without requiring pixel-level masks or heavy segmentation models. Compared with Watershed and K-Means, SLIC produced cleaner lesion boundaries and showed fewer failures under challenging conditions such as uneven illumination or background clutter.

SLIC was specifically chosen because it remains reliable under real field constraints such as variable lighting, overlapping leaves, dust, and complex scenes where conventional pixel-wise annotations or fully supervised segmentation networks become impractical to deploy. This makes SLIC a field-ready solution that supports robust severity estimation for downstream decision-making.

3.5 Environmental metadata integration

3.5.1 Climatic variables

The four climatic factors (temperature, relative humidity, wind speed and weather condition) were used as environmental metadata to support in determining the disease severity and to make the recommended spraying decisions. Previous studies [38, 39] have observed that these climatic factors influence the growth of pathogens, movement of spores, development of leaf wetness, and progression of pathogens.

3.5.2 Multimodal data fusion

The integration of climate data and sugarcane images (by linking location) allowed the analysis of disease-related factors to include the environment. This is important to know how temperature, humidity, and wind speed affect disease spread and efficacy of pesticides, are critical for generating a model for accurate spray recommendations. Spearman correlation is used to determine time-dependent relationships between disease severity and climate variables and therefore address the issue of nonlinear biological responses.

3.6 Multi-objective optimization for spray recommendation

To formalize the spray recommendation strategy, a multi-objective optimization (MOO) framework was developed that integrates agronomic constraints, approved chemical guidelines, and disease-yield relationships. The decision vector was structured following standard MOO formulations [40], ensuring that each variable reflects a practical and controllable field parameter:

$S=\{T, D, C\}$             (3)

where, $T$ is the timing of spray after disease detection (hours), $D$ is the dosage level (L ha$^{-1}$ or % solution), $C$ is the chemical type (active ingredient/formulation).

The optimization problem now follows the canonical MOO form:

$\min F(S)=\left[f_1(s), f_2(s), f_3(s)\right], s \in S$             (4)

This fixes earlier uncertainties and making sure that the method aligns with standard MOO procedures.

(1) Yield Protection: $f_1(s)$

Instead of using approximate measures, yield protection is estimated from a yield-loss curve based on the predicted disease severity (DS).

$f_1(s)=1-Y(D S, D, T)$                (5)

where, $Y(D S, D, T)$ represents yield preserved under a specific dose timing strategy. This method matches the way disease severity and yield loss are related, as shown in plant pathology research [41].

(2) Treatment Efficiency: $f_2(s)$

Treatment efficiency represents suppression of lesion progression as mentioned in the previous works [42].

$f_2(s)=E(D, T)$         (6)

where, $E(D, T)$ captures fungicide effectiveness as influenced by dosage and spray timing.

(3) Chemical Load: $f_3(s)$

Directly measurable and widely accepted metric to measure chemical load [43] has been used in this study.

$f_3(s)=D$           (7)

This allows the optimization to minimize chemical input without relying on unvalidated ecological models. This practice aligns with previous NSGA-II agricultural optimization studies [22]. Yield protection serves as the primary agronomic goal, implicitly capturing the trade-off: higher doses (higher cost) must produce proportionally higher yield benefits [21].

3.6.1 Incorporation of approved chemical types and dosage ranges

For each disease predicted by the CNN-Transformer model, the system automatically maps to a valid treatment option defined by previous studies [44, 45]:

C: ICAR-recommended chemical

D: label-approved dosage

T: recommended timing window

This ensures that optimization operates strictly within realistic and legally compliant boundaries.

Table 4 presents all chemical types and dosage ranges used.

Table 4. Disease and recommended chemical types and dosage ranges used

3.6.2 NSGA-II optimization and pareto-optimal outputs

The NSGA-II method has been used to find optimal trade-offs between yield protection, treatment effectiveness and chemical loads through an iterative process of non-dominated sorts and selecting spray schedules by crowding distance, resulting in a set of Pareto optimal spray schedules. These spray schedules account for climate conditions, disease severity, amount of chemicals allowed, as well as their timing, therefore providing farmers with context relevant and practical spray schedules that provide maximum agronomic value with the least amount of chemicals needed. The workflow for the NSGA-II is given in Figure 4 and Table 5 gives the settings for the NSGA II.

Figure 4. NSGA-II workflow for generating spray recommendations

Table 5. NSGA II Configuration parameters