This study establishes a framework for monitoring VW across vertically stratified cotton leaf layers based on hyperspectral data. Figure 1 illustrates the workflow encompassing in-situ hyperspectral data acquisition, vertical leaf layer feature optimization, and machine learning model construction, delivering layer-specific diagnostic precision across top, middle, and bottom leaf layers.

The trial was conducted in the field of VW disease at the Shihezi Institute of Agricultural Sciences (44.33°N, 86.05°E), Xinjiang (Figure 2). As Xinjiang is the most important cotton planting region in China, accounting for 23.1% of global cotton output and 90.2% of China’s cotton production (Feng et al., 2024). The region has a typical temperate continental climate with abundant sunlight. The national-level Verticillium wilt experimental field is cultivated with multiple candidate cotton germplasm lines and is artificially inoculated with the VW pathogen each year to maintain consistent disease pressure. Sowing was conducted from April 15-20, 2023, using ridge film mulching with on-film hole sowing and subsurface drip irrigation. The planting configuration followed a row spacing pattern of 10 + 66 + 10 + 66 cm. Irrigation was applied at intervals of 8–10 days, totaling 10–12 applications throughout the crop growth period. Fertilization included mono-ammonium phosphate (MAP) and urea, with potassium primarily supplied as available potassium (KCl), and potassium dihydrogen phosphate (KH₂PO₄) applied as a topdressing during the late growth stages. All fertilizers were dissolved in water and delivered via drip irrigation. The field is protected from external interference, and is irrigated and fertilized regularly to prevent drought stress or other stress that could affect VW monitoring.

The experiments were conducted from August 4th to September 10th, 2023. This period is the peak of cotton VW and covers the key growth period of flowering stage, experiencing the significant process from reproductive stages to vegetative stages. In this study, we randomly selected 11 distinct cotton plants from the experimental field and collected main stem leaves for analysis. Between August 4th and 10th, eleven cotton plants that had not shown symptom of VW were selected in the field. The typical main stem leaves at different leaf positions of these plants were used as experimental subjects to observe the full progression of the disease, from healthy to symptomatic, followed by leaf yellowing and senescence.

The occurrence of VW leads to a reduction in leaf pigment content, destruction of leaf cell structures, and a decrease in leaf water concentrations. These physiological changes are reflected in the spectral reflectance through light absorption and scattering interactions (Huang et al., 2018). In view of all layers, as shown in Figure 4 , the reflectance of cotton leaves consistently increased with disease severity from SL0 to SL4. As the disease progressed, the spectral differences between adjacent disease stages gradually expanded. Additionally, the envelope-removed spectra revealed that the absorption valleys across all spectral regions became more gradual with increasing disease severity, underscoring the progressive spectral alterations across all leaf layers. Compared to raw spectra, the envelope-removed spectra more effectively accentuated absorption-reflection features by eliminating background absorption effects, resulting in more distinct spectral patterns. The visible (400–700 nm) and red-edge region (≃ 705 nm) primarily reflected the degradation of the chlorophyll as VW infection progresses. In the blue (450–485 nm) and green regions (495–570 nm), the degradation of carotenoids and the accumulation of anthocyanins, respectively, become more evident with the severity of VW increased (Camino et al., 2021). VW-induced leaf structural degradation alters leaf thickness and dry matter content, impacting spectral reflectance in the red-edge and near-infrared plateau regions. Additionally, the absorption valleys in the near-infrared and short-wave infrared regions became progressively shallower, influenced by the decreasing leaf water concentration as the disease increased, leading to increased spectral reflectance.

Figure 5 illustrates the envelope-removed spectra of cotton leaves from different vertical layers (top, middle, and bottom) under different disease severities. Within the same disease severity level, spectral reflectance exhibits a decreasing trend as leaf layer moves higher. This trend is less pronounced in the near-infrared (NIR) region but becomes more apparent in the visible spectral region. Spectral reflectance increased with disease severity across all layers, with differences between adjacent severity stages becoming more distinct. Absorption positions were minimally influenced by leaf position, and the results align with those observed in the non-stratified dataset. However, notable differences in spectral patterns among layers were observed. From the top to the bottom layers, differences between disease severities became more pronounced under the same disease classification criteria. In the top layer, spectral differences between healthy, slightly and mildly infected disease stages were less evident, while differences between these stages and later stages (moderate and severe) were more distinct. As leaf position moved downward, spectral differences between healthy and mild disease stages gradually increased, along with those between moderate and severe stages. Among the three layers, the bottom layer exhibited the most prominent spectral differences across disease severities, followed by the middle layer, with the top layer showing the smallest differences, particularly under lower severity levels (SL=0-2).

As shown in Figure 6 , combinations of feature selection methods and classifiers exhibit significant variations in accuracy and computational efficiency for cotton VW monitoring. The RF-LightGBM combination achieves optimal performance, with a test accuracy of 0.69, surpassing other models. Specifically, RF-based feature selection outperforms Lasso and Relief-F, although Lasso and Relief-F demonstrated strong performance in specific models. When integrated with RF, LightGBM attains the highest training accuracy (0.82) and test accuracy (0.69), demonstrating its balance of precision and generalization. LightGBM also maintains stable test accuracies (0.66 and 0.67) when paired with Relief-F and Lasso, respectively, confirming its adaptability. While ANN under RF selection achieves comparable test accuracy (0.67), its runtime (103 seconds) triples that of LightGBM (36 seconds) and exists potential overfitting risks in small-sample scenarios. By contrast, XGBoost, SVM, and RF models exhibit lower test accuracies (<0.65) and prolonged training times. These results underscore the critical role of method compatibility in optimizing VW monitoring. The RF-LightGBM framework, with its high accuracy and efficiency, proves to be the most effective solution for precise, real-time monitoring of cotton leaf VW in field conditions.

Models for cotton VW monitoring were constructed across the top, middle, and bottom leaf layers using the RF feature selection method combined with the LightGBM classifier, as listed in Table 2 . The results revealed notable differences in model performance between the different leaf layers, with the bottom layer achieving the highest monitoring accuracy. This indicates that disease monitoring accuracy is influenced by the vertical stratification of cotton leaves.

The accuracy (precision) of the top layer model reached 0.83, while the middle layer performed similarly with an accuracy of 0.81. In contrast, the bottom layer achieved the highest accuracy of 0.91, making it the most effective for monitoring VW. The confusion matrices for different disease stages across the top, middle, and bottom leaf layers are shown in Figure 7 . For the monitoring of healthy, moderately infected, and severely infected leaves (SL=0, 3-4), the models across all three layers demonstrated high accuracy, indicating their effectiveness in identifying both healthy and advanced disease symptoms. For SL0 (healthy), the accuracy in the top, middle, and bottom layers were 1.00, 0.94, and 0.94, respectively. For SL4 (severely infected), all three layers achieved an accuracy of 1.00. However, when it comes to disease early-stage monitoring (SL1, slightly infected; and SL2, mildly infected), the performance varied significantly across layers. The bottom layer showed better accuracy for early-stage monitoring compared to the top and middle layers. The precision for SL1 (slightly infected) and SL2 (mildly infected) in the bottom layer reached 0.83 and 0.78, respectively.

The bottom leaf layer had the highest accuracy for monitoring cotton leaf VW across all disease severities. It’s performance in early disease stages also significantly outperformed the top and middle layers, highlighting the critical role of bottom leaves in early-stage disease monitoring. Additionally, the data used for modeling were derived from diverse cotton genotypes, further demonstrating the model’s robustness.

The importance of key features for monitoring cotton VW severity across different leaf layers and stages of disease development as shown in Figure 8 . For different disease severities, the top layer is primarily dominated by anthocyanin (Ant_Gitelson). This is followed by PRI₅₇₀, PSDNb and BF, which are closely linked to xanthophyll cycle, chlorophyll content and the blue band. In the middle layer, the blue band index (B), xanthophyll index (PRI₆₀₀), and anthocyanin index (Ant_Gitelson) are the most critical features. For the bottom layer, also xanthophyll index (PRI₅₇₀) plays a leading role, followed by the Skewness and the blue/red index (LIC). These findings indicate that anthocyanin, xanthophyll, and the blue band are effective features for VW monitoring across the top, middle, and bottom leaf layers. However, there are also distinct differences in features across leaf layers, due to variations in nutrient and light distribution across leaf layers, as well as differences in the physiological activities of leaves at distinct positions.

In the context of slightly infected monitoring, the importance of specific features across the three layers is also shown in Figure 8 . For the upper leaves, features representing the blue band, fluorescence, and chlorophyll content are key, with indices such as BRI, PRI₁, VOG₁, and PSDNb being the most critical. In the middle layer, blue band index (B) dominate, supported by entropy, and fluorescence index (CUR). For the bottom leaves, features related to water content (e.g., WSCT), Gitelson & Merzlyak (GM) and blue/red index (LIC) are critical. These findings highlight that the blue band is important in all three layers, and that key features for early monitoring across all layers are primarily related to pigments. Additionally, features linked to water content play an important role in VW early monitoring in the bottom layer.

In addition, signal features play a significant role in VW monitoring in middle and bottom layer, with entropy, and skewness ranking second, and fourth in importance for the middle, and bottom layers, respectively, as indicated by SHAP values. These signal features provide new perspectives for monitoring cotton VW, particularly during the slightly infected stage, by offering additional insights beyond traditional spectral indices.

The common approach for disease monitoring in cotton using remote sensing involves conducting spectral measurements on top layer leaves or mixed leaf samples. However, by the time disease is detected in top-layer leaves, it has often already reached an advanced stage, and the impact of leaf position on monitoring accuracy is overlooked. Considering the bottom-up spatial progression of VW, capturing spectral information from bottom leaves provides a more accurate representation of the practical early stage of infection in the cotton plant. This critical aspect is often ignored in existing studies (Lu et al., 2023). Affected by precipitation and temperature, the peak incidence period of VW in cotton in Xinjiang occurs between July and September each year. During this period, there are significant differences in pigment content among leaves at different positions, especially when nutrient transport is obstructed due to VW infection (Luo, 2010), which further exacerbates the differences between leaf positions. Additionally, the photosynthetic rate and nutrient allocation differ among leaves at different positions. These factors lead to spectral differences between leaves at different positions, which in turn affect the performance of cotton VW monitoring based on leaf spectra. Therefore, this study conducts experiments during the period when VW is prone to outbreaks, based on in situ leaf spectra from field experiments, focusing on analyzing the impact of different leaf layers on the monitoring of cotton VW.

As the disease progresses, lower leaves exhibit a more pronounced increase in spectral reflectance particularly in the visible region, and achieve higher VW monitoring accuracy compared to upper layers. These differences may be attributed to variations in light conditions and leaf age. On one hand, the distinct spectral responses among leaf layers originate from differences in light environments. The dense foliage architecture of cotton canopies, coupled with phototropism-driven leaf orientation, facilitates preferential PAR (photosynthetically active radiation) interception by middle and upper leaves, while creating persistent shading effects on lower canopy layers (Bertheloot et al., 2008). Longer wavelength light, with its stronger diffraction ability, is more likely to penetrate deeper into the canopy, while shorter visible wavelengths are predominantly intercepted by upper leaves (Li et al., 2022). Consequently, lower leaves adapted to low-light conditions develop thinner palisade tissues and lower chlorophyll a to b ratios (Yang et al., 2022a). This renders their spectral reflectance in the visible region potentially more sensitive to disease-induced chlorophyll degradation. On the other hand, age-related physiological decline exacerbates the vulnerability of lower leaves. Older lower leaves exhibit reduced photosynthetic efficiency (P_max), accelerated Rubisco degradation, and weakened stress resilience compared with younger upper leaves (Sassenrath-Cole et al., 1996). When VW infection occurs, pathogen effectors such as PevD1 further disrupt lower-leaf physiology by targeting senescence-associated genes (e.g., GhORE1), accelerating chlorophyll breakdown and cellular disorganization (Zhang et al., 2021), which amplify spectral variations under VW infection. In contrast, upper leaves with thicker cuticles, higher chlorophyll reserves, and active metabolic repair mechanisms tend to exhibit a “threshold effect” when facing damage from the VW pathogen, requiring larger lesion areas to trigger significant spectral changes. This physiological buffering may reduce spectral discriminability in upper layers during early infection stages. Thus, the synergistic effects of light limitation, age-related physiological decline, and pathogen-driven senescence in lower leaves may enhance spectral reflectance changes, enabling more precise VW severity monitoring in lower-canopy layers compared to upper ones.

Comparing the early (SL=0-2, including healthy, initial symptomatic, and early stages) and late (SL=3-4, including mid- and late stages) monitoring performance across top, middle, and bottom layers, late-stage monitoring showed consistently high accuracy across all layers. In contrast, monitoring during the slightly infected stage showed that the monitoring performance for VW in the bottom layer was significantly better than in the top and middle layers. Comparing spectral differences across leaves with varying disease severity shows that the spectral variations in the bottom leaves are greater than those in the middle and top leaves, which is particularly evident in the visible spectral region. The spectral reflectance in the visible bands is mainly influenced by pigments, especially chlorophyll content. The chlorophyll content varies significantly across leaf layers (Yang et al., 2022a). This vertical heterogeneity in pigment distribution leads to differences in key features for VW monitoring and affects the performance of monitoring models. Under VW infection, nutrient supply is remobilized from older bottom leaves to younger top leaves (Parkash et al., 2023). Consequently, even when leaves are classified into the same disease severity level based on lesion coverage in this study, the overall health condition of the bottom leaves is noticeably poorer than that of the top and middle leaves. Cotton bottom leaves tend to appear more yellow, with lower chlorophyll content, and have reduced photosynthetic activity (Parkash et al., 2023). Therefore, the spectral differences across bottom leaves with varying disease severity, particularly in the visible region, are more evident compared to those observed in middle and top leaves. At the mid-to-late stages of infection, extensive fungal invasion results in substantial physiological and optical changes in leaves across all layers. These changes become more uniform at severe disease stages, facilitating consistent VW monitoring regardless of leaf position. These findings demonstrate that bottom layer leaves have a distinct advantage in the monitoring of VW, especially during the slightly infected stages of the disease, compared to upper and middle leaves.

The key features for monitoring VW in cotton, especially vegetation indices derived from the spectral bands with short wavelengths, differ across leaf layers and disease stages. The results of this study indicate that the main indicators for monitoring VW in cotton are those related to chlorophyll degradation, carotenoid cycling, moisture content, fluorescence, and anthocyanins. The relevant spectral bands primarily include the blue, green, red-edge, and parts of the near-infrared region, particularly the absorption valleys affected by moisture content. These findings align with previous studies on crop diseases, which have established that visible and near-infrared bands are sensitive bands for identifying different crop diseases (Xie et al., 2018; Zhang et al., 2019; Lassalle, 2021; Feng et al., 2022; Shafik et al., 2023).

At five-levels cotton VW severity monitoring, important signatures for the top leaf layer include Ant_Gitelson, PRI₅₇₀, PSDNb, and BF, while the middle leaf layer is characterized by B, PRI₆₀₀, and Ant_Gitelson, all of which are closely associated with vegetation pigments. Plant physiology research indicates that chlorophyll degradation and anthocyanin increase occur in leaves infected by VW, mainly as a protective response to damage caused by the pathogen (Camino et al., 2021). In the top and middle leaf layers, exposure to higher light intensity and increased photosynthetic activity leads to more pronounced changes in these physiological parameters, making chlorophyll degradation and anthocyanin-related indices more effective for VW monitoring. Blue band-related indices (e.g., BF, B, LIC) and PRI derivatives (e.g., PRI₅₇₀, PRI₆₀₀) play a significant role in monitoring cotton VW, consistent with findings from previous researches (Calderón et al., 2015). These features are important across the top, middle, and bottom leaf layers. The infection of VW typically leads to chlorophyll degradation, with significant carotenoid changes, and a reduction in photosynthetic efficiency, which are driven by pathogen-host interactions but not influenced by leaf position. As a result, indices related to chlorophyll and carotenoid absorption in the blue band, along with photosynthetic efficiency-related indices, remain important across different leaf layers.

In early disease monitoring, key features identified differ somewhat from those at five disease severity levels. Chlorophyll fluorescence emission related indices, such as PRI₁ and CUR, play a significant role during the sightly infected stage, particularly in the top and middle layers. Fluorescence, as a by-product of photosynthesis, is closely tied to photosynthetic efficiency and photoprotection mechanisms, with potential to rapidly reflect plant stress states. In cotton plants, the average net photosynthetic rate decreases progressively from the upper to the lower leaves within the same growth stage (Li, 2007). In the slightly infected stage, upper leaves experience reversible inactivation of PSII reaction centers. Their exposure to higher light intensity further stimulates photoprotection mechanisms, including carotenoid cycling and non-photochemical quenching (NPQ) (Ruban, 2016), which leads to significant changes in fluorescence emissions. Similarly, middle leaves, benefiting from the cotton plant’s spiral leaf arrangement and canopy structure designed to optimize light utilization, also receive substantial light. This sufficient light exposure, during the slightly infected stages of disease, activates photoprotection mechanisms and results in significant changes in fluorescence emissions. In contrast, bottom leaves exposed to less light, exhibit weaker photoprotection changes, the fluorescence signals in these leaves are weaker and less responsive to early disease stress, making fluorescence-related indices less effective for monitoring early-stage disease in the lower canopy. However, the lower canopy exhibits a distinct advantage in early VW monitoring through water content-related indices (e.g., WSCT). Upper leaves, with stronger physiological activity, dynamically regulate water loss by closing stomata and reducing transpiration under early VW stress, thereby retaining water reserves (Pascual et al., 2010). In contrast, bottom leaves lack such regulatory capacity, leading to more pronounced water content variations during early infection.

It is noteworthy that certain bands play a critical role in monitoring cotton VW across different leaf positions and stages of disease development, particularly the blue and red-edge bands. In this study, a substantial number of features containing information from the blue and red-edge bands were identified as important for monitoring cotton VW. The red-edge band reflects both biochemical and biophysical factors, with its position shifting as chlorophyll declines and near-infrared reflectance rises due to structural changes (Barnes et al., 1992; PenUelas et al., 1995). It is consistently recognized as an effective wavelength for cotton VW monitoring (Yang et al., 2022b). However, this study explicitly indicates that, compared to the red-edge band, indicators related to the blue band, such as BF, B, BRI, and LIC, consistently prove to be highly significant for monitoring wilt severity across all stages of disease development, regardless of leaf layers. Also, these indicators demonstrate particularly strong performance in early monitoring. The blue band is highly sensitive to early disease stress due to changes in chlorophyll and carotenoid absorption. In the 450–499 nm region, chlorophyll strongly absorbs light, but its degradation during infection reveals carotenoid absorption features (Hennessy et al., 2020). This dual sensitivity to chlorophyll reduction and carotenoid prominence makes the blue band highly effective for monitoring subtle physiological changes during the slightly infected stage of disease. The blue band has consistently been considered as highly affected by disease stress at both the leaf and canopy scales (Poblete et al., 2020). Research on early monitoring of VW in olive trees has identified the blue/green index (BGI₁) and the blue/red indices (LIC₃ and BRI₁) as robust indicators for wilt monitoring (Calderón et al., 2015). The results of this study also confirm previous findings, specifically that the blue/red spectral index (i.e., BRI) exhibits differences between asymptomatic and early-stage infected vegetation (Zarco-Tejada et al., 2018; Poblete et al., 2020; Camino et al., 2021; Watt et al., 2023). This conclusion aligns with multiple studies on olive tree wilt caused by Verticillium, which share similar pathogenic bacteria, but has been overlooked in existing research on cotton VW monitoring. This study achieved this discovery by detecting subtle changes in cotton leaf characteristics during the early stages of symptom onset.

In addition to constructing vegetation indices, this study extends beyond vegetation index construction by systematically analyzing morphological characteristics across the full spectral domain. Thirteen signal features were extracted for monitoring cotton VW from a novel perspective. The results indicate that the signal features play a significant role in cotton VW monitoring, especially in middle and bottom layer. Specifically, entropy and skewness are crucial for disease monitoring, ranking fourth and second in importance for the middle and bottom layers, respectively, during the slightly infected stage. Entropy is crucial for the middle layer, as it reflects the more complex reflectance changes resulting from the combined effects of phototropism and disease-induced structural and biochemical alterations. For the bottom leaves, spectral differences in the visible and near-infrared regions are greater than in the shortwave infrared for leaves at different disease severity levels, which results in differences in the asymmetry of spectral distributions across disease stages, making skewness an important feature. These signal features demonstrate potential for early-stage monitoring of cotton VW.

Many studies have explored the combination of thermal infrared, RGB, leaf imaging spectra, and PSR point spectra for multi-source remote sensing data to collaboratively diagnose diseases, which is a popular approach in the field of disease monitoring (Zarco-Tejada et al., 2018; Camino et al., 2021). Some research on cotton VW has attempted to monitor the disease from the image and spectrum perspectives, respectively (Kang et al., 2022). However, the effectiveness of integrating spectral data and imagery for collaborative monitoring of cotton VW remains unclear. Existing studies have demonstrated the advantages of spatial-spectral combination models for plant trait estimation, suggesting that fusing images and spectral data can enhance robustness and accuracy (Li et al., 2020; Räsänen et al., 2020; Xu et al., 2020). Spatial information of images was not considered in this study, and further research should investigate whether combining image and spectral features could improve the early-stage monitoring accuracy of cotton VW.

This study demonstrates the capability of timely and accurate diagnosis of leaf symptoms using in situ ground-based hyperspectral data under natural solar illumination. The method, which captures whole-leaf reflectance spectra, is fast, simple, and representative in practice. Moreover, it offers potential for adaptation across various sensing platforms, providing technical support for field applications. However, environmental and plant growth conditions in the field, which can interfere with remote sensing monitoring of diseases, have not yet been considered. Further research should focus on time-series spectral data to assess the plant’s growing status in relation to environmental effects and disease infection. This will provide a theoretical basis and methods for crop disease monitoring in precision agriculture.