Luopu County, administered by Hotan Prefecture in the Xinjiang Uygur Autonomous Region, lies along the northern foothills of the Kunlun Mountains on the southern fringe of the Tarim Basin, where an extremely arid continental climate prevails, marked by large diurnal temperature variations, low humidity, annual evaporation of 2,226.2 mm, and frequent dust-laden winds (Liang et al., 2005; Liu and Wei, 2005; Xu et al., 2009). Mean annual temperature ranges from 7.8°C to 12°C (extremes: –24.6°C to 40.1°C), while average annual precipitation is only 35.2 mm. The study was conducted in an 8-ha orchard in Kuochaiker Aizik Village, Luopu County ( Figure 1 ), planted mainly with the ‘Zha 343’ walnut cultivar, with a smaller proportion of ‘Xinfeng’. Approximately 70% of trees are 13 years old, 20% are 7–13 years old, and 10% are younger than 7 years, at a spacing of 6 m × 8 m.

To comprehensively investigate the progression of juglans leaf necrosis (JLN) at different growth stages, a series of UAV data acquisitions and ground surveys were conducted under clear, light-wind conditions on five key dates: June 16, July 2, July 30, August 12, and August 30, 2024. These time points were selected to capture critical changes during the growing season. Special emphasis was placed on the peak disease period typically observed in early July, thereby enabling a comprehensive assessment of JLN onset, severity, and subsequent decline under field conditions.

During each survey, we selected five disease plots, with 40 trees sampled from each plot ( Figure 1D ). To account for potential spatial variability within the canopy, leaves were collected from five canopy positions—east, south, west, north, and center—with three leaves per position, totaling 15 leaves per tree. Subsequently, the disease severity of these 15 leaves was evaluated to comprehensively assess the prevalence and severity of the disease across the entire canopy. The classification criteria for leaf-level disease severity are detailed in Table 1 . Photographs of diseased leaves at different severity levels are shown in Figure 2 . The calculation formula for leaf disease severity is presented in Equation 1.

We use the random forest algorithm to classify the severity of juglans leaf necrosis. The basic principle of this method was to extract and analyze spectral features related to canopy plant stress (Tuominen et al., 2008). Feature selection was first performed to identify vegetation indices sensitive to plant health change. Then, a Random Forest classifier was trained on labeled examples of JLN severity, and its aggregated predictions produced a robust classification (Vanguri et al., 2024; Yang et al., 2015). The RF classifier was implemented in ENVI with 500 trees, a maximum tree depth of 20, bootstrap sampling enabled, and Gini impurity as the splitting criterion. Class boundaries for severity levels were not manually set; instead, they were determined by the RF’s decision tree splits, ensuring a data-driven thresholding process. The performance of the Random Forest model was validated using a confusion matrix and a suite of evaluation indicators—including accuracy, precision, and recall—based on a sample of JLN severity data from ground surveys (Rissati et al., 2020). After classification, the results were refined to eliminate obvious errors and exclude patches located outside the walnut canopy. The severity of JLN was divided into five levels: Grade 0 (healthy) indicates no symptoms; Grade 1 (mild) indicates mild damage; Grade 2 (moderate) indicates larger, more continuous patches of leaf damage.; Grade 3 (severe) indicates severe crown damage; and Grade 4 (critical) indicates that most of the crown was affected (Wang and Jin, 2023). The classification standards are shown in Table 1 . To validate UAV-based disease detection, we conducted systematic ground surveys concurrent with UAV flights. A total of 1,000 georeferenced sampling points were established across the study orchards, ensuring representation of all disease severity levels (Grade 0–4). At each point, three to five walnut trees were visually assessed using the standardized juglans leaf necrosis (JLN) severity scale. Symptomatic and asymptomatic trees were recorded, and severity grades were assigned based on leaf chlorosis, necrosis percentage, and canopy damage patterns. All ground samples were geotagged using a sub-meter accuracy GNSS receiver and spatially matched to UAV-derived hyperspectral and RGB orthomosaics. The 1,000 samples were randomly split into a training set (80%) and validation set (20%) for supervised classification. Spatial and statistical analyses were conducted in ArcGIS 10.8 (Environmental Systems Research Institute, Inc., Redlands, CA, USA; https://www.esri.com)

To evaluate classification performance, we used a confusion matrix along with Overall Accuracy (OA), accuracy, precision, recall, and Cohen’s kappa coefficient (k). Cohen’s kappa provides a chance-corrected measure of agreement between the classified map and reference data. It is calculated as:

where ρ o     is the observed agreement and ρ e is the expected agreement by chance. The interpretation of k follows Landis and Koch (1977): values <0.00 = poor; 0.00–0.20 = slight; 0.21–0.40 = fair; 0.41–0.60 = moderate; 0.61–0.80 = substantial; and 0.81–1.00 = almost perfect agreement.

Hyperspectral data were processed in ENVI software (Exelis Visual Information Solutions, USA; https://envi.geoscene.cn/), and three types of indices were derived: greenness indices, pigment indices, and canopy water/photochemical efficiency indices. The greenness indices primarily reflect vegetation vigor by detecting variations in reflectance between red and near-infrared (NIR) bands, with healthy vegetation typically exhibiting higher values; we calculated nine such greenness indices, including NDVI, EVI, and various red-edge indices, as listed in Table 2 (Galvão et al., 2013; Liu et al., 2007). The leaf pigment indices, which depends on reflectance in the green and red bands, estimates leaf pigment content (especially chlorophyll) closely linked to photosynthetic performance and overall plant vitality; this study determined four pigment-related indices to provide insights into the nutritional status and health of walnut trees, also shown in Table 2 (Ahmad et al., 2020; Velichkova and Krezhova, 2018). Meanwhile, the canopy water or photosynthetic efficiency indices characterize vegetation water content and photosynthetic capacity, offering critical understanding of plant resilience under stress; here, four such indices were computed to assess potential water deficits and photosynthetic function decline in the orchards, as indicated in Table 2 (Oliveira et al., 2019; Lu et al., 2015).

When analyzing five stages of high-resolution UAV images of juglans leaf necrosis (JLN) in the experimental orchard, the dynamic changes in its occurrence and evolution could be clearly observed ( Figure 4 ). In the first and second stages, JLN had just occurred, and symptoms were not obvious. From locally magnified UAV images, the lesions did not exhibit obvious color or texture changes and appeared mainly healthy green ( Figures 4A, B ). This may be because the leaf damage area was still small in the early stage, and spectral reflectance characteristics had not yet changed significantly. As time passed to the third experimental stage, symptoms were significantly aggravated, and affected leaves showed large damaged area. High-resolution images revealed marked changes in the walnut crowns, manifested as dry leaves and decreased reflectivity in diseased parts, especially in the visible band. From the locally enlarged image at Stage 3 ( Figure 4C ), crowns displayed a grayish-white appearance, which contrasted with Stages 1 and 2. This change in texture and color provided a reliable remote sensing indicator for JLN.

In the fourth and fifth stages, the degree of juglans leaf necrosis was further aggravated. Large areas of leaves showed scorch characteristics, and the diseased area expanded features. During this period, the spatial distribution of disease features became clearer. In addition, some walnut trees were observed to have begun to sprout new leaves. This may be due to the recovery growth after infection. As a result, images displayed the coexistence of new and dead leaves, creating complex spectral mixing in multispectral imagery ( Figures 4D, E ).

From the analysis results, we found that the complexity of detecting JLN increases as the disease progresses. Specifically, in the early stage of the disease, symptoms were subtle and difficult to distinguish from healthy crowns in RGB UAV imagery. In later stage, due to the recovery of the walnut itself, newly grown leaves may mask the infection characteristics, making field-based assessments and traditional RGB remote sensing increasingly unreliable. Therefore, more advanced technologies are needed to more accurately detect and diagnose the occurrence and development of juglans leaf necrosis.

To rigorously validate UAV-derived hyperspectral classification of juglans leaf necrosis (JLN) severity, we conducted five dedicated field campaigns contemporaneous with flight missions. A stratified sampling design yielded 1000 Geo-referenced crown centers, where visual assessments followed the national JLN severity standard (grade 0–4). For every sample point reflectances corresponding to the 17 pre-selected vegetation indices (VIs) were extracted from the atmospherically and radiometrically corrected hyperspectral orthomosaic. The random forest algorithm was used to calculate the importance of each VI for JLN severity ( Figure 5 ). As can be seen from Figure 5 , MRESRI, CRI1 and PRI clearly dominated the feature-importance ranking, indicating strong sensitivity of red-edge and carotenoid-related metrics to JLN stress. In contrast, indices such as MRENDVI, WBI and VREI1 contribute marginally, suggesting limited additional discriminatory power once these indices are included.

The confusion matrix ( Figure 6 ) demonstrated an overall accuracy (OA) of 86% (172/200) and a Cohen’s k of 0.825, denoting substantial agreement between RF predictions and ground observations. Table 3 details class-wise statistics. Grades 0 and 4 achieved both high precision (≤ 6% false alarm) and high recall (≥ 85%), indicating that healthy and severely stressed crowns were rarely confused with other. Most of the 28 errors occurred between adjacent intermediate grades (1, 2 ↔ ↔, 3), which is expected given overlapping visual and spectral cues. Among them, Grade 1 had the lowest classification accuracy. Importantly, no misjudgments across more than two grades (such as 0→3 or 1→4), confirming reasonable decision boundaries. From the overall indicators ( Table 4 ), balanced/macro average recall was 0.86, macro precision was 0.865, and Macro F1-score was 0.861, further confirm the balanced performance of the model across all severity stages. The high OA, robust k statistic and biologically coherent error structure indicated that the VI-based RF model was well suited for operational mapping of JLN severity. In particular, the dominance of red-edge and pigment-related indices highlighted their value as early-warning proxies for canopy stress.

The classification results are shown in Figure 7 . From a temporal and spatial perspective, the classification results showed a clear trend in the occurrence and development of juglans leaf necrosis (JLN) in the experimental orchard, on which more detailed studies can be conducted to better understand the dynamics of the disease. The results also showed that it was feasible to calculate vegetation indices based on drone hyperspectral remote sensing data and used the random forest method to classify the severity of juglans leaf necrosis. The classification results of this method can fully help to understand the law of JLN spread and aggravation over time, as well as its spatial distribution evolution in different areas of the orchard.

Figure 7 shows the hyperspectral classification maps, while Figure 8 provides class-area statistics. Together, they show a clear escalation of JLN across the five monitoring stages.

To investigate the progression of juglans leaf necrosis (JLN) across different severity levels and to characterize the transition from mild to severe infection, we constructed transition matrices between disease severity levels at different time points. These matrices were derived from the statistical analysis of 3.2 classification results in ArcGIS 10.8 and are used to quantify the dynamic changes in disease severity during the monitoring period. To illustrate the transitions between different severity classes visually, Sankey diagrams were employed. As shown in Figure 9 , the diagrams clearly depict the flow of infected walnut trees between severity levels over time, providing intuitive insight into the development trajectory of JLN. This approach enables a better understanding of disease dynamics, highlighting the rates and directions of progression, persistence, or potential remission under field conditions.

From stage 1 to stage 2, 66.13% of the trees originally classified as severity grade 0 (healthy) remained stable, while 12.16% of those at grade 1 also maintained their status. Notably, 9.55% of trees transitioned from grade 0 to grade 1, and 3.64% advanced to grade 2. Additionally, 1.5% of trees at grade 1 progressed to grade 2. These findings indicate that the predominant trend during this stage was the early development of disease symptoms, primarily characterized by the progression from grade 0 to grade 1 and 2.

Between stage 2 and stage 3, the transition became more pronounced. Only 33.82% of grade 0 trees remained healthy, whereas the retention rates for grades 1 and 2 dropped significantly to 4.79% and 1.68%, respectively. Meanwhile, 17.41% of grade 0 trees progressed to grade 1, 15.67% to grade 2, and 6.37% to grade 3. Furthermore, 5.60% of grade 1 trees advanced to grades 2 and 3, while 0.54% of grade 2 trees transitioned to grade 3. This stage marked a rapid escalation in disease severity, with widespread movement toward higher severity levels.

From stage 3 to stage 4, the proportion of trees maintaining their original severity level continued to decline. Only 27.59% of grade 0 trees remained unchanged, followed by 6.58%, 9.15%, and 0.79% retention for grades 1, 2, and 3, respectively. Transitions from grade 0 to higher severity grades were observed at 7.02% (to grade 1), 3.79% (to grade 2), 2.88% (to grade 3), and 3.35% (to grade 4). Similarly, trees at grade 1 transitioned to grades 2 (4.55%), 3 (1.17%), and 4 (1.31%). Trees at grade 2 continued to progress, with 1.49% and 1.97% moving to grades 3 and 4, respectively, while 0.98% of grade 3 trees advanced to the most severe grade (grade 4). This stage reflected an acceleration in disease progression toward the advanced stages.

Between stage 4 and stage 5, disease advancement persisted, though at a slightly reduced rate. The stability rates for grades 0 through 4 were 26.31%, 6.57%, 9.15%, 0.79%, and 0.97%, respectively. Trees at grade 0 continued to show signs of infection, with 6.59%, 4.31%, 4.18%, and 5.13% transitioning to grades 1 through 4, respectively. Grade 1 trees progressed to grade 2 (3.47%), grade 3 (2.16%), and grade 4 (2.19%). Transitions from grade 2 to 3 (2.17%) and 4 (2.31%) were also observed, along with 1.25% of grade 3 trees reaching grade 4. These results indicate a steady but ongoing upward shift in disease severity, with increasing proportions of trees reaching moderate to severe grades in later stages.

The use of vegetation indices based on hyperspectral data proved essential and effective in assessing the severity of JLN. In this study, greenness indices, leaf pigment indices, and canopy moisture or photosynthetic efficiency indices were calculated to capture different physiological aspects of plant health. Among them, greenness indices such as NDVI, EVI, and RENDVI effectively detected changes in overall canopy health. This is consistent with a study on chlorosis in the Shulayar Reserve Forest in Kerala, which showed that such indices are sensitive to vegetation vigor and chlorophyll content and can be used to monitor forest canopy health (Ahmad et al., 2020). Leaf pigment indices such as CRI1 and ARI1 can be used to identify early signs of stress related to chlorophyll degradation and carotenoid accumulation, which can indicate the presence of JLN. This aligns with findings that the chlorophyll-to-carotenoid ratio is a key indicator of physiological stress and that hyperspectral indices effectively track pigment changes during plant development (Song and Wang, 2022). Canopy moisture indices such as PRI and WBI were particularly valuable for assessing water stress commonly associated with JLN, providing additional insights into the underlying causes of the disease. Similar results were obtained in studies on Scots pine and sorghum, where PRI and red-edge-based indices reliably captured carotenoid activity and water stress dynamics (Miettinen et al., 2025; Song et al., 2023).

In addition, studies on strawberry and sweet corn also showed that chlorophyll fluorescence indices and red-edge-based vegetation indices (e.g., CIREDEDGE, REIP) are very effective in detecting the early stages of heat, water, and nitrogen stress, which strengthens the role of using hyperspectral indices to distinguish different types of physiological stress (Poobalasubramanian et al., 2022; Sellami et al., 2022). These findings also validate that combining multiple vegetation indices targeting chlorophyll, carotenoids, and water content can provide a comprehensive and early assessment of JLN. Future work could further refine this approach, as demonstrated in studies using deep learning and hyperspectral signatures to monitor tree and crop stress (Moley et al., 2024; Deng et al., 2023; Bai et al., 2024; Otone et al., 2024).

Our classification analysis showed that the combination of MRESRI, CRI1, and PRI achieved the highest performance across various metrics (classification accuracy, precision, recall, and Kappa coefficient). This highlights the synergistic value of integrating indices targeting different physiological traits. Studies on squash and tomato diseases using UAV hyperspectral imaging also supported the advantages of combined metrics for early disease stage classification, achieving more than 90% accuracy even under field conditions (Abdulridha et al., 2020a, Abdulridha et al., 2020b).

However, we found that misclassification occurred mainly between the “healthy” (0) and “mild” (1) categories, likely due to overlapping spectral signatures. This challenge was echoed in studies of wheat rust and rubber tree disease, where the early stages of the disease are the most difficult to distinguish, even with advanced models (Zeng et al., 2024; Abdulridha et al., 2023). We also found that the winds in the Hotan area are strong, and the leaves of walnut trees often have dust attached and accumulated, which can affect the accuracy of classification and also increases the difficulty of early disease monitoring. Despite these problems, the high accuracy in classifying “healthy” (0) and “severe” (3) categories, as well as strong overall performance, demonstrates that UAV hyperspectral imaging, especially when combined with field validation, remains a powerful and reliable method for disease monitoring in walnut orchards and other areas.

High-resolution UAV imagery and hyperspectral classification showed clear spatio-temporal progression of JLN. In the early stages (stages 1 and 2), visual symptoms of JLN were mild and difficult to detect with traditional remote sensing methods, including high-resolution RGB imagery. Slight textural changes in the canopy observed in RGB imagery indicate early stress, but symptoms were not obvious. This limitation is echoed by other studies that highlight the lack of spectral sensitivity of RGB imaging to detect physiological stress before it becomes apparent (Kouadio et al., 2023).

As the disease progresses to stage 3, visible symptoms such as browning and curling and scorching of canopy leaves became more evident. At this point, hyperspectral imaging became increasingly effective in capturing disease-related changes. A combination of vegetation indices (MRESRI, CRI1, and PRI) provided the best choice for classification, consistent with other studies in which hyperspectral imagery outperformed traditional methods in detecting disease progression. For example, hyperspectral sensors with neural networks achieved accurate classification of Huanglongbing in citrus by detecting early canopy reflectance anomalies (Deng et al., 2020). This further proves that hyperspectral remote sensing combined with physiological indicators can reliably detect and quantify the disease in early and intermediate stages, even before visible symptoms fully appear (Kuswidiyanto et al., 2022).

The study found that the spatial distribution of JLN in the study area also showed different patterns. The results showed that hotspots of JLN occurrence occurred more frequently near roads and low-lying areas. These areas may be affected by microclimatic conditions such as increased temperature, altered airflow, or poor drainage. Similar spatial patterns have been found in remote sensing studies of citrus and palm diseases, where the disease was found to be more severe near human infrastructure or waterlogged areas (Alsadik et al., 2024). This further validates the role of local environmental variables in driving disease onset and highlights the importance of drone data for spatially targeted monitoring.

In-depth analysis of the temporal dynamics of JLN (Section 3.4) showed that disease severity had a significant and rapid progression between early and late monitoring stages. Initially, more than 70% of trees were healthy or only slightly affected with indistinct symptoms (grades 0-1), but by stage 3, moderate symptoms (grade 2) became common and severe infections (grade 3) rose to nearly 9%, with affected walnut trees showing obvious symptoms and concentrated in patches. By stages 4 and 5, the proportion of trees showing severe to critical symptoms increased dramatically, indicating that JLN has the ability to intensify in a short period of time. Similar time series drone studies have revealed similar trends, with scholars using continuous hyperspectral monitoring and tracking of biotic stress in vineyards and wheat fields, finding that biotic stress in the study area also expanded rapidly in a short period of time (Nguyen et al., 2021; Zhang et al., 2025). Interestingly, during the development of JLN, new leaves will grow, and some tree crowns will show temporary recovery, which will mask the disease signal and reduce the detection accuracy. This phenomenon has also been observed in related studies when using chlorophyll fluorescence and solar-induced fluorescence index for stress monitoring. This physiological improvement may produce false negatives in automatic models (Chang et al., 2020). These findings emphasize the importance of continuous and high-frequency monitoring using drone hyperspectral, which can not only capture the trend of disease escalation but also potential mitigation or recovery trends.

We used Sankey diagrams in Section 3.5 to illustrate the dynamic path of JLN from one severity level to another in the study area. During the early monitoring intervals (e.g., stage 1 to stage 2), most healthy trees (grade 0) remained in the same state, while only a small number of trees transitioned to grade 1 or 2. In contrast, in the mid-to-late intervals (stage 3 to stage 4 and stage 4 to stage 5), the severity accelerated significantly, with a considerable number of grade 1 and 2 trees progressing to grades 3-4. This pattern not only highlights the nonlinear progression of JLN but also indicates the importance of early management, where timely intervention management can prevent mild infection from evolving into large areas of burnt areas. Similar nonlinear disease progression patterns were observed in a drone-based maize study, where researchers used hyperspectral and wavelet features to capture disease transitions across time intervals. It was found that disease transitions at different levels in different periods were not linear, and the transitions were significantly accelerated in the later stages (Bai et al., 2024).

Notably, a small number of trees either remained stable in the mild category or transitioned downward from the moderately severe category to the mild category, a reversal that was often associated with new leaf growth. This reflects the heterogeneity of host responses, and some specific factors, such as genetic resilience or microclimate buffering, may attenuate or delay disease progression. This finding is consistent with the results of a time series study of Fusarium solani in cotton, which used deep learning and temporal modeling to explain changes in disease trajectories. The study found that some plants showed partial recovery or delayed symptom escalation during disease development (Abdalla et al., 2024).

For orchard managers, these findings imply that a one-size-fits-all approach to disease control may not be sufficient in orchard management and that more precise management, focused on “hotspots” or severely affected areas, is needed to produce cost-effective and timely results. This is consistent with recommendations from precision disease management studies that advocate the use of drones for spatial analysis and severity modeling, followed by site-specific precision interventions (Heidarian Dehkordi et al., 2020; Kouadio et al., 2023).

While drone-based hyperspectral imaging offers significant advantages for JLN monitoring, several limitations and challenges must be considered. A major limitation is the difficulty in accurately detecting early JLN due to subtle spectral differences between healthy and mildly infected trees. Even with optimized vegetation indices, distinguishing early symptoms from healthy canopies remains a challenge. This issue has also been noted in studies detecting early root rot effects on grapevines, where asymptomatic and diseased plants showed only slight spectral differences, and although machine learning approaches can improve classification accuracy, reflectance overlap between healthy states can lead to early misclassifications (Calamita et al., 2021).

Canopy complexity has also been found to complicate disease detection further. Heterogeneous canopies caused by mixed leaf ages, new sprouts, or overlapping branches can significantly distort spectral readings and increase classification errors. These challenges have been demonstrated in studies of oil palm diseases, where mixed classes within a single canopy often led to false positives, especially under shaded or partially obscured foliage (Anuar et al., 2021). Similarly, in apple orchards, studies have found that shadow pixels significantly reduce the accuracy of leaf area index (LAI) and chlorophyll content detection. Therefore, when using hyperspectral data for leaf area index (LAI) and chlorophyll content detection, shadow correction is required to reduce the error caused by it (Zhang et al., 2024).

Environmental conditions during drone flight (such as wind, light changes, and altitude changes) will also introduce noise into hyperspectral data sets, especially changes in light, which will affect the quality of hyperspectral data collection. Although the collected data requires further atmospheric correction and radiation calibration to reduce the errors caused by environmental conditions, the quality of the data may be limited by the constraints of on-site environmental conditions. This was also found in the citrus mold detection study, which showed that changing light conditions reduced the accuracy of the model during daytime flights, while the night vision-enhanced model provided more consistent results (Apacionado and Ahamed, 2023).

Although drones provide valuable insights, not all growers have access to this technology, especially those with limited financial resources or technical expertise. Similar concerns arise in a variety of crop systems, including viticulture and grassland management, where high-end sensors significantly outperform low-cost options but remain out of reach for most users (Hall and Lara, 2022). Therefore, reducing the cost and complexity of drone monitoring systems is critical for their wider adoption in precision agriculture.