Spectral characterization and severity assessment of rice brown planthopper damage using multivariate models

Abstract

Brown planthopper (BPH) is a serious rice pest that threatens global food security by causing yield losses of up to 80%. Conventional methods for assessing BPH infestation are labour-intensive and lack real-time precision. This study evaluates hyperspectral remote sensing as a rapid, non-invasive approach for quantifying BPH population severity in three rice varieties: Pusa Basmati-1509, Pusa Basmati-1121, and TN-1. Leaf-level spectral measurements (350–2500 nm) acquired using a portable spectroradiometer effectively differentiated BPH population severity levels. Among 28 spectral indices evaluated, Structural Insensitive Pigment Index (SIPI), Pigment Specific Normalized Difference Index (PSND) for chlorophyll b, Pigment Specific Simple Ratio (PSSR a) for chlorophyll a, and (PSSR b) for chlorophyll b, showed high sensitivity to BPH infestation. Multivariate Regression models, including Partial Least Squares Regression (PLSR), Support Vector Machine (SVM), and Random Forest (RF), were developed for severity prediction. Among the tested models, RF achieved the highest accuracy for vegetation indices-based estimation (R² = 0.99), while PLSR showed strong relationships between hyperspectral data and BPH population severity (R² = 0.62) and key biochemical parameters, including chlorophyll (R² = 0.84), carotenoids (R² = 0.77), and protein (R² = 0.84). In contrast, flavonoids exhibited weak predictability (R² = 0.34). Field validation confirmed model robustness, with vegetation index-based predictions achieving R² values ranging from 0.72 to 0.86. Overall, the results demonstrate the potential of hyperspectral sensing combined with machine learning for early, non-destructive detection and monitoring of BPH stress, supporting precision pest management in rice.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-42245-0.

Introduction

With the global population increasing, improving crop productivity is essential to meet rising food demands. Cereals, particularly rice, form a staple diet for much of the world’s population, especially in Asia. However, insect pests pose significant threats to rice production, causing an estimated 25% yield loss and quality deterioration¹. Among these pests, the rice brown planthopper (Nilaparvata lugens), is one of the most destructive². It thrives in dense, nitrogen-rich vegetation and damages crops from panicle initiation to post-flowering stages. Both nymphs and adults feed on phloem sap, inducing symptoms such as yellowing, hopper burn, and sooty mold, often resulting in barren plants and yield losses ranging from 20–80%³. In addition, BPH acts as a vector for grassy stunt and ragged stunt viruses, further impairing plant growth and causing yield losses of up to 60%⁴.

Conventionally, BPH infestation is assessed through visual observation and field inspection by specialists or experienced farmers. However, these approaches are labour-intensive, subjective, and impractical for large-scale monitoring. Moreover, visible damage often appears only after substantial yield losses have already occurred, making early intervention difficult. Consequently, BPH remains a “hidden enemy”, emphasizing the need for timely detection and effective management. In this context, remote sensing technologies offer promising alternatives by enabling early, non-destructive detection of vegetation stress⁵. Advances in spectroscopy and remote sensing techniques have further supported the development of complementary approaches to traditional crop management^6,7. In recent years, hyperspectral remote sensing has gained attention for pest and disease monitoring due to its ability to capture plant chemical composition⁸, and subtle physiological changes associated with early plant stress^9,10.

Vegetation indices (VIs) derived from spectral reflectance are widely used to assess plant health and detect pest-induced stress. Indices such as Normalized Difference Vegetation Index (NDVI) and Modified Red Edge Normalized Difference Vegetation Index (MNDVI-705) effectively reflect changes in vegetation vigour and chlorophyll content under pest pressure^11,12, while Modified Chlorophyll Absorption in Reflectance (MCARI1), and Plant Senescence Reflectance Index (PSRI) are sensitive to capture physiological, water-related, and senescence-associated stress responses^13–15. Detection of spectral signatures associated with biotic stress supports effective pest management strategies⁵, and hyperspectral remote sensing in the Visible and Near-Infrared Region (VNIR) has proven effective for identifying BPH-induced stress in rice^16,17.

Despite the detailed information provided by hyperspectral data across the optical and near-infrared regions, their application is often constrained by high dimensionality and multicollinearity^18,19. Partial Least Squares Regression (PLSR) effectively addresses these challenges and has been widely applied to relate spectral reflectance with plant biophysical traits^20,21. Previous studies have reported strong relationships between hyperspectral reflectance and pest or disease severity in rice using regression and machine-learning approaches. Linear and multivariate models have successfully linked spectral responses in the green, red, and NIR regions with leaf folder damage, blast severity, rust infection, and Bipolaris oryzae disease, with techniques such as PLSR, Support Vector Machine (SVM), and Gaussian Process Regression (GPR) showing high predictive accuracy^22–26. Machine-learning approaches such as SVM and RF have been shown to improve predictive performance in high-dimensional spectral datasets^27,28.

However, studies integrating vegetation indices and biochemical parameters with machine-learning techniques for evaluating BPH population severity remain limited. To date, no study has systematically correlated biochemical parameters or compared multiple multivariate models for assessing BPH population severity across different rice varieties, including Pusa Basmati-1509, Pusa Basmati-1121, and the susceptible check TN-1. The present study addresses this gap by systematically comparing PLSR, SVM, and RF models for non-invasive assessment of BPH population severity in rice. The specific objectives were to develop and compare spectral models linking between vegetation indices, spectral data, biochemical parameters (chlorophyll, carotenoids, protein, and flavonoids), and with BPH population severity; and to evaluate the predictive accuracy of BPH population severity across three multivariate modelling approaches.

Results

Rice BPH scoring

The extent of damage caused by BPH infestation was assessed based on the proportion of plant tissue showing yellowing and complete drying. In this study, BPH population severity was quantified exclusively by using the number of BPH nymphs released per plant rather than visual damage scores. Infestation treatments were imposed by releasing predefined nymph densities (0, 5, 10, 20, 40, 80, 100, and 200 nymphs per plant), representing progressive severity from healthy plants (control) to complete hopper burn.

Impact of BPH population severity on leaf reflectance and the red-edge region

Increasing BPH population severity resulted in distinct and progressive spectral responses at both 20 (Fig. 1a–c) and 40 days after infestation (DAI) (Fig. 1d–f) across all three rice varieties. Reflectance differences between healthy and infested plants were analysed over the 350–2500 nm. Reflectance differences were minimal in the visible region (350–730 nm) but became more pronounced in the near-infrared (740–925 nm), followed by SWIR 1 (1400–1800 nm) and SWIR 2 (1950–2500 nm). These effects intensified from 20 to 40 DAI, with severe infestation causing reflectance patterns to approach soil spectral signatures (Fig. 1d–f).

Fig. 1.

Reflectance spectra of rice plants at different wavelengths in relation to variable BPH infestation in pot experiment at 20 DAI: (a) PB-1509, (b) PB-1121, (c)TN-1 and at 40 DAI: (d) PB-1509, (e) PB-1121, (f)TN-1.

Correlation analysis revealed a strong wavelength-specific relationship (r = 0.6–0.9) between reflectance and BPH population severity, particularly at 20 DAI. These sensitive spectral regions showed significant potential for early stress prediction, indicating 20 DAI as the stage for as optimal stage for early stress detection. In contrast, advanced damage at 40 DAI reduced spectral sensitivity.

The red-edge region (680–760 nm) exhibited high sensitivity to BPH population severity. Red-edge position (REP) consistently shifted downward with increasing infestation in PB-1509, PB-1121, and TN-1 at both sampling times. Regression analysis between red-edge value (REV) and BPH population severity yielded higher R² values. At 20 DAI, REV amplitudes in control and severely infested plants were 0.0171 and 0.0045 for PB-1509, 0.0128 and 0.0079 for PB-1121, and 0.0117 and 0.0067 for TN-1 (Fig. 2). At 40 DAI, REV amplitudes further declined to 0.0139 and 0.0026 in PB-1509, 0.0129 and 0.0012 in PB-1121, and 0.0015 and 0.0006 in TN-1 (Fig. 3). The largest REV reduction occurred in TN-1 under severe infestation, indicating greater varietal susceptibility to BPH-induced stress.

Fig. 2.

Red edge amplitude values of PB-1509, PB-1121, TN-1 at 20 days after infestation. The rice varieties were exposed to 0, 5, 10, 20, 40, 80, 100 and 200 BPH nymphs in Control, T1, T2, T3, T4, T5, T6, T7 treatments respectively.

Fig. 3.

Red edge amplitude values of PB-1509, PB-1121, TN-1 at 40 days after infestation. The rice varieties were exposed to 0, 5, 10, 20, 40, 80, 100 and 200 BPH nymphs in Control, T1, T2, T3, T4, T5, T6, T7 treatments respectively.

BPH population severity prediction using vegetation indices

In the current work, A total of 28 vegetation indices was derived from canopy reflectance (equations provided in S Table 1). All indices showed significant correlations with BPH population severity (S Table 2). Among them, SIPI, PSND-Chl-b, PSSR a, PSSR b exhibited the highest correlation coefficients (r ≥ 0.7–0.9) across varieties and sampling stages. These indices were therefore selected for multivariate model development.

Multivariate model development and validation

Three modelling frameworks were implemented: (i) vegetation indices vs. BPH population severity, (ii) spectral data vs. BPH population severity, and (iii) spectral data vs. biochemical parameters, using RF, SVM, and PLSR algorithms. Datasets from all three rice varieties were pooled across 20 and 40 DAI to improve model robustness and capture consistent physiological responses across genotypes and infestation stages. The feasibility of pooling was supported by the observation of similar spectral response patterns and monotonic declines in biochemical traits across varieties and infestation times. All models were externally validated using an independent field dataset to assess generalizability under natural infestation conditions.

Prediction of BPH population severity using vegetation indices and spectral data

The four highly correlated vegetation indices (SIPI, PSND-Chl-b, PSSR a, and PSSR b) were used for model development. The RF model achieved the highest predictive accuracy (; R² = 0.99; RMSE = 4.65; RPD = 0.02), followed by SVM (; R² = 0.96; RMSE = 9.26; RPD = 0.05) and PLSR (; R² = 0.92; RMSE = 14.72; RPD = 0.07) (Fig. 4). Although the models exhibited high R² values, the comparatively low RPD values can be explained by the statistical nature of the metric. RPD is calculated as the ratio of the standard deviation of the reference data to RMSE and is therefore strongly influenced by the variability and size of the dataset. The relatively narrow range of BPH population severity, together with the limited sample size, may have constrained RPD despite the strong agreement observed between predicted and measured values. Field validation showed strong agreement between predicted and observed BPH population severity, with R² values of 0.86 (SIPI), 0.72 (PSND-Chl-b), 0.86 (PSSR a), and 0.83 (PSSR b) (Fig. 5a-d) confirming model reliability under field conditions.

Fig. 4.

Prediction of rice BPH severity levels using multivariate regression models (RF, SVR, PLSR): Vegetative Indices vs. Population severity.

Fig. 5.

Highly correlated vegetative indices obtained from field data validation.

Spectral-based BPH population severity models showed moderate and comparatively lower predictive performance. PLSR produced the best spectral-based prediction (; R² = 0.62; RMSE = 31.99; RPD = 0.16), outperforming RF (R² = 0.57; RMSE = 33.92; RPD = 0.31) and SVM (R² = 0.22; RMSE = 45.75; RPD = 0.23) (Fig. 6). Field validation yielded a stable but moderate relationship between predicted and observed severity (R² = 0.59; Fig. 7), based on averaged field spectra (S Table 3; S Fig. 1).

Fig. 6.

Prediction of rice BPH severity levels using multivariate regression models (RF, SVR, PLSR): Spectral data vs. Population severity model.

Fig. 7.

Correlation of BPH population with average spectral data.

Prediction of biochemical parameters using spectral data

Biochemical parameters, including chlorophyll, carotenoids, protein, and flavonoids, declined progressively with increasing BPH population severity at both sampling times (S Fig. 2a-h). These parameters were measured from potted plants maintained under controlled glasshouse conditions, ensuring uniformity in infestation levels and treatment conditions. Predictive models were developed for each parameter using RF, SVM, and PLSR.

Stronger predictive relationships were observed for BPH population severity with chlorophyll, carotenoids, and protein. Among the tested models PLSR achieved the highest accuracy for chlorophyll prediction (; R² = 0.84; RMSE = 0.40; RPD = 0.10), followed by RF (; R² = 0.80; RMSE = 0.44; RPD = 0.11) and SVM (; R² = 0.77; RMSE = 0.47; RPD = 0.12) (Fig. 8). Carotenoid prediction was also best by PLSR (; R² = 0.77; RMSE = 0.07; RPD = 0.13), compared with SVM (; R² = 0.72; RMSE = 0.08; RPD = 0.14) and RF (; R² = 0.47; RMSE = 0.11; RPD = 0.20) (S Fig. 3). Protein estimation showed similar trends, with PLSR achieving (; R² = 0.85, RMSE = 4.01, and RPD = 0.11), outperforming RF (; R² = 0.80; RMSE = 4.66; RPD = 0.13) and SVM (; R² = 0.82; RMSE = 4.37; RPD = 0.12) (S Fig. 4). In contrast to flavonoid prediction exhibited substantially weaker performance across all models. The PLSR model yielded (; R² = 0.34, RMSE = 0.58; RPD = 0.21) while RF (; R² = 0.24; RMSE = 0.62; RPD = 0.23 ) and SVM (3; R² = -0.016; RMSE = 0.72; RPD = 0.26) showed even lower predictive capability (S Fig. 5). These results indicate limited predictive capability for flavonoids, under BPH stress conditions, and flavonoid-related models are therefore considered exploratory rather than robust.

Fig. 8.

Prediction of rice BPH severity levels using multivariate regression models (RF, SVR, PLSR): Spectral data vs. Biochemical parameter (chlorophyll).

Field validation demonstrated good agreement between predicted and measured biochemical parameters (S Fig. 6). Chlorophyll showed the strongest validation performance (R² = 0.80), followed by carotenoids (R² = 0.70), protein (R² = 0.69), and flavonoids (R² = 0.46) (S Fig. 7). These field validations represent an initial external assessment within the same season and agro-climatic region and therefore reflect preliminary applicability rather than full generalization.

Discussion

Spectral response to BPH population severity

This study demonstrated clear and systematic hyperspectral responses to increasing BPH population severity, quantified as graded nymph densities per plant, across rice varieties PB-1121, PB-1509, and TN-1. Control plants exhibited physiologically healthy conditions with intact pigments, cellular structure, and water content, whereas plants infested with 200 BPH nymphs exhibited severe stress symptoms. Among the tested varieties, TN-1 displayed the highest susceptibility, with rapid damage progression, while PB-1121 and PB-1509 exhibited relatively moderated responses.

Distinct spectral variations were consistently observed across the visible (blue, green, and red), red-edge, NIR, and SWIR regions, enabling reliable differentiation among BPH severity levels. In healthy rice plants, visible reflectance is typically low due to strong chlorophyll absorption, while NIR reflectance is high because of intact internal leaf structure and multiple scattering²⁹. BPH infestation disrupts these properties through pigment degradation, cellular collapse, and moisture loss, resulting in predictable spectral shifts with increasing pest density.

Reflectance in the red region increased progressively with infestation severity, with severely infested plants consistently showing higher red reflectance than healthy plants, a pattern reported previously for pest and disease-induced stress in crops^30,31. These changes were accompanied by reduced near-infrared reflectance, consistent with earlier studies^30,32.

Blue-green and red-edge wavelengths were particularly effective in distinguishing infestation severity across varieties. Williams et al.³³ similarly reported genotype-specific stress responses under biotic and abiotic stress in raspberry. Collectively, these results reinforce that sensitivity of the visible and red-edge spectral regions to BPH-induced stress in rice, supporting hyperspectral sensing as a non-destructive, and genotype-independent diagnostic approach.

Recent studies further highlight the growing role of machine-learning approaches in pest detection and forecasting. Algorithms such as Convolutional Neural Networks (CNNs), Random Forest (RF), and Support Vector Machines (SVM)and other deep-learning models have been shown to significantly enhance pest identification, outbreak prediction, and decision support in agriculture^34,35. These findings support our results, where RF and PLSR demonstrated strong predictive capability for BPH severity, aligning with the broader trend of ML-driven pest monitoring systems improving early detection and management efficiency³⁶.

In the SWIR domain, severely infested plants exhibited increased reflectance in SWIR-1 (1400–1800 nm) and SWIR-2 (1950–2500 nm) regions at both 20 and 40 DAI. This response reflects substantial moisture loss and leaf desiccation caused by prolonged sap feeding at the plant base, where BPH populations typically congregate.

At the highest infestation level, spectral signatures converged toward soil reflectance across VIS–NIR–SWIR regions, indicating severe loss of pigments, water content, and structural integrity. Such soil-like spectral convergence under extreme stress is physiologically expected and has been widely reported under conditions of severe moisture depletion and pest damage. Lesaignoux et al.³⁷ demonstrated that reduced moisture content drives reflectance patterns toward soil-like behaviour across the 0.4–14 μm domain. Similar soil-vegetation spectral convergence has been observed in rice blast and other crop-pest systems^26,38, highlighting a common stress-induced optical response under physiological degradation. These patterns were also reported for BPH in rice and grain aphids in wheat^39,40. Moreover, characteristic spectral responses have been documented for crop injuries caused by diverse insect pests, including leafhoppers, Solenopsis mealybugs, and aphids^41–43, reinforcing the broad applicability of hyperspectral techniques for agricultural pest monitoring.

Red-edge sensitivity and performance of vegetation indices

The amplitude of the red-edge peak declined consistently with increasing BPH population severity, highlighting the high sensitivity of the red-edge region to pest-induced stress. Comparable red-edge responses have been reported for rice blast²⁶, grey mould infection, and nitrogen stress in wheat^45,46. Here shifts in the Red Edge Position (REP) and Red Edge Value (REV) consistently reflected BPH damage severity across rice varieties⁴⁴.

Regression analysis revealed strong relationships between REV and BPH population severity, with R² values ranging from 0.75 to 0.96 across PB-1121, PB-1509, and TN-1 at both 20 and 40 DAI (Fig. 9a, c and e) and 40 DAI (Fig. 9b, d and f). TN-1 consistently exhibited the lowest REV values, confirming its higher susceptibility, whereas PB-1121 and PB-1509 maintained relatively higher REV values under comparable infestation levels. A steady decline in first-derivative reflectance within the 670–780 nm range accompanied increasing pest severity, consistent with reports from fungal pathogens studies such as Cercospora beticola and Uromyces betae⁴⁷.

Fig. 9.

Red edge technique and 1st derivative of spectral reflectance in a pot experiment for PB-1509, PB-1121, TN-1 at 20 and 40 DAI.

Evaluation of 28 vegetation indices demonstrated that indices related to green biomass, chlorophyll content, carotenoids, and dry matter were most effective for predicting BPH severity. These indices collectively capture key physiological disruptions including structural collapse, and moisture loss, rather than acting as independent indicators. This behaviour is consistent with earlier hyperspectral stress studies^48–50.

Among the evaluated indices, SIPI, PSND-Chl-b, PSSR a, and PSSR b showed the strongest relationships with BPH population severity. These indices effectively captured declines in photosynthetic efficiency caused by BPH-induced cellular damage, making them reliable indicators of pest severity. In contrast, several other indices exhibited weaker associations, indicating limited suitability for BPH severity estimation.

Our findings align with previous studies^51,52, reporting strong correlations between spectral indices and Diuraphis noxia abundance. Indices such as the Red-Edge Vegetation Stress Index (RVSI)⁵³, Yellowness Index (YI)⁵⁴, Anthocyanin Reflectance Index (ARI)⁵⁵, Carotenoid Reflectance Index (CRI)⁵⁶, and Water Band Index (WBI) have similarly been shown to detect stress^57,58. Spectral reflectance responses near 700 nm reflect biochemical alterations including chlorophyll, carotenoids, protein, and flavonoids, while bands around 800–900 nm region capture structural changes and moisture variations under stress^12,59,60.

Biochemical drivers of spectral changes

Biochemical analyses confirmed that increasing BPH population severity induced consistent and quantifiable physiological deterioration across all rice varieties. BPH infestation significantly reduced soluble protein and sugar contents, indicating impaired photosynthesis, assimilate translocation, and overall metabolic activity, consistent with earlier findings⁶³. Although contrasting biochemical responses have been reported under different pest systems, BPH feeding is known to suppress carbohydrate metabolism and pigment synthesis through sustained phloem extraction and physiological stress^64,65.

Chlorophyll content declined progressively with infestation, confirming its sensitivity to BPH feeding and in agreement with previous studies⁶¹. Among chlorophyll fractions, total chlorophyll exhibited the strongest decline, whereas chlorophyll b was comparatively less affected. Carotenoid content also declined with increasing BPH pressure, reflecting pigment degradation and reduced photoprotective capacity. This response aligns with studies highlighting the protective role of carotenoids against oxidative stress⁶². Uninfested plants consistently maintain higher carotenoid concentrations than infested plants across all sampling stages.

In contrast, flavonoids exhibited a non-linear, stage-dependent response, increasing at moderate infestation (20 DAI) and declining under severe stress (40 DAI). This biphasic behaviour reflects early inducible defense activation followed by metabolic collapse and aligns with molecular evidence linking flavonoids to BPH resistance mechanisms such as NlCDK1 kinase inhibition^66,67. Consequently, flavonoids behave as dynamic defense metabolites rather than stable stress indicators explaining their weaker spectral predictability.

Overall, most biochemical traits exhibited consistent infestation-dependent declines (S Fig. 2), providing a physiological basis for hyperspectral responses, while flavonoids reflected transient defense dynamics. These findings reinforce the value of integrating biochemical responses, hyperspectral sensing, and machine learning approaches. Traditional pest assessment approaches, typically rely on vegetation indices^55,68,69, population severity³², and biochemical indicators⁶⁴, often derived from limited visible-spectrum feature spaces. Recent advances in machine learning have substantially improved pest and disease detection and prediction^70,71, particularly when integrated with hyperspectral remote sensing, enabling more accurate quantification of physiological stress across infestation gradients.

Machine learning models’ comparison

Integration of hyperspectral data, vegetation indices, and machine learning models demonstrated strong potential for predicting BPH population severity. Among the evaluated approaches, the Random Forest (RF) achieved the highest accuracy, (R² = 0.99; Fig. 4), highlighting its ability to effectively capture complex non-linear relationships between spectral indices and pest population severity. Partial Least Squares Regression (PLSR) also performed robustly, particularly when linking full hyperspectral information with physiological traits and biochemical traits.

PLSR models showed consistent associations with BPH population severity (R² = 0.62; Fig. 6) and key biochemical parameters, including chlorophyll (R² = 0.84; Fig. 8), carotenoids (R² = 0.77; S Fig. 3), and protein (R² = 0.84; S Fig. 4). In contrast, flavonoid prediction exhibited substantially weaker performance (R² = 0.34; S Fig. 5), indicating limited spectral sensitivity under BPH stress conditions. Model performance was evaluated using complementary statistical metrics, including R², RMSE, and residual predictive deviation (RPD). R² reflects the strength of agreement between observed and predicted values, RMSE quantifies the average magnitude of prediction error, and RPD expresses prediction error relative to the variability (standard deviation) of the reference data. Because RPD is directly influenced by the spread of the response variable, low RPD values can occur even when R² is high if the dataset spans a restricted dynamic range or includes a limited number of samples, as previously reported in hyperspectral modelling studies^72,73. This explains the apparent discrepancy between R² and RPD observed in some models, particularly evident under controlled stress gradients.

The comparatively low predictive performance for flavonoids, reflected by both low R² and low RPD values (R² = 0.34; S Fig. 5), suggests limited modelling reliability and indicates that flavonoid-related results should be interpreted as exploratory rather than robust. In contrast, the comparatively stronger and more consistent predictive relationships observed for BPH population severity, chlorophyll, carotenoids, and protein suggest clearer spectral–physiological linkages and more stable modelling performance relative to flavonoids. These findings confirm the effectiveness of integrating hyperspectral sensing with machine-learning approaches for assessing BPH-induced stress in rice.

Practical implications for early pest detection

Data pooling across rice varieties and infestation stages was deliberately adopted as a modelling strategy to capture shared physiological responses to BPH feeding across genotypes, rather than cultivar-specific spectral signatures. Consequently, the developed models are intended to be variety-agnostic, supporting generalized assessment of BPH population severity across diverse rice cultivars. While this approach enhances model robustness by incorporating a broad physiological stress gradient, it may smooth finer genotype or stage-specific spectral differences. Accordingly, the present findings should be interpreted as a proof of concept for generalized BPH severity assessment, with further validation across additional genotypes, seasons, and environments required to refine cultivar-level sensitivity and improve broader generalizability.

Field validation confirmed the robustness and transferability of the developed models. Vegetation index-based models achieved R² values ranging from 0.72 to 0.86 (Fig. 5), while spectral reflectance-based models achieved an R² of approximately 0.59 (Fig. 7), for predicting BPH population severity. PLSR models linking hyperspectral data with biochemical parameters achieved R² values ranging from 0.34 to 0.85 (S Fig. 7), demonstrating their capacity to capture physiologically meaningful stress responses under field conditions.

Collectively, these results demonstrate the feasibility of integrating hyperspectral remote sensing and machine-learning approaches for early, non-destructive detection of BPH-induced stress. Similar success has been reported for pest monitoring across crops using other deep-learning frameworks³⁶, as well as PLSR-based severity prediction for rice diseases such as tungro, bacterial leaf blight, and blast⁷⁴. Regression based approaches often outperform classification models under overlapping severity gradients, a trend also reported for BPH, late blight, and powdery mildew³⁰. Support vector machines (SVM) have also demonstrated high predictive accuracy for multiple plant diseases, and when combined with principal component analysis (PCA), have achieved classification accuracies exceeding 90% in disease discrimination tasks^75,76. Together, these findings highlight the robustness and practical utility of machine-learning-enabled hyperspectral analysis, particularly RF and PLSR, for early pest identification and decision support in precision crop protection systems.

Limitations and future scope

Although the proposed framework demonstrated strong predictive capability, the experiments were conducted within a single growing season under controlled glasshouse conditions. Consequently, the independent field dataset collected within the same agro-climatic region represents an initial external validation of model transferability, rather than evidence of full generalizability. Pooling data across varieties and infestation stages enabled robust, variety-agnostic modelling by capturing shared physiological responses to BPH stress; however, this approach may have smoothed finer genotype- or stage-specific spectral differences. As a result, cultivar-level sensitivity and temporal variability were not explicitly resolved in the current framework. Future studies should therefore incorporate multi-season and multi-location field experiments, along with stratified modelling approaches to rigorously assess robustness across diverse environments. In addition, the integration of low-cost hyperspectral or multispectral imaging sensors with UAV or satellite platforms would enhance scalability and support practical relevance. Such deployment, however, should follow comprehensive multi-environment validation to ensure reliable performance under heterogeneous field conditions.

Methods

The experiment was conducted during the June to October (Kharif season) of 2022 at the ICAR-Indian Agricultural Research Institute (IARI), New Delhi (28°36’36’’N, 77°13’48” E), under controlled glasshouse conditions with a temperature of 28 ± 2 °C; 65 ± 2% relative humidity). Two widely cultivated basmati rice varieties, Pusa Basmati-1509 (PB-1509) and Pusa Basmati-1121 (PB-1121), and along with the susceptible check variety, TN-1^77,78 were used for the study. Nurseries for all varieties were prepared in 1-meter square areas, and 30-day-old seedlings were transplanted into 12-inch diameter pots. Pots were enclosed in a mylar cage to prevent any external infestation. Gravid BPH females were collected from paddy fields at ICAR-IARI and reared in controlled conditions on TN-1 (25 ± 5 °C, 70–75% relative humidity) until the second instar nymphs were obtained.

Infestation treatment

Rice plants at 20 days after transplanting were subjected to controlled infestations using second-instar BPH nymphs. BPH population severity, defined as the number of nymphs per plant, was imposed at eight predefined levels: 0, 5, 10, 20, 40, 80, 100, and 200 nymphs per plant. These nymph counts served as the quantitative severity variable in all statistical and machine learning models. These nymph counts were used directly as severity values in all analyses. Each infestation treatment was replicated five times across the three rice varieties.

Plants were also visually scored using the (0 = no damage; 9 = complete hopper burn) INGER scale⁷⁹ to validate the relationship between increasing nymph density and plant injury (S Fig. 8). Visual symptoms ranged from healthy green foliage (score 0) to extensive yellowing and complete drying of leaves (scores 7–9), confirming progressive damage with increasing infestation levels.

Each plant was enclosed in an individual cylindrical mylar cage with a muslin top to allow airflow and prevent insect escape. Symptoms such as leaf yellowing and drying were recorded qualitatively, throughout the infestation duration. Among the varieties tested, TN-1 served as a global susceptible check for pest resistance studies⁸⁰ and has a crop duration of 120 days. At the time of insect release, all varieties were at the vegetative tillering stage.

Spectral reflectance measurement

BPH damage on rice plants was evaluated using leaf reflectance spectra collected over 350 to 2500 nm. Spectral reflectance data were obtained from PB-1509, PB-1121, and TN-1, using a portable spectroradiometer (ViewSpec Pro, Analytical Spectral Devices [ASD]) with a leaf clip assembly equipped with wireless communication capabilities, allowed data acquisition and processing were performed using ASD software⁸¹. Measurements were taken at two infestation stages of BPH population severity at 20 and 40 days after infestation (DAI). For each replication, three spectral readings were collected, resulting in a total of 720 spectra (8 treatments × 5 replications × 3 readings × 3 varieties). Prior to measurements, the instrument was calibrated using a Spectralon reference panel.

Each observation consisted of an average of 50 spectra collected simultaneously. The leaf clip assembly provided illumination using an internal tungsten-halogen quartz lamp. Reflectance data were exported as tab-delimited text files using ViewSpec Pro software (version 4.05) (support.asdi.com/products.aspx) to enable data transfer and sharing. At the first measurement, PB-1509 and TN-1 were at panicle initiation, while PB-1121 remained at maximum tillering. At the second measurement, PB-1509 and TN-1 had progressed to the booting stage, whereas PB-1121 was at panicle initiation.

Spectral data pre-processing

Spectral data were pre-processed to reduce noise and improve multivariate model performance. A Savitzky-Golay smoothing filter was applied using a moving polynomial fit with a window size of (2n + 1), where n represents the half-window width⁸². This approach enhanced the signal-to-noise ratio and minimized high-frequency noise.

Spectral derivative and red edge analysis

Spectral derivative and red-edge analyses were performed to assess BPH-induced physiological changes. First-derivative reflectance was calculated, and red-edge characteristics were evaluated across infestation levels. The amplitude (dr_re) of the red edge peak, was estimated by fitting a second-order polynomial equation to the red and infrared slopes, followed by linear interpolation. Key parameters, including dr_re, λ_re and Σ (dr670-780) were analyzed, and the area under the red-edge curve, were analyzed to quantify spectral responses to increasing BPH population severity⁸³.

Prediction of BPH population severity using regression models

Vegetation indices vs. BPH population severity

Twenty-eight narrow-band vegetation indices were computed from leaf reflectance data to assess plant health and stress response (Index formulae provided in S Table 4). Indices showing strong associations with BPH population severity were selected for model development.

Spectral data vs. BPH population severity

Spectral reflectance measurements were analyzed to establish correlations with the extent of BPH infestation at two intervals: 20 and 40 days after infestation (DAI). These analyses aimed to establish relationships between reflectance patterns and BPH population severity, enabling prediction across a gradient of infestation intensities.

Spectral data vs. biochemical parameters

Biochemical assays were performed for all treatments, including severely dried plants. Remaining leaf tissue from dried plants was processed using identical extraction protocols. Chlorophyll and carotenoids were quantified using the DMSO method⁸⁴, with absorbance readings at 645, 663, and 480 nm. Protein content was measured using the Bradford method⁸⁵, Flavonoids were estimated using the aluminium chloride method⁸⁶. Data were analyzed using ANOVA and Tukey’s test (p = 0.05) with SPSS software (Version 20.0).

Field validation using an independent dataset

An independent field dataset was collected during Kharif 2022 at ICAR-IARI, New Delhi, to assess model robustness under natural infestation conditions. Rice fields were monitored for BPH incidence, and plants were individually tagged. BPH populations were recorded at 60, 70, 80, and 90 days after transplanting. Spectral measurements were collected at 90 days after transplanting using the same instrument settings as in the glasshouse experiment. Plants were grouped into ten population-severity classes based on BPH nymph density. For each class, vegetation indices and biochemical traits (chlorophyll, carotenoids, protein, and flavonoids) were estimated following the same protocols used in the glasshouse study. This dataset was used to externally validate three model categories: (i) vegetation indices vs. BPH population severity, (ii) spectral data vs. BPH population severity, and (iii) spectral data vs. biochemical parameters.

Multivariate machine learning techniques

Spectral reflectance, vegetation index, and biochemical datasets from all varieties (PB-1509, PB-1121, and TN-1) collected at 20 and 40 days after infestation were pooled prior to model development to capture the full physiological gradient of BPH population severity and reduce overfitting to variety- or stage-specific models.

Data partitioning

Datasets were randomly split into training (80%) and test (20%) subsets. The training set was used for model calibration and hyperparameter tuning, while the test set served as an independent evaluation dataset. This strategy was applied consistently across all models.

Model evaluation metrics

Model performance was evaluated using three standard metrics: coefficient of determination (R²), Root Mean Square Error (RMSE), and Residual Predictive Deviation (RPD). R² quantified the strength of agreement between observed and predicted values, RMSE measured prediction error, and RPD assessed predictive reliability relative to the variability of reference data. These metrics were applied uniformly across all modelling frameworks, including vegetation indices vs. BPH population severity, spectral data vs. BPH population severity, and spectral data vs. biochemical parameters.

where SD_reference represents the standard deviation of the reference values and RMSEP denotes the root mean square error of prediction. As R² and RPD capture different aspects of model performance, moderate RPD values may occur even when correlation is weak, particularly for biochemical traits exhibiting low spectral sensitivity or high biological variability^72,73. All spectral datasets were uniformly pre-processed using the Savitzky–Golay smoothing filter described above, and the identical preprocessing and validation workflows were applied consistently across PLSR, SVM, and RF models to ensure robustness of model comparisons.

Random Forest (RF)

RF is an ensemble learning method designed to improve predictive accuracy by aggregating multiple decision trees for regression analysis⁸⁷. In the present study, RF regression was implemented using the randomForest function from the “randomForest” package⁸⁸ in R software (version 3.5.0). The number of trees (ntree) was fixed at 500. The number of variables randomly sampled at each split (mtry) was optimized using single 5-fold cross-validation on the training dataset by minimizing the root mean square error (RMSE). Default hyperparameter settings were not used.

Support Vector Machine (SVM)

SVM is a margin-based learning algorithm widely applied to high-dimensional datasets such as Vis-NIR spectral data⁸⁹. In this study, SVM was employed in a regression framework to model continuous response variables. A linear kernel was used, and the regularization parameter (C) was optimized using a grid-search approach combined with single 5-fold cross-validation, with the optimal value selected by minimizing prediction error. SVM models were implemented using the “svm” function from the “e1071” package ⁸⁴ in R software (version 3.5.0).

Partial Least Square Regression (PLSR)

PLSR is a robust multivariate statistical method commonly used to address multicollinearity in hyperspectral datasets^90,91. PLSR reduces the predictor space into orthogonal latent variables (LVs) while maximizing covariance with the response variable. In the present study, the optimal number of latent variables was selected by minimizing the Root Mean Square Error of Cross-Validation (RMSECV) using single 5-fold cross-validation⁹². No fixed upper limit was imposed on the number of latent variables; LV selection was governed by data dimensionality and RMSECV minimization to avoid overfitting. PLSR modelling was implemented using the “plsr” function from the “pls” package⁹³ in R software (version 3.5.0).

All models were trained and evaluated under an identical preprocessing, cross-validation, and data-partitioning framework to ensure reproducibility and unbiased comparison of model performance. Spectral and biochemical datasets from PB-1509, PB-1121, and TN-1 collected at 20 and 40 days after infestation (DAI) were pooled to capture the full physiological gradient of BPH population severity, and the resulting models were designed to be variety-agnostic for broader operational applicability. Multivariate regression models (PLSR, SVM, and RF) were developed to relate vegetation indices, spectral reflectance, and biochemical parameters to BPH population severity, and model performance was validated using independent field data. The overall workflow and associated experimental procedures are illustrated in Fig. 10.

Fig. 10.

The overall workflow of the present study from BPH infestation to development of prediction models like PLSR, SVR, RFR.

Conclusion

This study demonstrates the effectiveness of hyperspectral remote sensing combined with multivariate modelling for quantifying brown planthopper (BPH) population severity in rice. Sensitive spectral regions in the red-edge and near-infrared domains, together with pigment-based vegetation indices (SIPI, PSND-Chl-b, PSSR a, and PSSR b), enabled reliable detection of pest-induced physiological stress. Comparative evaluation of modelling approaches revealed complementary strengths, with Random Forest performing best for vegetation index-based severity estimation and PLSR showing superior capability when integrating full hyperspectral information with biochemical traits. Importantly, the integration of spectral and physiological responses enables early, non-destructive detection of BPH stress prior to severe crop damage. While the models showed strong predictive performance and successful external field validation, the findings should be interpreted as an initial proof-of-concept evaluated within a single season and agro-climatic region. Overall, this work lays a foundation for future refinement and broader validation, with potential to support precision pest monitoring and early warning systems for sustainable rice production.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

EVM conducted the experiment and authored the draft of the manuscript. RNS, SC, SSS, and SR were responsible for designing and facilitating the research and revising the manuscript. SR, JSR and SP contributed to analyzing spectral data and revising the manuscript. All authors reviewed and approved the final version of the manuscript.

Funding

The research work is financially supported by ICAR-Network Program on Precision Agriculture (NePPA), Indian Agricultural Research Institute, New Delhi, India.

Data availability

The data supporting the findings of this study are available from the first author and the corresponding author upon reasonable request.

Declarations

Competing interests

The authors declare no competing interests.

Ethics approval

This research was conducted following the ethical guidelines of the institute.