Detection of Pine Wilt Disease Using a VIS-NIR Slope-Based Index from Sentinel-2 Data

Abstract

Pine wilt disease (PWD), caused by Bursaphelenchus xylophilus Steiner & Buhrer (pine wood nematodes, PWN), impacts forest carbon sequestration and climate change. However, satellite-based PWD monitoring is challenging due to the limited spatial resolution of Sentinel’s MSI sensor, which reduces its sensitivity to subtle biochemical alterations in foliage. We have, therefore, developed a slope product index (SPI) for effective detection of PWD using single-date satellite imagery based on spectral gradients in the visible and near-infrared (VNIR) range. The SPI was compared against 15 widely used vegetation indices and demonstrated superior robustness across diverse test sites. Results show that the SPI is more sensitive to changes in chlorophyll content in the PWD detection, even under potentially confounding conditions such as drought. When integrated into Random Forest (RF) and Back-Propagation Neural Network (BPNN) models, SPI significantly improved classification accuracy, with the multivariate RF model achieving the highest performance and univariate with SPI in BPNN. The generalizability of SPI was validated across test sites in distinct climate zones, including Zhejiang (accuracyZ_Mean = 88.14%) and Shandong (accuracyS_Mean = 78.45%) provinces in China, as well as Portugal. Notably, SPI derived from Sentinel-2 imagery in October enables more accurate and timely PWD detection while reducing field investigation complexity and cost.

1. Introduction

Bursaphelenchus xylophilus Steiner & Buhrer, commonly known as the pine wood nematode (PWN), is a destructive species that infects conifers and causes pine wilt disease (PWD) [1]. Native to North America, PWN has spread worldwide through the transportation of infected pinewood, with regions in the northern hemisphere being particularly affected, including China [2,3], Japan [4], Portugal [5] and Spain [6,7]. Favorable living conditions and the absence of natural predators in these areas further promote their rapid spread [8], threatening the balance of pine forest resources and ecosystems [9,10]. PWD is known for its high mortality rates, with infected trees typically dying within approximately two months [11]. If managed improperly, the disease can rapidly spread and devastate pine forests within 3 to 5 years [12]. Therefore, accurate identification and reliable mapping of PWD-infected areas are essential for guiding timely management responses and minimizing further ecological damage. Currently, remote sensing offers a practical approach for monitoring vegetation health and detecting potential PWD outbreaks [13]. However, variations in climate conditions can affect the reliability of traditional vegetation indices, while the high costs of data acquisition may hinder their large-scale applicability across different regions. These limitations highlight the need for a more robust and transferable spectral indicator to monitor the large spatial progression of PWD.

In recent decades, remote sensing technologies have been extensively applied in forest pest prevention due to their large-scale monitoring, high efficiency, and simplicity [14,15]. PWD disrupts water and nutrient transport in trees by obstructing their conduits and impairing photosynthesis, ultimately leading to wilting and death of trees [12,16]. After infection, pine trees exhibit significant morphological changes, with needle color changing from green to grayish-green and yellow, and eventually, the entire canopy turns reddish-brown [17]. These modifications of spectral properties can be detected through optical remote sensing techniques even before the symptoms become visible to human eyes [18]. Previous studies have attempted to monitor PWD using medium-resolution satellite data. For instance, Li [19] applied a random radiation transfer model to estimate infected areas using moderate-resolution imagery, achieving R² values between 0.48 and 0.57. Kuang [2] compared the ability of Landsat and Sentinel-2 to monitor PWD. The results showed that Sentinel-2 had a better recognition ability than Landsat, with classification accuracy of 78.91% and 77.87%, respectively. This difference was attributed to the advantage of Sentinel-2 in spatial resolution. The limited performance of these approaches is largely attributed to the coarse spatial resolution of medium-resolution satellites, where individual pixels often contain mixed signals from surrounding heterogeneous land cover components. This spectral mixing effect reduces the purity of pixel-level information and hampers the accurate detection of PWD.

To overcome these issues, research has increasingly focused on close-range sensing, particularly unmanned aerial vehicles (UAVs) equipped with RGB cameras or hyperspectral sensors [20,21,22]. UAV-based remote sensing, characterized by high spatial resolution, has significantly improved the accuracy of PWD detection, with reported R² values reaching up to 0.9 [23,24]. However, it often faces challenges such as high operational costs, weather dependency, restricted airspace assessment, and the complex topography of forested areas. These constraints limit its feasibility for large-scale or long-term applications. Consequently, commercial satellites with high spatial and temporal resolution have emerged as a promising alternative for PWD detection. Takenaka [25] monitored PWD using the Worldview data (0.46 m), achieving an overall accuracy of up to 0.98, while Zhou [26] employed Beijing-2 satellite imagery (1 m) and reported the recognition accuracy of the test data set as high as 0.99. These studies demonstrate that commercial satellite imagery can effectively combine the wide-area capabilities of satellite platforms with the fine spatial detail typically associated with UAVs. Nevertheless, the commercial nature of such data imposes substantial financial costs, which restrict their accessibility and widespread adoption. As a result, achieving high-accuracy PWD detection at a low cost remains a critical challenge in the field of forest pest management.

There are two prominent methods in PWD detection, including image-based visual recognition technology and classification methods based on remote sensing indices. Most computer vision methods rely on RGB imagery, which typically requires large annotated training datasets and high-resolution inputs to achieve accurate detection results [23,27]. Regarding index-based classification methods, traditional vegetation indices derived from fixed band ratios or differences may be easily affected by objective factors such as environmental differences [28], which reduces their reliability in detecting subtle physiological changes associated with PWD. In response, Kim [28] proposed the green-red spectral area index (GRSAI), which quantifies the area under the spectral curve between the green and red bands to detect infection-induced changes. The results indicated that leveraging broader spectral regions can enhance index stability and reduce temporal variability compared to traditional indices based on discrete spectral bands. However, indices such as GRSAI lack integration of spectral information from the red-edge and near-infrared (NIR) bands, which are known to be sensitive to vegetation stress, may constrain their effectiveness in identifying spectral changes related to PWD infection.

Based on these facts, our study aims to develop a novel vegetation index derived from open-access satellite data, utilizing spectral gradient features within the visible to near-infrared (VNIR) range. The index is designed to capture key spectral responses associated with chlorophyll anomalies, thereby improving the reliability and accuracy of PWD detection. Given the increasing availability of hyperspectral satellite missions, the proposed index’s scalability to hyperspectral data underscores its potential for broader application. The higher spectral resolution is expected to optimize its performance further, enabling more accurate and sensitive PWD monitoring. A further goal is to conduct cross-validation of the developed monitoring model for PWD for two different sites in China and one in Portugal. By assessing the model’s performance in these two countries, we will evaluate the transferability and generalization of the monitoring method under varying environmental conditions, thus providing a robust and widely adaptable remote sensing tool for international forest health monitoring.

2. Data and Methods

2.1. Study Area

Figure 1 shows the investigated regions in China and Portugal. Lishui (28.685° N–28.704° N, 119.332° E–119.355° E) is one of the Chinese locations situated in the southwest of Zhejiang Province. It is characterized by a mountainous landscape with a subtropical monsoon climate and four distinct seasons. The average elevation is around 1500 m, and the region is covered by nearly 450,000 hectares of pine monocultures. Weihai (37.367° N–37.531° N, 121.851° E–122.044° E), the second study area, is located on the eastern edge of the Shandong Peninsula, bordered by the sea to the north, east, and south. It has nearly 80,700 hectares covered by pure pine trees. A north temperate continental monsoon climate, characterized by mild temperatures and ample rainfall, provides favorable conditions for the growth and spread of PWD. To evaluate our model’s performance under different climatic conditions, we further chose a test site in Portugal. It is located in south-western Europe (36.967° N–42.133° N, 6.15° W–9.169° W), bordered by Spain in the east and north and by the Atlantic Ocean in the west and south. Portugal is characterized by extensive forest districts, with pine forests covering nearly one million hectares, about one-third of the country’s total forest area. In Europe, the PWD first appeared in 1999 in Portugal [5]. Despite ongoing control efforts, pine trees in Portugal remain relatively vulnerable, and the disease spreads more rapidly there than in North America, where native host species tend to be more tolerant and less susceptible to PWD infestation [29].

2.2. Data Used

2.2.1. Satellite Data

We used data recorded by the multispectral instruments (MSIs) aboard the Sentinel-2A and 2B satellites from ESA’s (European Space Agency) Earth observation mission as part of the Copernicus Program. Both satellites can alternately observe the Earth’s surface with a revisit time of five days. Therefore, due to their relatively high spatial and spectral resolution, Sentinel-2 has offered significant advantages for change detection purposes since 2015. We used Level-2A (L2A) products, which provide surface reflectance (bottom of atmosphere (BoA) data). We selected bands B2 to B7 for our approach since the VNIR region, especially the bands covering the red edge, is particularly sensitive to subtle changes in plant physiology related to stress, disease, and chlorophyll content. To ensure consistency in spatial resolution, we resampled all bands to a 10 m ground sampling distance (GSD). To investigate the influence of spectral signatures of droughts on those induced by PWD before 2015, we had to select data from the Operational Land Imager (OLI) sensor aboard the US Landsat-8 satellite from 10 October 2014, as a substitute. Given that the spatial resolution of the OLI sensor is lower than that of Sentinel-2, our analysis focuses on spectral information rather than spatial detail. We adopted bilinear interpolation to derive a comparable GSD of 10 m, as it better preserves the spectral characteristics of the original pixels. This method was selected to avoid potential spectral distortions or artefacts. The detailed data used are shown in

Table 1. Acquisition dates of the used satellite data.

Study Area	Satellite Data	Auxiliary Data
	Sentinel-2	Sentinel-2	GF-5B	Landsat-8
Zhejiang Province, China	10 October2022	3 May 2022, 7 June 2022, 17 June 2022, 27 June 2022, 17 July 2022, 27 July 2022, 1 February 2022, 6 August 2022, 11 August 2022, 16 August 2022, 21 August 2022, 26 August 2022	1 October 2022, 28 July 2022	*
Zhejiang Province, China		5 September 2022, 10 September 2022, 20 September 2022, 25 September 2022, 30 September 2022
Shandong Province, China	25 October 2022	25 September 2022, 28 May 2022. 1 October 2019	24 September 2022, 31 May 2022	10 October 2014
Portugal	27 September 2016, 7 October 2016	*	*	*

Note: * denotes the absence of data.

.

To measure the chlorophyll content and evaluate the sensitivity of various spectral indices, we obtained AHSI (visible to short-wave infrared hyperspectral camera) level 1 data from the Chinese GF-5B hyperspectral satellite. The GF-5 is the first international full-spectrum hyperspectral satellite comprising GF-5A and GF-5B and was designed for comprehensive observations of the atmosphere and land surfaces. While GF-5A was the first to be launched in 2018, it encountered operational failures, and GF-5B was subsequently launched in 2021 as a replacement to continue the mission. The AHSI spans the wavelength range from 400 to 2500 nm, based on 310 usable contiguous bands with a spectral resolution of 5 nm. The provided GSD of 30 m was also resampled to 10 m to align with the spatial resolution of the MSI and OLI data. As the AHSI data are delivered as raw data (L1), we utilized ENVI 5.6 software in conjunction with a provided satellite plugin to accomplish radiometric calibration, atmospheric corrections and orthorectification.

2.2.2. Field Measurements in China

In 2022, field measurements were conducted in the provinces of Zhejiang and Shandong from 7 October to 27 October. At this time of the year, the PWD symptoms, including canopy discoloration and wilting, are more prominent. Additionally, the disease has often progressed to a stage where its impact on tree health is more pronounced, making detection and assessment easier. In Lishui, Zhejiang, a total of 196 samples were measured, comprising 98 from diseased trees and 98 from healthy trees. In the Shandong province, a total of 493 samples were collected, comprising 246 from diseased trees and 247 from healthy trees. The locations of all sample points were recorded using a handheld GPS within 10 × 10 m plots, with a spacing of approximately 30 m between plots to ensure spatial separability and compatibility with the resolution of the selected satellite imagery.

All samples were processed according to this procedure. A total of 70% of the 196 samples from Lishui, Zhejiang were randomly selected as a training set (n = 137) and 30% as a validation set (n = 59). Simultaneously, all 493 samples from Weihai, Shandong, were used as test sets to assess model generalization ability. The spatial distribution of sample points is shown in Figure 2.

2.2.3. Observed Data of PWD in Portugal

PWD exhibits strict host specificity, affecting only pine trees. To assess the robustness of the monitoring model under varying climatic and landscape conditions, we further examined observational data on the extent of PWN provided by Portuguese authorities, namely the Instituto da Conservação da Natureza e das Florestas [30]. This dataset assessed the impact of an invasive forest pest in the Natura 2000 network of protected areas in Portugal during October 2016 and depicts the spatial distribution of the areas infected by PWD in Portugal. Therefore, we used the ‘FROM_GLC10′ land cover dataset of 2017 [31] for existing pine forests in Portugal and the survey data of the Natura 2000 network for the diseased trees as a reference. Based on these two datasets, we identified infected trees using our developed index and validated our proposed method.

2.3. Methods

2.3.1. Genetic Analysis of Samples for PWN Identification

Samples were collected from different parts of the trees, such as bark, wood (especially near the infection site), and roots with visible symptoms and signs of potential infection. They were placed into sterile test tubes or sealed bags and stored under appropriate cold conditions for transportation. For genetic analyses, we employed a widely accepted molecular detection technology known as the Polymerase chain reaction (PCR). Known for their high sensitivity and accuracy, they allow for the amplification of specific deoxyribonucleic acid (DNA) segments, enabling the detection of target DNA even at very low concentrations. Sample preparation was performed using the Baermann funnel method [32] and inverted stereoscopic microscope observations. Therefore, an effective separation and identification of nematodes from the collected samples is ensured, and the nematodes of interest can be isolated for further analysis. Subsequently, the DNA was extracted from samples using the cetyltrimethylammonium bromide (CTAB) method. This step is crucial for obtaining high-quality DNA and accurate downstream molecular detection. The PCR was performed on the internal transcribed spacer (ITS) region of ribosomal DNA (rDNA) from the nematode to increase the amount of DNA fragments available for analysis using Taq DNA Polymerase (Thermo Fisher Scientific, Waltham, MA, USA). Gel electrophoresis (Bio-Rad Laboratories, Inc., Hercules, CA, USA) was utilized to analyze the PCR-amplified DNA fragments of the PWN, allowing for the separation and verification of the target DNA based on its size. The gel electrophoresis process validated the presence of the nematode by resolving specific DNA bands that correspond to the ITS region of ribosomal DNA, thus confirming the identification of the PWN [33].

2.3.2. Design of the Slope Product Index (SPI)

PWD disrupts the vascular system of infected trees, impairing water transport and triggering rapid chlorophyll degradation [34]. This physiological response results in a noticeable decline in chlorophyll content and a shift in foliage color from green to reddish-brown as the disease progresses. Given its close association with disease severity, chlorophyll content has become one of the most promising indicators for detecting physiological responses in PWD-infected pines. Therefore, when designing a new index, it is essential to fully deploy the detailed spectral characteristics of the relevant species to optimize the detection of the chlorophyll response. Spectral bands in the VNIR regions were selected for analysis due to their high sensitivity to vegetation biochemical and structural changes under stress conditions. Spectral curve analyses show that the green band (G, center wavelength = 560 nm) exhibits minimal change or a slight increase under vegetation stress, as chlorophyll absorbs relatively little green light. The red band (R, center wavelength = 665 nm) shows a more pronounced increase because chlorophyll strongly absorbs red light for photosynthesis. Thus, less red light is absorbed when chlorophyll content declines, resulting in higher red reflectance. In the red-edge band (Re, wavelength = 783 nm) region, the spongy mesophyll tissues are disrupted under stress conditions, reducing internal scattering and thus lowering NIR reflectance [35]. During the progression of PWD, chlorophyll degradation reduces green reflectance while increasing red reflectance, resulting in a steeper positive slope between the green and red bands (S_GR). Meanwhile, reflectance in the red band increases while near-infrared reflectance decreases slightly, leading to a flatter or reduced slope (S_RRe) between the red and NIR bands, as illustrated in Figure 3.

Therefore, we propose the SPI to simultaneously capture spectral slope changes in both the green-red and red-red edge intervals. The structure of the SPI design is conceptually grounded in an approximate integral framework, where a weighted summation of discrete slope estimates emulates the integration of spectral gradients across selected wavelength regions. Specifically, the SPI consists of two components (S_GR and S_RRe) corresponding to slopes between green-red and red-red edge bands, respectively. These components are computed using a combination of a Gaussian weighting function and a central difference approximation of the first derivative. This design emphasizes spectral transitions that are strongly associated with chlorophyll degradation and physiological damage induced by pine wilt disease (PWD). By integrating slope dynamics in both directions, SPI enhances sensitivity to subtle changes in chlorophyll, even in single-date imagery. Theoretically, as disease severity increases, red band reflectance rises more rapidly than green band reflectance. Consequently, S_GR (the slope from green to red) increases from a negative value toward zero. Once chlorophyll degradation is advanced and red reflectance exceeds green reflectance, S_GR becomes positive. For S_RRe (the slope from red to near-infrared), the concurrent increase in red reflectance and decrease in NIR reflectance causes the slope to decline from a positive value toward zero. As a result, the SPI value transitions from negative to positive, reflecting the spectral signature of chlorophyll degradation. In cases of severe infection, SPI may slightly decrease due to saturation effects. However, a positive SPI already marks a significant spectral shift indicative of advanced physiological stress. This transition serves as a reliable threshold for classification, capturing critical changes before the most extreme damage occurs. The corresponding calculation formulas are as follows:

$${\mathrm{S}}_{\mathrm{G}\mathrm{R}}={\sum }_{\mathrm{i}=2}^{\mathrm{n}-1}({\mathrm{e}}^{-{\displaystyle \frac{{\mathsf{\lambda }}_{\mathrm{i}}-{\mathsf{\lambda }}_{\mathrm{n}/2}}{2{\mathsf{\delta }}^2}}}\times {\displaystyle \frac{{\mathsf{\rho }}_{\mathrm{i}+1}-{\mathsf{\rho }}_{\mathrm{i}-1}}{{\mathsf{\lambda }}_{\mathrm{i}+1}-{\mathsf{\lambda }}_{\mathrm{i}-1}}})$$

(1)

$${\mathrm{S}}_{\mathrm{R}\mathrm{R}\mathrm{e}}={\sum }_{\mathrm{t}=2}^{\mathrm{m}-1}({\mathrm{e}}^{-{\displaystyle \frac{{\mathsf{\lambda }}_{\mathrm{t}}-{\mathsf{\lambda }}_{\mathrm{m}/2}}{2{\mathsf{\delta }}^2}}}\times {\displaystyle \frac{{\mathsf{\rho }}_{\mathrm{t}+1}-{\mathsf{\rho }}_{\mathrm{t}-1}}{{\mathsf{\lambda }}_{\mathrm{t}+1}-{\mathsf{\lambda }}_{\mathrm{t}-1}}})$$

(2)

$$\mathrm{S}\mathrm{P}\mathrm{I}={\mathrm{S}}_{\mathrm{G}\mathrm{R}}\times {\mathrm{S}}_{\mathrm{R}\mathrm{R}\mathrm{e}}$$

(3)

where ρ represents the reflectance of the respective band, n represents the number of spectral bands in the range (from 560 nm to 665 nm), m represents the number of spectral bands in the range (from 665 nm to 783 nm), i represents the index of spectral channels, λ represents the wavelength and δ represents the single-sided bandwidth of the Gaussian function. In this study, δ was set to 1 to balance sensitivity and stability in slope estimation. For multispectral sensors such as Sentinel-2, where the green-red and red-edge intervals contain only a few bands, δ = 1 allows slope computation between adjacent bands, which is critical for capturing any spectral gradient. In the case of satellite hyperspectral sensors, where spectral resolution typically ranges from 5 to 10 nm, δ = 1 corresponds to a local slope calculation over 10–20 nm. Within such narrow spectral ranges, a smaller δ retains more detailed spectral information. This setting minimizes the influence of edge inflection points and enhances the expression of physiologically relevant changes, such as chlorophyll degradation.

2.3.3. Selection of Common Indices for Comparative Analyses

To comprehensively evaluate the performance of the proposed SPI, we compared it against 15 widely used spectral indices in remote sensing, selected for their proven effectiveness in assessing vegetation health, chlorophyll content and stress-related physiological changes, which are closely associated with the degradation of pine trees affected by PWD. Specifically, these include vegetation indices such as the green chlorophyll index (GCI, also well known as CI_green) [36], the vegetation index of the green band (VI_green) [37], and the structure insensitive pigment index (SIPI) [38], all regularly used to assess the chlorophyll content of canopies. The atmospherically resistant vegetation index (ARVI) [39] and the visible atmospherically resistant index (VARI) [37] are designed to enhance the ability to resist atmospheric interference. The difference vegetation index (DVI) [40], the ratio vegetation index (RVI) [41] and the NDVI [42] are proposed to identify and monitor biomass and vegetation health. Furthermore, there are spectral features of the green band (G), red band (R), red edge band (Re), the maximum red edge slope (RES), the SDI and the newly proposed S_GR, S_RRe, SPI, totaling 16 different approaches. The definition of each index is shown in

Table 2. Calculation basis of the 16 selected indices for comparison with the proposed SPI.

Spectral Index	Definition	Vegetation Index	Definition
SGR	slope of the curve from the green band to the red band (see Equation (1))	GCI	${\displaystyle \frac{{\mathsf{\rho }}_{\mathrm{N}\mathrm{i}\mathrm{r}}}{{\mathsf{\rho }}_{\mathrm{G}}}}-1$
SRRe	slope of the curve from the red band to the near-infrared (see Equation (2))	VIgreen	${\displaystyle \frac{{\mathsf{\rho }}_{\mathrm{G}}-{\mathsf{\rho }}_{\mathrm{R}}}{{\mathsf{\rho }}_{\mathrm{G}}+{\mathsf{\rho }}_{\mathrm{R}}}}$
SPI	${\mathrm{S}}_{\mathrm{G}\mathrm{R}}\times {\mathrm{S}}_{\mathrm{R}\mathrm{R}\mathrm{e}}$	SIPI	${\displaystyle \frac{{\mathsf{\rho }}_{\mathrm{N}\mathrm{i}\mathrm{r}}-{\mathsf{\rho }}_{\mathrm{B}}}{{\mathsf{\rho }}_{\mathrm{N}\mathrm{i}\mathrm{r}}-{\mathsf{\rho }}_{\mathrm{R}}}}$
SDI	${\mathrm{S}}_{\mathrm{R}\mathrm{R}\mathrm{e}}-{\mathrm{S}}_{\mathrm{G}\mathrm{R}}$	ARVI	${\displaystyle \frac{{\mathsf{\rho }}_{\mathrm{N}\mathrm{i}\mathrm{r}}-2\times {\mathsf{\rho }}_{\mathrm{R}}+{\mathsf{\rho }}_{\mathrm{B}}}{{\mathsf{\rho }}_{\mathrm{N}\mathrm{i}\mathrm{r}}+2\times {\mathsf{\rho }}_{\mathrm{R}}-{\mathsf{\rho }}_{\mathrm{B}}}}$
G	${\mathsf{\rho }}_{560\mathrm{n}\mathrm{m}}$	VARI	${\displaystyle \frac{{\mathsf{\rho }}_{\mathrm{G}}-{\mathsf{\rho }}_{\mathrm{R}}}{{\mathsf{\rho }}_{\mathrm{G}}+{\mathsf{\rho }}_{\mathrm{R}}-{\mathsf{\rho }}_{\mathrm{B}}}}$
R	${\mathsf{\rho }}_{665\mathrm{n}\mathrm{m}}$	DVI	${\mathsf{\rho }}_{\mathrm{N}\mathrm{i}\mathrm{r}}-{\mathsf{\rho }}_{\mathrm{R}}$
Re	${\mathsf{\rho }}_{783\mathrm{n}\mathrm{m}}$	RVI	${\displaystyle \frac{{\mathsf{\rho }}_{\mathrm{N}\mathrm{i}\mathrm{r}}}{{\mathsf{\rho }}_{\mathrm{R}}}}$
RES	$\mathrm{M}\mathrm{a}\mathrm{x}({\displaystyle \frac{{\mathsf{\rho }}_{\mathrm{i}+1}-{\mathsf{\rho }}_{\mathrm{i}-1}}{{\mathsf{\lambda }}_{\mathrm{i}+1}-{\mathsf{\lambda }}_{\mathrm{i}-1}}})(2\le \mathrm{i}\le \mathrm{n})$	NDVI	${\displaystyle \frac{{\mathsf{\rho }}_{\mathrm{N}\mathrm{i}\mathrm{r}}-{\mathsf{\rho }}_{\mathrm{R}}}{{\mathsf{\rho }}_{\mathrm{N}\mathrm{i}\mathrm{r}}+{\mathsf{\rho }}_{\mathrm{R}}}}$

Note: ρ represents the reflectance of the red band (ρ_R), green band (ρ_G), blue band (ρ_B) and near-infrared band (ρ_Nir), n represents the number of spectral bands in the range from 665 nm to 783 nm, i represents the index of spectral bands, and λ is the wavelength. The newly proposed variables are displayed in bold to distinguish them from conventional indices.

.

2.3.4. Evaluation of Index Effectiveness

To assess the accuracy of the SPI in monitoring PWD compared to traditional indices, we analyzed it quantitatively using a number of methods. As mentioned before, the chlorophyll content in pine trees infected with PWD shows a more pronounced decrease [34]. Therefore, examining the response or correlation of each index to the respective chlorophyll content can serve as a method to assess its precision. However, it is difficult to measure canopy chlorophyll content directly because of the considerable height of pine trees. Studies have shown a strong correlation between the normalized area over reflectance curve (NAOC) and the respective chlorophyll content [43]. For this purpose, GF-5B data were processed using the NAOC index to obtain a relative estimate of the chlorophyll content. These values are used as the true values to assess the feasibility of our proposed SPI index. The equation of the NAOC index is expressed as follows:

$$\mathrm{N}\mathrm{A}\mathrm{O}\mathrm{C}=1-{\displaystyle \frac{{\int }_{\mathrm{a}}^{\mathrm{b}}\mathrm{R}\mathrm{d}\mathsf{\lambda }}{{\mathrm{R}}_{\mathrm{m}\mathrm{a}\mathrm{x}}(\mathrm{b}-\mathrm{a})}}$$

(4)

$$\mathrm{C}\mathrm{h}\mathrm{l}=-3.8868+101.94\times \mathrm{N}\mathrm{A}\mathrm{O}\mathrm{C}$$

(5)

where $\mathrm{R}$ is the reflectance, $\mathsf{\lambda }$ is the wavelength. ${\mathrm{R}}_{\mathrm{m}\mathrm{a}\mathrm{x}}$ is the maximum reflectance in the near-infrared wavelength region. $\mathrm{a}$ and $\mathrm{b}$ correspond to the wavelengths 643 nm and 795 nm, respectively. In this context, ${\mathrm{R}}_{\mathrm{m}\mathrm{a}\mathrm{x}}$ equals the reflectance at wavelength b, denoted as ${\mathrm{R}}_{795}$. Chl is the chlorophyll content.

Traditional correlation methods primarily rely on linear, nonlinear and machine learning models. However, under the influence of complex natural environments, most indices do not exhibit ideal mathematical relationships with the true values [44]. On the other hand, data mining capabilities of machine learning models may introduce unexplainable effects on the data correlations, making it difficult to accurately describe the relationship between the index and the data. Thus, we analyzed the overall correlation concept by dividing the evaluation into three components: trend consistency (TC), magnitude consistency (MC), and sensitivity to change (SC). The TC describes the degree of correlation between the trend of the index and the trend of the true values, i.e., whether the index increases or decreases in response to the corresponding true values. The MC describes the degree of correlation between the changes in the index and the true values, i.e., whether the maximum (or minimum) change in the true values corresponds to the maximum (or minimum) change of the index. The SC refers to the average relative rate of change for each index when the true values experience the same rate of change (e.g., 1%).

For the trend consistency (TC) and magnitude consistency (MC) evaluations, we employed the Spearman rank correlation coefficient to quantify the monotonic relationship between each vegetation index and the corresponding chlorophyll content. The calculation involves two key steps: (1) independently ranking each set of values (e.g., vegetation index and chlorophyll content) in ascending order. (2) applying the Pearson formula to these rank values. This approach measures whether one variable consistently increases or decreases in relation to the other, regardless of the actual data magnitudes. TC reflects the consistency of directional trends (i.e., whether higher index values are associated with higher or lower chlorophyll content), while MC evaluates the consistency in relative magnitude between the index and chlorophyll variations. This rank-based method is robust to outliers and nonlinear effects, making it more reliable for datasets that violate the assumptions of linearity or normal distribution. The TC was calculated using Sentinel-2 imagery acquired on 25 and 30 September 2022, based on the chlorophyll reference values derived from GF-5B data on 24 September and 1 October 2022. For MC calculation, we quantified the differences between chlorophyll content (estimated from two dates of GF-5B data) and the corresponding index values derived from Sentinel-2 images acquired close to these two dates.

Both the TC and MC effectively evaluate the correlation between the direction and magnitude of changes in the index and the chlorophyll content. However, they do not fully capture the degree to which the index responds to changes in the true value, specifically the sensitivity and responsiveness of the index to variations. In other words, these two metrics can reveal the consistency between the changes in the index and the true values, but do not provide a comprehensive reflection of how the index responds to changes in the true values. To address this limitation, we used the SC to quantify the responsiveness of the index with changes in the true values by calculating the average relative rate of change. Compared to the TC and MC, the average relative rate of change offers a more specific characterization of the index’s sensitivity to variations in the true values. Therefore, we calculated the relative average rate of change for each index corresponding to a 1% change in chlorophyll content as the quantitative representation of sensitivity to change. The formula is as follows:

$${\mathrm{R}}_{(\mathrm{i}\mathrm{n}\mathrm{d}\mathrm{i}\mathrm{c}\mathrm{a}\mathrm{t}\mathrm{o}\mathrm{r},\mathrm{t}\mathrm{r}\mathrm{u}\mathrm{t}\mathrm{h})}={\displaystyle \frac{{\mathrm{y}}_{\mathrm{f}\mathrm{i}\mathrm{n}\mathrm{a}\mathrm{l}(\mathrm{i}\mathrm{n}\mathrm{d}\mathrm{i}\mathrm{c}\mathrm{a}\mathrm{t}\mathrm{o}\mathrm{r},\mathrm{t}\mathrm{r}\mathrm{u}\mathrm{t}\mathrm{h})}-{\mathrm{y}}_{\mathrm{i}\mathrm{n}\mathrm{i}\mathrm{t}\mathrm{i}\mathrm{a}\mathrm{l}(\mathrm{i}\mathrm{n}\mathrm{d}\mathrm{i}\mathrm{c}\mathrm{a}\mathrm{t}\mathrm{o}\mathrm{r},\mathrm{t}\mathrm{r}\mathrm{u}\mathrm{t}\mathrm{h})}}{{\mathrm{y}}_{\mathrm{i}\mathrm{n}\mathrm{i}\mathrm{t}\mathrm{i}\mathrm{a}\mathrm{l}(\mathrm{i}\mathrm{n}\mathrm{d}\mathrm{i}\mathrm{c}\mathrm{a}\mathrm{t}\mathrm{o}\mathrm{r},\mathrm{t}\mathrm{r}\mathrm{u}\mathrm{t}\mathrm{h})}}}\times 100\%$$

(6)

$$Ratio={\displaystyle \frac{{\sum }_{i=1}^n|\frac{R_{indicator}}{R_{truth}}|}{n}}$$

(7)

where ${\mathrm{R}}_{(\mathrm{i}\mathrm{n}\mathrm{d}\mathrm{i}\mathrm{c}\mathrm{a}\mathrm{t}\mathrm{o}\mathrm{r},\mathrm{t}\mathrm{r}\mathrm{u}\mathrm{t}\mathrm{h})}$ represents the relative rate of change between the index and the true value, ${\mathrm{y}}_{\mathrm{i}\mathrm{n}\mathrm{i}\mathrm{t}\mathrm{i}\mathrm{a}\mathrm{l}}$ represents the data value at the start time and ${\mathrm{y}}_{\mathrm{f}\mathrm{i}\mathrm{n}\mathrm{a}\mathrm{l}}$ represents the data value at the end time. $\mathrm{R}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}$ indicates the average relative rate of change for each index corresponding to a 1% change in the true value, and $\mathrm{n}$ represents the number of samples.

To further integrate the three evaluation metrics, we employed the entropy weight method to compute a composite performance score for each spectral index. The entropy weight method is a widely used objective weighting technique in multi-criteria decision analysis, particularly suitable when no subjective or expert-defined weighting is available. It assigns weights based on the degree of variability in each metric across all indices. Metrics with greater variability (i.e., lower entropy) are considered to carry more information and are thus assigned higher weights, while metrics with lower variability contribute less. This data-driven approach ensures that the final composite score reflects the informational value of each metric rather than assuming equal importance by default. By avoiding arbitrary weight assignment, the method provides a more objective and balanced aggregation, enabling a comprehensive and multidimensional comparison of index performance. The formula for calculating the entropy weights is as follows:

$${\mathrm{x}}_{\mathrm{i}\mathrm{j}}^{\prime }={\displaystyle \frac{{\mathrm{x}}_{\mathrm{i}\mathrm{j}}-\mathrm{m}\mathrm{i}\mathrm{n}({\mathrm{x}}_{\mathrm{j}})}{\mathrm{m}\mathrm{a}\mathrm{x}({\mathrm{x}}_{\mathrm{j}})-\mathrm{m}\mathrm{i}\mathrm{n}({\mathrm{x}}_{\mathrm{j}})}}$$

(8)

$${\mathrm{p}}_{\mathrm{i}\mathrm{j}}={\displaystyle \frac{{\mathrm{x}}_{\mathrm{i}\mathrm{j}}^{\prime }}{{\sum }_{\mathrm{i}=1}^{\mathrm{m}}{\mathrm{x}}_{\mathrm{i}\mathrm{j}}^{\prime }}}$$

(9)

$${\mathrm{H}}_{\mathrm{j}}=-{\displaystyle \frac{1}{\mathrm{l}\mathrm{n}\left(\mathrm{m}\right)}}{\sum }_{\mathrm{i}=1}^{\mathrm{m}}{\mathrm{p}}_{\mathrm{i}\mathrm{j}}\mathrm{l}\mathrm{n}\left({\mathrm{p}}_{\mathrm{i}\mathrm{j}}\right)$$

(10)

$${\mathrm{w}}_{\mathrm{j}}={\displaystyle \frac{1-{\mathrm{H}}_{\mathrm{j}}}{{\sum }_{\mathrm{j}=1}^{\mathrm{n}}(1-{\mathrm{H}}_{\mathrm{j}})}}$$

(11)

where ${\mathrm{x}}_{\mathrm{i}\mathrm{j}}$ represents the original data matrix, $\mathrm{i}$ is the sample index and $\mathrm{j}$ is the metric index. ${\mathrm{x}}_{\mathrm{i}\mathrm{j}}^{\prime }$ represents the standardized data matrix, ${\mathrm{p}}_{\mathrm{i}\mathrm{j}}$ represents the proportion of the ith sample in the jth metric. $\mathrm{m}$ and $\mathrm{n}$ represent the total number of samples and the total number of metrics, respectively. ${\mathrm{H}}_{\mathrm{j}}$ is the entropy value of the jth metric, and ${\mathrm{w}}_{\mathrm{j}}$ is the weight of the jth metric.

2.3.5. Machine Learning-Based Classification Models

The above analysis is based on the correlation between each index and chlorophyll content, which provides a solid mechanistic explanation for the response relationship between the index and the true value changes. To assess the contribution of each index in machine learning, we conducted the Shapley additive explanations (SHAP) analysis method. It is a technique for explaining machine learning model predictions, particularly suitable for high-dimensional and complex models such as deep learning and ensemble methods. SHAP combines the Shapley values from game theory (hereafter referred to as SHAP) to provide a method for measuring the contribution of each feature to the model’s output. By quantifying the importance of each feature as a relative value, SHAP helps to understand the decision-making process of the model. SHAP also provides both global and local explanations. Global explanations summarize the SHAP values across all samples to help identify the most important features. Local explanations allow us to analyze individual predictions and how specific features drive them.

Given a prediction model $\mathrm{f}$, the SHAP value represents the contribution of a feature to the model’s predicted result. Its calculation is based on the average contribution difference of the model across all possible feature combinations. The formula is as follows:

$${\mathsf{\varphi }}_{\mathrm{i}}(\mathrm{f})=\sum _{\mathrm{S}\subseteq \mathrm{N}\setminus \left(\mathrm{i}\right)}{\displaystyle \frac{\left|\mathrm{S}\right|!\left(\mathrm{n}-\left|\mathrm{S}\right|-1\right)!}{\mathrm{n}!}}[\mathrm{f}(\mathrm{S}\cup (\mathrm{i}))-\mathrm{f}(\mathrm{S})]$$

(12)

where ${\mathsf{\varphi }}_{\mathrm{i}}(\mathrm{f})$ represents the contribution of feature, $\mathrm{i}$ to the model output $\mathrm{f}$, i.e., the SHAP value. $\mathrm{S}$ is the set of feature subsets, $\left|\mathrm{S}\right|$ represents the number of elements in the subset, $\mathrm{N}$ is the set of all features, and $\mathrm{N}\setminus (\mathrm{i})$ represents the set obtained by removing feature $\mathrm{i}$ from $\mathrm{N}$. $S\subseteq N\setminus (i)$ represents the subset of features excluding feature $\mathrm{i}$, and $\mathrm{n}$ is the total number of features. $\mathrm{f}(\mathrm{S})$ and $\mathrm{f}(\mathrm{S}\cup (\mathrm{i}))$ represent the model outputs when predicting using only the subset $\mathrm{S}$ and when predicting with the subset $\mathrm{S}$ augmented by feature $\mathrm{i}$, respectively.

• Random forest (RF)

Since the PWD can manifest in various ways depending on the respective environmental conditions, it is essential to use a method capable of capturing these complexities. The RF model comprises multiple decision trees, each trained on a random subset of training samples, and the classification of samples is determined by the average or majority voting of the results from these decision trees [45]. To ensure that the RF model achieves a balance between computational efficiency, classification accuracy, and generalization capability, we optimized its key hyperparameters using a grid search strategy. Specifically, a parameter grid was constructed with the number of decision trees ranging from 50 to 150 (in increments of 10), and the maximum number of leaf nodes ranging from 8 to 40 (in increments of 5). For each parameter combination, three-fold cross-validation was conducted, and model performance was assessed by the average classification accuracy across all folds. As a result, the parameter combination that yielded the highest validation accuracy consisted of 100 decision trees and a maximum of 32 leaf nodes.

2.

Back-propagation neural network (BPNN)

BPNN can recognize intricate patterns in the spectral data, even if those patterns are subtle or hard to detect with traditional methods. A BPNN includes an input, hidden, and output layer. The process begins with forward propagation using training samples, where the error between the output layer and the actual values is calculated. Then, the error is back-propagated to the hidden and input layers. Gradients of the error with respect to the weights are computed and then updated in the opposite direction of the gradients to minimize the error [46]. To ensure a balance between model complexity and generalization, we adopted a grid search strategy to optimize key hyperparameters, including the number of hidden layer neurons and learning rate. Given the relatively small training sample size, we employed a simple neural architecture with only one hidden layer to reduce the risk of overfitting while retaining sufficient representational power. The number of neurons in the hidden layer was tested from 4 to 16 with a step size of 2 and a learning rate of 0.001, 0.005, and 0.01. Each parameter combination was evaluated using three-fold cross-validation, and the optimal configuration was selected based on the highest average validation accuracy across folds. The final model employed eight neurons in the hidden layer and a learning rate of 0.01. The activation function of the hidden layer was set to tansig, while the output layer used softmax for classification. To further prevent overfitting, an early stopping mechanism was applied. The training was terminated if the validation loss failed to decrease for 20 consecutive epochs. The maximum number of training epochs was set to 1200, with a goal error of 1 × 10⁻⁶, and mini-batch training was used for improved convergence efficiency. This configuration provided a robust and accurate model that effectively generalizes across datasets, making efficient use of the limited available training data.

2.3.6. Model Performance Assessment

To comprehensively assess the model’s classification performance, we employed accuracy, F1-score, and the Kappa coefficient as evaluation metrics. Accuracy is a straightforward indicator representing the percentage of correctly classified samples out of the total dataset. The F1-score, particularly useful for imbalanced datasets or when the costs of false positives and false negatives differ, is the harmonic mean of precision and recall, thus balancing these two error types. The Kappa coefficient measures the agreement between predicted and true classifications beyond chance, providing a more robust evaluation than accuracy alone. These metrics were calculated as follows:

$$\mathrm{A}\mathrm{c}\mathrm{c}\mathrm{u}\mathrm{r}\mathrm{a}\mathrm{c}\mathrm{y}={\displaystyle \frac{\mathrm{T}\mathrm{P}+\mathrm{T}\mathrm{N}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}+\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{N}}}\times 100\%$$

(13)

$$\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}={\displaystyle \frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}}}$$

(14)

$$\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}={\displaystyle \frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N}}}$$

(15)

$$\mathrm{F}1={\displaystyle \frac{2·\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}·\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}}{\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}+\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}}}$$

(16)

$${\mathrm{P}}_{\mathrm{e}}={\displaystyle \frac{\left(\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}\right)\left(\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N}\right)+(\mathrm{F}\mathrm{N}+\mathrm{T}\mathrm{N})(\mathrm{F}\mathrm{P}+\mathrm{T}\mathrm{N})}{{\left(\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}+\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{N}\right)}^2}}$$

(17)

$$\mathrm{K}\mathrm{a}\mathrm{p}\mathrm{p}\mathrm{a}={\displaystyle \frac{{\mathrm{P}}_0-{\mathrm{P}}_{\mathrm{e}}}{1-{\mathrm{P}}_{\mathrm{e}}}}$$

(18)

where $\mathrm{T}\mathrm{P}$ represents the number of true positives (healthy samples correctly classified as healthy), and $\mathrm{F}\mathrm{P}$ represents false positives (infected samples incorrectly classified as healthy). $\mathrm{T}\mathrm{N}$ denotes true negatives (infected samples correctly classified as infected), and $\mathrm{F}\mathrm{N}$ denotes false negatives (healthy samples incorrectly classified as infected). In this context, the positive class refers to healthy trees, while the negative class refers to infected trees. ${\mathrm{P}}_0$ represents the observed agreement, equivalent to overall accuracy.

In addition, five-fold cross-validation was employed in the ablation study of the SPI to enhance the robustness and generalizability of the results. The entire dataset was evenly distributed into five subsets. In each iteration, one subset served as the validation set, while the remaining four were used for model training. This procedure was repeated five times, ensuring that each subset was used once for validation. The final performance metrics were obtained by averaging the results across all folds, thereby mitigating the influence of random variation due to sample partitioning and enabling a more objective assessment of SPI’s independent contribution.

3. Results

3.1. Analysis of Index Correlation and Model Impact

3.1.1. Correlation Analysis Based on Multimetric Weighting

First, we calculated the TC and MC using the Spearman rank correlation coefficient to quantify their consistency with the chlorophyll content values. The results show that the proposed SPI index is responsive to changes in the chlorophyll content and shows a high correlation with it (Figure 4). In the analysis of the TC (Figure 4a) and the MC (Figure 4b), the SPI ranked third (r = −0.86, p < 0.01) and second (r = −0.82, p < 0.01), respectively. These results suggest that the SPI effectively captures data trends and reasonably reflects the magnitude of changes, highlighting its solid predictive potential.

Calculating the SC (Figure 5a), the results indicate that the SPI displays a change rate of 11.241% for every 1% change in chlorophyll content. This is significantly higher than that of other indices, such as the second-ranked index, showing a change rate of only 7.409%. Therewith, the result demonstrates that the SPI has a very high sensitivity and responsiveness to the changes in the chlorophyll content. It thus enables fundamental changes to be reflected more clearly, providing a more dependable reference for subsequent analysis and applications.

The results (Figure 5b) based on the entropy weight method visually highlight the superiority of the SPI index compared to other indices. Regarding the TC and MC, the SPI shows a strong correlation with chlorophyll changes, comparable to that of several widely used traditional indices, indicating its ability to capture both trends and magnitudes of change. Furthermore, the SPI also outperforms other indices in terms of sensitivity to change (SC), making it the most accurate for detecting variations in the chlorophyll content. This advantage strengthens the applicability of the SPI in PWD monitoring, providing a more reliable tool for disease assessment.

3.1.2. The Contribution of Indices to Machine Learning Models

In the previous section, we conducted a series of correlation analyses to explore the relationship between the indices and chlorophyll content, which reflects the condition of an infection with PWD, confirming the SPI’s strong responsiveness to changes in chlorophyll. However, for monitoring PWD, it is also necessary to evaluate the contribution of each index to the overall performance of the classification model.

The SHAP analysis of the training set showed that the SPI contributes the most to the RF and BP models. Figure 6(a1) illustrates the influence of the SPI across all samples in the RF model. Compared to other indices, whose impact is narrower around the central part, the influence of the SPI is more prominent and distributed to high SHAP values. Figure 6(a2) further depicts that the SPI dominates in almost all samples, with its contribution nearly equal to the combined influence of all other indices. This, in turn, indicates that it has a considerable decision-making capability in the RF model, confirming its importance for this model. In contrast, Figure 6(b1,b2) display the SHAP analysis results for the BPNN model. Although the SPI has a high impact, its contribution is more balanced relative to the other indices. This difference in contribution is closely related to the mechanisms of the two models. Specifically, tree-based models such as the RF tend to rely on specific features for splitting decisions, while the BPNN processes multiple features through complex nonlinear transformations. As a result, the BPNN model distributes the influence of each feature more evenly.

3.2. Classification Accuracy of Univariate Machine Learning Algorithms

To further validate this result, we designed univariate machine learning models for the classification of PWD, with each index being modeled individually. The classification performance of these models is represented by the maximum and mean accuracy, as shown in

Table 3. Classification accuracies of univariate machine learning algorithms for two test sites.

Study Area	Lishui, Zhejiang Province				Weihai, Shandong Province
Model	RF		BPNN		RF		BPNN
Accuracy (%)	Max	Mean	Max	Mean	Max	Mean	Max	Mean
SPI	93.22	82.39	96.61	84.07	79.51	74.52	81.92	76.41
SDI	81.36	73.73	79.66	75.32	76.09	72.61	75.11	71.30
SGR	88.13	79.07	91.53	82.02	78.91	73.70	80.09	74.08
SRRe	62.73	56.99	61.11	59.69	58.44	55.13	61.27	56.44
G	69.17	65.24	62.55	60.43	65.81	63.09	61.44	58.49
R	63.22	62.50	53.14	47.92	60.73	59.47	54.08	47.17
Re	72.91	66.41	69.43	65.95	68.44	62.17	68.96	64.33
RES	77.97	68.22	73.05	69.95	49.90	49.72	50.10	49.62
GCI	74.64	61.88	72.88	56.46	59.02	55.62	66.13	53.32
VIgreen	79.66	64.42	79.66	64.81	51.27	50.10	50.91	50.34
SIPI	62.50	58.43	60.16	53.74	58.97	56.44	55.31	51.07
ARVI	74.75	70.59	76.27	64.71	64.30	61.35	64.91	57.45
VARI	68.73	61.24	65.79	62.33	63.22	60.17	60.94	58.43
DVI	77.97	67.54	76.36	68.83	73.02	64.90	72.06	65.23
RVI	63.47	58.29	60.47	55.58	59.41	56.04	59.91	56.48
NDVI	74.96	68.37	66.44	59.09	70.11	63.19	60.01	56.19

Note: 30% of the samples in Lishui, Zhejiang (validation) and 100% of the samples in Weihai, Shandong (testing) were chosen to validate the respective model’s performance. The model with the highest accuracy in both study areas is highlighted in bold. Mean and Max represent the average and maximum classification accuracy after one hundred iterations.

.

The results of the classification accuracy based on univariate models partially confirm our findings above. The developed SPI exhibits the best classification performance in both study areas (Accuracy_{Z_Max} = 96.61%, Accuracy_{Z_Mean} = 84.07%, Accuracy_{S_Max} = 81.92%, Accuracy_{S_Mean} = 76.41%). Despite showing a higher influence in the RF model with a SHAPRF concentrated around 0.2, the SPI performance in the BPNN model is about 2% better. A similar trend is observed for the S_GR index. These differences in classification accuracy between the two models can be attributed to their distinct computational mechanisms. Unlike RF, which is based on decision trees and may face challenges with high-dimensional, nonlinear data, BPNN leverages its multi-layered architecture to automatically extract relevant features, allowing it to better accommodate the complexities of the data. Furthermore, the variation in classification performance across different regions can largely be attributed to regional heterogeneity, such as differences in vegetation type, canopy structure, infection severity, and atmospheric conditions. These factors may affect the spectral and structural expression of PWD symptoms, leading to variability in model generalization and performance across regions. Moreover, S_GR follows the SPI, with maximum classification accuracies of approximately 91% and 78%, illustrating that the slope-based methods play a significant role in this study.

In addition to overall accuracy, model performance was further evaluated using the F1-score and Kappa coefficient, which better capture classification robustness, particularly in identifying infected samples. As presented in

Table 4. Other Classification indicators of univariate machine learning algorithms for the two test sites.

Study Area	Lishui, Zhejiang Province				Weihai, Shandong Province
Model	RF		BPNN		RF		BPNN
Indicators	F1	Kappa	F1	Kappa	F1	Kappa	F1	Kappa
SPI	0.84	0.66	0.85	0.70	0.75	0.49	0.76	0.53
SDI	0.78	0.49	0.75	0.49	0.73	0.45	0.72	0.43
SGR	0.79	0.59	0.83	0.62	0.74	0.47	0.73	0.48
SRRe	0.60	0.16	0.61	0.19	0.54	0.10	0.56	0.13
G	0.64	0.29	0.62	0.22	0.63	0.26	0.56	0.17
R	0.61	0.26	0.49	−0.05	0.59	0.19	0.45	−0.05
Re	0.66	0.32	0.69	0.32	0.62	0.24	0.95	0.29
RES	0.68	0.36	0.70	0.39	0.48	−0.01	0.50	−0.01
GCI	0.59	0.25	0.55	0.12	0.55	0.11	0.53	0.07
VIgreen	0.66	0.29	0.62	0.29	0.51	0.01	0.50	0.00
SIPI	0.58	0.15	0.56	0.09	0.55	0.13	0.52	0.02
ARVI	0.73	0.43	0.66	0.29	0.63	0.23	0.59	0.15
VARI	0.62	0.22	0.61	0.26	0.60	0.20	0.57	0.17
DVI	0.71	0.36	0.68	0.39	0.66	0.30	0.65	0.31
RVI	0.56	0.15	0.55	0.12	0.56	0.12	0.56	0.13
NDVI	0.67	0.36	0.57	0.19	0.64	0.27	0.55	0.12

Note that the highest F1-scores and Kappa values are highlighted in bold for clarity.

, the proposed SPI index consistently yielded the highest F1-scores (0.84–0.85 in Lishui; 0.75–0.76 in Weihai) and Kappa values (0.66–0.70 in Lishui; 0.49–0.53 in Weihai) across both the RF and BPNN models. These metrics exceed widely accepted benchmarks (F1 > 0.75; Kappa > 0.6), indicating that SPI maintains a strong balance between precision and recall and achieves substantial agreement with ground-truth labels.

In contrast, other indices such as S_RRe, R, and SIPI yielded considerably lower F1 scores and Kappa values. These values fell below 0.6 and occasionally approached zero, highlighting their limited ability to discriminate between healthy and infected samples, despite achieving a moderate overall accuracy. This discrepancy confirms that accuracy alone may overstate the reliability of classification. Overall, the F1-score reflects the model’s sensitivity in detecting infected samples, balancing precision and recall, particularly in minimizing false negatives. The Kappa coefficient provides a measure of agreement that accounts for chance, offering a more reliable evaluation than accuracy alone. Together, these metrics provide a more comprehensive and objective assessment of model performance, supporting the superior effectiveness of the SPI index in identifying PWD.

3.3. Classification Accuracy of Multivariate Machine Learning Algorithms

In practical applications, investigated features are often not independent of each other. Compared to univariate models, which emphasize the individual performance of each index and reflect its classification ability, multivariate models can simultaneously consider the relationships between multiple features. These models are capable of utilizing the information in the data more comprehensively and account for potential interactions between features. This can further enhance the classification performance of a model. Therefore, we designed a multivariate machine learning model that incorporates all the indices to validate the best performance of the two models. The recorded results are shown in

Table 5. Classification accuracies of multivariate machine learning models including the SPI.

Accuracy (%)	Lishui, Zhejiang		Weihai, Shandong
	RF	BPNN	RF	BPNN
Max	96.61	93.22	82.15	80.73
Mean	88.14	84.75	78.45	72.82

Note that the highest max and mean accuracy are highlighted in bold for clarity.

. In addition, we also designed multivariate models that exclude the SPI, aiming to make its impact more concrete. The results are shown in

Table 6. Classification accuracies of multivariate machine learning models excluding the SPI.

Accuracy (%)	Lishui, Zhejiang		Weihai, Shandong
	RF	BPNN	RF	BPNN
Max	83.26	76.34	78.44	75.49
Mean	74.91	68.23	69.73	62.17

Note that the highest max and mean accuracy are highlighted in bold for clarity.

.

The classification accuracies of the multivariate models above indicate that the RF model exhibits an excellent performance and generalization ability, with the highest classification accuracy (accuracy_Max = 96.61%, accuracy_Mean = 88.14%). A comparison of the results shows a slight decrease in cross-validation accuracy in different areas, but the overall accuracy and model generalization remain acceptable. Compared to the univariate models, the accuracy of the RF model increased by approximately 4% (84.07% to 88.14%), while the accuracy of the BPNN model did not show a significant improvement and even declined in some parts of the Weihai test site. That aligns with findings from other PWD studies [47]. When the SPI is removed from the model, we observe a significant drop in classification accuracy (approximately 10%) (Table 6). In this scenario, the performance of the multivariate model is even worse than that of the univariate model based solely on the SPI, highlighting the advantage of the SPI index in monitoring PWD.

The results of the five-fold cross-validation are presented in

Table 7. Five-fold cross-validation classification results.

	Including SPI						Excluding SPI
Model	RF			BPNN			RF			BPNN
Indicators	OA	F1	Kappa	OA	F1	Kappa	OA	F1	Kappa	OA	F1	Kappa
Fold1	85.50	0.86	0.71	76.08	0.74	0.52	71.73	0.74	0.44	63.76	0.61	0.27
Fold2	80.43	0.79	0.61	73.91	0.74	0.48	68.84	0.67	0.38	59.42	0.60	0.19
Fold3	81.88	0.82	0.64	77.53	0.76	0.55	68.11	0.70	0.37	63.76	0.68	0.32
Fold4	86.23	0.86	0.72	80.43	0.81	0.61	71.01	0.69	0.42	61.59	0.63	0.23
Fold5	82.61	0.81	0.65	75.36	0.76	0.51	65.21	0.68	0.31	60.86	0.63	0.22
Mean	83.33	0.83	0.67	76.67	0.76	0.53	68.98	0.67	0.38	61.88	0.63	0.25

Note: In this experiment, samples from both regions were combined into a single dataset, with each fold containing 20% of the samples (n = 138). OA denotes overall accuracy. The performance metrics from the ablation experiments, with and without SPI, were evaluated using a two-tailed paired t-test, yielding p-values less than 0.01, indicating statistically significant differences.

. Incorporating the SPI index led to substantial improvements in classification performance across all folds. For the RF model, the average Overall Accuracy (OA), F1-score, and Kappa coefficient increased by 14.35%, 0.16, and 0.29, respectively. For the BPNN model, the corresponding improvements were 14.79%, 0.13, and 0.28. In contrast, models that excluded SPI consistently showed inferior performance and, in some cases, performed worse than the univariate SPI model. Paired two-tailed t-tests conducted on the cross-validation results confirmed that the performance improvements attributed to SPI were statistically significant (p < 0.01), suggesting that the observed gains are unlikely to result from random variation. These findings underscore the critical role of SPI in enhancing model sensitivity, robustness, and overall consistency in PWD detection.

In the training set, SHAP analysis primarily focuses on evaluating the influence of each feature during the model learning process. In the test and validation sets, however, SHAP analysis shifts its focus to assessing the model’s performance on unseen data. This shows what role the indices play when they encounter new samples and thereby reflects the model’s generalization capability (Figure 7). The results indicate that, similar to the findings described in Section 3.1 for the training set, our proposed index, the SPI, continues to play a crucial role in the test and validation sets. Its top ranking proves its excellent generalization ability and applicability to different regions. The SHAP value distributions are generally consistent with those in the training set. In the RF model, the SPI occupies a dominant decision-making role, while in the BPNN model, the SHAP values exhibit a more balanced distribution, reflecting its stability.

The analysis of the SHAP plots depicts that, in the RF model, the inclusion of additional indices provides more splitting node options, thereby supporting the decision-making process of the SPI. In the BPNN model, the neural network’s transformation properties balance the influence of each index, which somewhat diminishes SPI’s strong decision-making ability. Other indices fail to provide better decision-making capabilities, leading to a slight decrease in the overall accuracy of the BPNN model. Therefore, the RF model is more suitable for classification tasks with multiple variables because it handles many input features effectively and is less sensitive to irrelevant or redundant variables. However, the BPNN model using the SPI index achieves an accuracy comparable to the multivariate model, with only about a 5% difference. Therefore, we selected the SPI with the BPNN model for PWD monitoring in our third test area in Portugal.

3.4. Inversion of PWD in Portugal Based on the SPI

In Figure 8a, the extracted forest areas generally coincide with the conclusions (Figure 8b) of Alegria [48]. Portugal’s landscape is primarily characterized by maritime pine forests (Figure 8b), with lush vegetation in the northern regions and a comparatively sparse coverage in the south. Figure 8e presents our monitoring results overlaid with the PWD infection areas reported by de la Fuente [30], demonstrating the centralized consistency of the PWD distribution, except for the southern area. After closely examining high-resolution historical images from Google Earth, it is evident that the southern area of Portugal is less densely vegetated, which is also verified by Alegria [48]. In 2016, the PWD was mainly found in Coimbra, Leiria, Aveiro, Viseu, Guarda, and Castelo Branco in west-central Portugal, as well as in Santarem, Setubal, and Beja. Notably, there was an infection-free zone in the central region, mainly concentrated in Santarem. The density map (see Figure 8f) aligns closely with the maritime pine distribution reference image (Figure 8b), highlighting these prevalence areas. The results show a more severe infection with PWD in the northern areas than in the southern regions. Moreover, the high-density areas in the upper-middle region exhibit a distinct clustering pattern, while the overall infected areas are more widely distributed. The infected areas in the south are more scattered with lower densities than in the north, a pattern not identified by de la Fuente [30]. This variation could be explained by the dense vegetation in the northern forests, which may facilitate the rapid spread of the PWD. Meanwhile, in the sparser southern forests, the areas affected by PWD are primarily concentrated near the initial detection sites, likely due to the timing of infection development.

In summary, our PWD monitoring method yielded reliable results and provided high-resolution and detailed information on infections in Portugal. Furthermore, our research model demonstrates specific pan-regional monitoring capabilities and universality, although we could not accomplish detailed fieldwork as we conducted in China.

4. Discussion

4.1. The Ability of the SPI to Distinguish Between PWD and Other Stress Factors

Various biotic and abiotic stressors, such as fungal infections and drought, can also lead to chlorophyll degradation in pine trees, complicating the task of attributing chlorophyll decline exclusively to PWD. However, PWD tends to cause a more rapid and severe reduction in chlorophyll content compared to other stress conditions. Zhang [34] conducted a controlled inoculation experiment in which Masson and black pine trees were subjected to PWN infestation, fungal pathogens, and drought treatments. The results showed that the chlorophyll content dropped more severely after being inoculated with the PWD than when affected by the fungal attack or droughts. Figure 9 was adapted and reproduced from Zhang [34], demonstrating that indices designed for monitoring PWD could account for variations in chlorophyll content to differentiate the PWD from different stress conditions.

However, drought poses a significant challenge in PWD detection, as both conditions are associated with a notable decline in chlorophyll content. To verify whether the SPI can capture this variation, we analyzed the distribution properties of the index under the influence of droughts and PWD. Li [49] illustrated that the Weihai study area experienced severe drought events in 2014 and 2019. In this context, the PWD began to spread significantly in 2015, with a large-scale outbreak in 2019. Therefore, we selected a total of around three hundred samples from October 2014 and 2019 to construct a drought group and a drought + PWD group. Combined with the 2022 PWD group, we used the Kolmogorov–Smirnov (K–S) test [50] to analyze the distribution of the SPI index across these three sample groups. The results are shown in Figure 10.

The results revealed a significant difference in the SPI distribution between the PWD-affected groups and the drought-only group (p < 0.001). This indicates that when PWD occurs, the SPI value differs from drought-only conditions, confirming the index’s ability to distinguish between drought and PWD. Although the K–S test indicates a significant difference between the drought + PWD group and the PWD group (p < 0.05), their data distribution is relatively similar. However, it is worth noting that a small portion of data errors still overlap, making it difficult to distinguish between drought and PWD. This error arises when the later stages of drought coincide with the mid-phase of PWD, leading to similar chlorophyll degradation values. Consequently, future work should focus on collecting more data samples that cover specific PWD occurrence periods to further enhance monitoring accuracy.

Furthermore, we gathered information on several common pine diseases worldwide, including PWD, pine processionary caterpillar (PPC) [51], pine needle scale (PNS) [52], and dothistroma needle blight (DNB) [53]. Analyzing their onset times, symptoms, and progression rates, as shown in

Table 8. Overview of various common pine diseases.

Type	Stressors	Time	Symptom		Speed
			Change Color	Defoliation
Pest	PWD	May to Oct.	reddish-brown	No	Rapid
	PPC	Jan. to Aug.	Yellow	Yes	Moderate
	PNS	May to Sep.	White specks	Yes	Moderate
Fungus	DNB	Late summer	Half-needle scorch, defoliation, especially on lower, interior branches	Yes	Slow
Environment	Drought	Any season	Brown	Yes

.

It can be observed that most diseases occur in late spring and summer, and early autumn when the temperature and a moist environment promote their spread. Based on descriptions of symptoms in the literature, we also found that these diseases are not indistinguishable. Compared to other diseases, the most distinctive feature of the PWD is that the affected leaves (needles) do not fall off but remain hanging on the tree, preserving the canopy’s texture. In contrast, other diseases, especially the PPC, lead to substantial leaf shedding as the pest feeds on the pine needles, accelerating canopy loss. This suggests that late winter, when pests enter dormancy and defoliation becomes prominent, may provide an optimal observation period for distinguishing the PWD from other pine diseases. Different researchers have described leaf discoloration with varying degrees of specificity. The “reddish-brown” color, which is considered a sign of the most severe stress, is only associated with the PWD. This distinctive symptom reflects a substantial decline in chlorophyll content, underscoring the potential of chlorophyll-sensitive indices as effective tools for distinguishing PWD from other stress-induced conditions in pine forests.

Regarding the progression speed, the PWD exhibits the fastest onset. Given the potential timing of disease outbreaks, we hypothesize that, with timely human intervention each year, significant discoloration detectable by remote sensing imagery is highly likely to represent PWD. However, the real situation is more complex. These diseases do not exist in isolation, and pine forests are often simultaneously affected by multiple stressors. In satellite data, this overlapping information could lead to misinterpretations in monitoring results. Therefore, future research should consider a broader range of sample types, incorporating data from multiple stress factors, to more accurately assess the severity of PWD stress and to develop more targeted indices, thus enhancing monitoring accuracy.

4.2. Feasibility of Early Detection of PWD Using Indices That Respond to Chlorophyll Content

As soon as pine trees are infested with PWD, the disease typically develops rapidly and spreads quickly, causing the trees to die within a short period of time. Due to the fast spread of the PWD, early detection is crucial, as it helps identify potential risks and enables timely intervention and control measures. Therefore, we used Sentinel-2 images from early May to late September to calculate and analyze the trend of index changes of vital and infected sample points in satellite imagery. Taking the proposed SPI index as an example, we selected some representative time series from our samples to demonstrate the separability of healthy and infected trees (Figure 11).

From the time series, we can learn that while the index values of healthy trees fluctuate, they generally remain at a low level, which is distinctly different from the behavior of infected trees. The infected trees 1, as shown in Figure 11, exhibits a large and rapid change in the index over a short period of time (from DOY 208 to 218), while the infected trees 2 shows a continuous, slower change over a longer period after a certain time point, and with a smaller amplitude. We noticed from Google Earth that these two patterns correspond to different tree density distributions. In areas with high tree density, the tight spacing accelerates the spread of the PWD, causing a rapid onset of symptoms in many trees simultaneously. This results in a sharp and significant change in the index over a short time period, as seen in the first pattern. Conversely, in areas with sparse tree distribution, the spread of the PWD is slower, but the infection persists for a longer period, leading to gradual, continuous changes in the index over time, corresponding to the second pattern. Additionally, by comparing the initial values of the indices (in May, before, or with minimal infection with PWD), we observed that in the first pattern, the initial index values were lower, suggesting that the pixels contained more trees and the spectral characteristics of the vegetation were more pronounced, further confirming our analysis. For the first pattern, we can directly identify the period when the index shows a significant change and attribute it to the onset of the PWD infection. Using Sentinel-2 satellite data, this change typically takes two overflights, approximately 10 days. For the second pattern, more temporal data might be needed to detect early infection, but since the tree distribution is sparse, the spread of the PWD is slower, providing us with more time to assess and implement responses.

Theoretically, using the trend of index changes for early PWD detection is feasible, especially with change-sensitive indices like the SPI. Unfortunately, we lack sufficient early-stage infection samples to fully validate this approach. Therefore, in future research, we will focus on collecting additional field samples that represent the early stages of PWD infection, which are currently underrepresented in our dataset. These early-stage samples are essential for verifying whether variations in vegetation indices, particularly those identified by the SPI index, can effectively indicate the onset of infection before visible symptoms appear. In addition, with the increasing availability of hyperspectral remote sensing data, we plan to explore whether indices derived from higher spectral resolution can provide improved sensitivity and specificity in detecting early physiological changes associated with PWD. Such data may enhance the compatibility and overall performance of our index-based method, especially in identifying subtle stress signals during the initial stages of disease development. These efforts will contribute to the refinement of remote sensing-based early warning systems for monitoring forest health.

4.3. Issues That Need to Be Addressed Further

Although satellite remote sensing offers substantial advantages for large-scale monitoring of PWD, several issues specific to this study warrant further consideration. One challenge arises from the presence of deciduous trees in mixed forests. In principle, deciduous and coniferous trees may exhibit similar spectral characteristics due to overlapping canopy structures and planting densities. As the proportion of deciduous species increases, their spectral signals may mask the actual chlorophyll condition of pine trees, thereby reducing detection accuracy. In this study, although forested areas were extracted during the preprocessing stage, pine trees were not specifically identified during model training. Incorporating species-level classification using higher spatial resolution imagery, which enables the recognition of individual tree crowns, could enhance the accuracy of disease detection in complex forest environments. In addition, GF-5B hyperspectral imagery was used to assist in validating chlorophyll-related indices, while its current configuration reflects a common trade-off in spaceborne hyperspectral sensors between high spectral resolution and limited spatial resolution. This trade-off may constrain its effectiveness in capturing fine-scale forest canopy variations, especially in heterogeneous or mixed-species environments. Despite these limitations, GF-5B still provided valuable spectral information for evaluating chlorophyll-sensitive indices at the landscape level. In future studies, integrating physical radiative transfer models, such as PROSPECT or PROSAIL [54,55], with hyperspectral data may further enhance the reliability of chlorophyll estimation and disease detection in complex forest ecosystems. While the outbreak data used for validation in Portugal is both authoritative and representative due to the large scale and early timing of the events, it may not fully capture recent infection dynamics. In recent years, strengthened regulatory efforts have significantly reduced the occurrence of widespread outbreaks, which limits opportunities for updated field validation. The lack of up-to-date ground-truth samples from international sites constrains the robustness of cross-regional validation. Future research should incorporate worldwide observations from areas experiencing recent PWD emergence to improve the generalizability of the proposed method. These limitations highlight potential uncertainties in our current work and suggest opportunities for future improvement. Addressing these issues could enhance the applicability, generalizability, and overall reliability of SPI-based monitoring in operational forest health assessments.

5. Conclusions

Pine wilt disease (PWD) remains a major global threat to forest ecosystems, posing significant risks to both ecological balance and economic stability. Effective monitoring and timely intervention are critical to limit its spread and mitigate its impacts, especially in the context of climate change. This study introduced the Slope Product Index (SPI) as a novel approach for monitoring PWD based on a single Sentinel-2 image, combining it with machine learning models to enhance predictive accuracy. The SPI index was applied and validated using data from the provinces of Zhejiang and Shandong in China and from the entire country of Portugal.

The results show that SPI outperforms other indices by providing detailed and robust chlorophyll values based on spectral gradients in the VNIR, which are directly associated with PWD infections. Using a BPNN model, the SPI achieved a mean accuracy of 84.07%, slightly below the model that integrates 16 indices (88.14%), but still highly effective. Notably, removing the SPI from the model significantly decreased the accuracy by around 10%, highlighting its critical role. Cross-validation within different geographical regions further demonstrated the model’s feasibility and effectiveness in a diverse environmental context. Regarding the Portugal test area, the model effectively delineated regions likely affected by PWD and revealed regional variations in infection levels.

In conclusion, this study contributes a reliable, satellite-based index and method for the monitoring of PWD that can be applied worldwide to support forest health management. Future research should focus on refining the model with more early-stage data and expanding its application to other affected regions.