Assessing Sirex noctilio Fabricius (Hymenoptera: Siricidae) Damage in Pine Plantations Using Remote Sensing and Predictive Machine Learning Models

Abstract

Early detection and monitoring of invasive forest pests are crucial for effective pest management, particularly in preventing large-scale damage, reducing eradication costs, and improving overall control effectiveness. This study investigates the potential of machine learning models and remote sensing at various spatiotemporal scales to assess forest damage caused by the woodwasp Sirex noctilio in pine plantations. A Random Forest (RF) model was applied to analyze Planetscope satellite images of Sirex-affected areas in Neuquén, Argentina. The model’s results were validated through accuracy analysis and the Kappa method to ensure robustness. Our findings demonstrate that the RF model accurately classified Sirex damage levels, with classification accuracy improving progressively over time (overall accuracy of 87% for five severity categories and 98% for two severity categories). This allowed for a clearer distinction between healthy and Sirex-infested trees, as well as a more refined categorization of damage severity. This study highlights the potential of machine learning models to accurately assess tree health and quantify pest damage in plantation forests, offering valuable tools for large-scale pest monitoring.

1. Introduction

Invasive forest pests are a major problem for plantation forests, drastically affecting the provision of goods and services worldwide [1]. The increasing arrival pressure of non-native species, coupled with changes in the dynamics of established pests, demands considerable management efforts to minimize their impacts on both private owners and government agencies. Among these efforts, early detection and monitoring of pest infestations are essential for effective pest management, particularly in preventing large-scale damage, reducing the costs associated with eradication and containment, and improving overall pest control effectiveness [2].

To assess forest health, a combination of in situ field methods and remote sensing techniques is typically used [3,4]. Field data provide local and specific biological information but can be costly and struggle to capture broad spatial patterns. In contrast, remote sensing (RS) offers cost-effective and efficient options for assessing forest vegetation health based on physical–chemical parameters, covering larger regions and overcoming the challenges of monitoring in remote locations [5]. Remote sensing techniques have proven to be valuable for identifying forest disturbances such as windstorm damage and pest infestations at different spatial scales [3,6,7]. These methods are indispensable for acquiring accurate data on pest distribution, spread, and intensity, facilitating the development of improved management strategies adapted to vulnerable forest areas.

Recent advances in remote sensing technologies, including unmanned aerial vehicles (UAVs) and satellite-based platforms, offer valuable tools for forest pest monitoring. Several studies have explored the potential of UAV-based hyperspectral and multispectral imagery for detecting pest-damaged trees. UAVs are mainly implemented at the stand scale due to the costs of image acquisition and processing, as well as the potential to create a long time series (TS) of images for the same areas of interest, but when monitoring programs need to cover larger areas, satellite-based RS approaches become essential [8]. Processing images by means of Machine Learning (ML) algorithms, including K-nearest neighbors, Random Forest (RF), Decision Trees, Artificial Neural Networks (ANN), and Support Vector machines (SVM), have been successfully employed in forest damage assessments, achieving accuracies ranging from 55% to 90% [6,9,10]. The use of these RS technologies in conjunction with these advanced techniques provides broader spatial and temporal resolutions, which are critical for mapping changes in forest pests over time.

One of the most promising satellite platforms is Planet, which operates a constellation of over 175 standardized CubeSat 3U nanosatellites, capturing high-resolution multispectral images daily. Planet’s satellites offer a spatial resolution of 3–4 m and capture data in red, green, blue, and near-infrared (NIR) spectral bands. These high-resolution images have been used successfully to detect forest disturbances, such as bark beetle infestations, by differentiating healthy trees from those in early stages of infestation [8,11]. The ability to create time series of such data makes satellite-based RS a powerful tool for assessing forest health and facilitating timely decision making in pest management programs. Remote sensing studies conducted by Villaverde (2020) [12] using Landsat and Sentinel platforms showed an upward trend in the values of MSI and NDVI spectral indices over a five-year period. These changes indicate a gradual deterioration of plant health and a decrease in plant vigor. The correlation between these results and the increase in water stress levels, coupled with the absence of adequate silvicultural practices, suggests an increase in the susceptibility of forests to attack by Sirex, a pest that finds optimal development conditions in weakened trees [12].

Time series analysis allows for the early detection of stress or damage to forests. By tracking the temporal evolution of spectral indices, researchers can identify early signs of decline before visible symptoms appear. TS analysis involves processing images acquired monthly to detect changes in the land surface caused by the infestation damage and distinguish them from changes due to other environmental factors, such as the natural behavior of light reflectance during the year. Several studies [7,13,14,15] use TS analysis to assess the impact of specific disturbances, such as insect outbreaks or climate change, on forest ecosystems. By comparing the temporal patterns of spectral indices before and after a disturbance, researchers can quantify the extent of damage and the rate of recovery. To capture the fine-scale spatial details necessary for symptom identification, high-resolution remote sensing data are required. Consequently, researchers have utilized sensors such as IKONOS, QuickBird, RapidEye, WorldView-2, ADS FieldSpec ProFR, and LIDAR to map insect-related forest disturbances [14]. By analyzing TS of vegetation indices or biophysical variables derived from satellite data, we can investigate changes in vegetation dynamics [7].

The invasive woodwasp Sirex noctilio Fabricius (Hymenoptera: Siricidae) is one of the most damaging insect pests affecting planted pine forests worldwide [16,17,18]. Severe outbreaks have been recorded in several invaded regions, with tree mortality rates reaching up to 75% in some areas of southern Argentina [10,18]. The population dynamics of S. noctilio are characterized by eruptive, pulse-like patterns, driven by both density-independent factors—such as the availability of weakened trees due to prolonged droughts—and density-dependent mechanisms [10,17,18]. While populations can remain at low, endemic levels for extended periods, attacking mostly stressed trees, outbreak events can lead to widespread mortality of healthy trees, resulting in significant economic consequences.

In non-native regions, management strategies for S. noctilio aim to reduce tree mortality through a combination of legal, biological, and silvicultural measures. Early detection of infestation outbreaks is critical for facilitating timely decision-making and improving the success of these management strategies [18,19]. While local-scale monitoring methods, such as visual surveys, trap trees, or kairomone intercept traps, can effectively detect the presence of Sirex [19], large-scale monitoring poses significant challenges. For government agencies and large forestry operations, having robust data on the pest’s distribution and damage over large areas is essential for quantifying the problem’s spatio-temporal extent and developing an appropriate intervention protocol.

In this study, we aimed to develop a remote sensing framework to identify tree mortality patterns caused by S. noctilio in pine plantations. Our underlying hypothesis was that spectral profile comparisons across different Sirex damage categories could serve as the basis for developing AI-based predictive machine learning models capable of differentiating between levels of infestation severity from remote sensing images. Specifically, our objectives were to (a) characterize Sirex damage categories based on spectral signatures, (b) apply machine learning algorithms to classify field plots into damage categories, (c) create time series of damage categories using selected indices, and (d) evaluate and validate the classification models across the study area. Given the increasing impact of forest pests worldwide, we hope this study will contribute to the development of new monitoring tools that facilitate the acquisition and comparison of data on pest damage across large regions, ultimately supporting the design and implementation of effective management strategies.

2. Materials and Methods

2.1. Study Areas

We used the regional S. noctilio aerial classification damage database maps, generated in February 2019 and 2023 as part of the official S. noctilio Monitoring and Control Programme led by the General Directorate of Plant Health and Agricultural Emergency, Ministry of Production and Industry, Neuquén Province, Argentina (see open data in: https://produccioneindustria.neuquen.gov.ar/programa-de-control-de-avispa-sirex/, accessed on 24 April 2024). This monitoring was extended over the entire planted area of Neuquén in northern Patagonia, Argentina, covering nearly 75 thousand hectares of both public and private pine plantations forest. Damage classification was carried out through visual assessments of Sirex-affected trees by qualified personnel during aerial helicopter surveys. Each pine plantation was assigned an attack level based on tree mortality ranges (i.e., proportion of Sirex-affected trees) representative of the different stages of Sirex population dynamics [17]. The categories of Sirex damage were classified as follows: no attack (no visible signs of Sirex-affected trees); incipient (affecting less than 5% of trees); mild (affecting 5–10% of trees); moderate (affecting 10–50% of trees); severe (affecting more than 50–70% of trees); and extreme (affecting more than 70% of trees) (see aerial survey categories in Appendix A).

To our study, we selected two Sirex-infested pine plantation areas for field data collection, characterized by similar tree ages and forest management approaches, including only partial early pruning and thinning. Firstly, to train our machine learning algorithm we chose a forest area located in Las Ovejas, a rural town in the northern region of Neuquén Province (hereafter referred to as the Las Ovejas plantation; 36.87°S–70.88°W). This area is planted with 6053 hectares of pure stands of Pinus contorta and Pinus ponderosa, ranging from 28 to 32 years old, with all levels of Sirex damage severity (from no damage to extreme), as identified in the 2019 aerial survey.

Additionally, to generate field validation to apply the ML algorithm selected in the previous step, we chose a second pine plantation area affected by Sirex, located in Junín de Los Andes city, in the southern region of Neuquén Province (hereafter referred to as the Junín plantation; 39.92°S–71.09°W—Figure 1). This area covers nearly 220 hectares, predominantly planted with 28-year-old lodgepole pines (Pinus contorta), with some patches of ponderosa pine (Pinus ponderosa). We generated an orthomosaic of the entire Junín planted area from images captured by a drone equipped with multispectral camera. Using QGIS [20], we generated circular sampling plots of 400 square meters regularly spaced by 100 m, in which we identified and counted in the field all healthy and Sirex-infested trees.

In both Las Ovejas and Junín areas, according to the proportion of Sirex-infested trees, we labeled each sampling plot according to the damage levels to compare with the result obtained from ML algorithm classification. We assume that all pixels within a plot are homogeneous and representative of the assigned label. Within each plot spectral variability is minimal, ensuring accurate classification. For instance, in a plot labeled as no damage, all pixels exhibit spectral values in RGB+NIR bands that are distinctly different from those labeled as severe, incipient, or moderate.

To address potential inaccuracies in visually assignment of intermediate damage categories, we expanded the labeling scheme beyond the original aerial monitoring approach (with five original categories—5LA) by implementing a Combined Labels Approach (CLA). Firstly, to produce classification maps of Sirex damage categories, we applied the Four-Labels Approach (4LA), which included the following labels: no damage, incipient, moderate (a combination of mild and moderate), and severe. Given the absence of areas with severe damage in the 2023 aerial survey data, we adopted a three-group categorization (3LA) that included the labels: no damage, incipient, and moderate (a combination of mild and moderate). Additionally, to generate presence/absence Sirex maps, we employed a Two-Labels Approach (2LA) to differentiate between no damage and affected, encompassing all categories of Sirex damage levels (see all labeling approaches in Appendix A).

2.2. Remote Sensing Data Acquisition and Processing

Remote sensing data were obtained from PlanetScope Ortho Tile products [21] with less than 20% cloud coverage for the Las Ovejas plantation. Images acquired between January 2019 and December 2023 (excluding the months from May to September due to the presence of snow) were used to create an orthomosaic by merging multiple scenes, which were orthorectified into a single strip and divided into a 25 × 25 km grid. All Images were subjected to radiometric, geometric, and sensor corrections. The analytical format consisted of a 4-band, 16-bit multispectral image (RGB-NIR) with an orthorectified spatial resolution of 3.125 m. Daily review was conducted, and 32 spectral indices (SI’s) were calculated for each image to assess vegetation greenness and stress levels (Appendix B). A similar approach was implemented for the Junín plantation, using images acquired in April 2023. Image pre-processing and processing were performed using ENVI (version 5.7) [22], and post-processing was performed with QGIS (version 3.16.3) [20] and R software (version 3.3.1) [23].

2.3. Approach to Land Cover and Forest Density Classification

For land cover and forest density classification, we utilized the approach proposed by Roy et al. (1996) [24], which isolates the vegetation feature space using the Canopy Shadow Index (SI), Bare Soil Index (BI), and Advanced Vegetation Index (AVI). These indices are the most useful for detecting and estimating canopy density, as they reduce the effects of atmospheric disturbances and vegetative influences [25]. The separation of canopy density and bare soil is significant in the color composite of the AVI + SI + BI combination for the orthomosaic. Pixels showing values of 0 or null for some of the components of this combination are eliminated, resulting in pure pixels with vegetation values.

2.3.1. Bare Soil Index

Bare soil index (Equation (1)) enhances the detection of bare soil areas, fallow lands, and vegetation with distinct background responses. Like the AVI, BI is a normalized index that differentiates between reflective (IR and B) and absorption (MIR and R) bands. This index aids in distinguishing vegetation with varying background conditions, including completely bare, sparse canopy, and dense canopy [24].

BI = {(MIR + R) − (IR + B)}/{(MIR + R) + (IR + B)}

(1)

2.3.2. Canopy Shadow Index

Crown arrangement in forest stands influences shadow patterns, affecting spectral responses. Young, even-aged stands have lower canopy shadow indices (Equation (2)) compared to mature natural forest stands. Mature forest stands exhibit a flatter and lower spectral axis compared to open areas. SI is directly proportional to forest canopy density, resulting in higher SI values for denser forests. The vegetation feature space is further stratified based on SI, considering the textural variations introduced by canopy shading in forest stands [25].

SI = [(255 − B) ∗ (255 − G) ∗ (255 − R)]^1/3

(2)

where B is the blue band, G is the green band, and R is the red band

2.3.3. Advanced Vegetation Index

The NDVI has been shown to be sensitive to canopy foliage activity. Anon et al. (1993) suggested that subtle differences can be enhanced by using a power transformation of the infrared response. The resulting index, termed the advanced vegetation index (Equation (3)), has proven to be more sensitive to forest density and physiognomic vegetation classes [13].

AVI = [IR ∗ (255 − R) ∗ (IR − R)]^1/3

(3)

where IR is the infrared band response and R is red band response.

We conducted a pixel-by-pixel classification of the image based on damage labeling generated during the 2019 aerial survey, which was further refined using a computer workstation. In this process, we created 776 circular plots with a radius of 20 m (1.256,64 m²), positioned on the orthomosaic using QGIS to capture representative regions within the study area. To enhance the accuracy of our analysis, we eliminated pixels from each plot that lacked values in any of the combined indices (AVI + SI + BI) (Figure 2). This step was crucial for reducing noise caused by ground refraction in areas exhibiting some degree of damage, ensuring that only pixels representing vegetation cover were retained. Furthermore, to maintain the integrity of our labels, we excluded plots where more than 10% of the pixels contained missing data and reassigned the remaining data as necessary. Since the final pixels were labeled according to plot damage category, a pixel-level model was fitted, providing pixel-specific results.

Mean values of spectral reflectance bands (4) and SI’s (32) for each plot class were extracted using the “extrac” function from the terra package in R.

2.4. Data Analysis

2.4.1. Feature Selection: Variable Importance Selection

The number of variables useful as predict input was reduced using permutation importance (PIMP) proposed by Altmann et al. (2010) [26] for reducing bias, based on a permutation test, returning significant p-values for each feature (Appendix B). PIMP is available in the vita package of R software [23]. The PIMP-derived spectral indices used as input for ML models was reduced using the Jeffries–Matusita (J–M) separability test [7]. Based on IS’s derived from PIMP, the J-M value was calculated for each severity class using the training set. J-M ranges from 0 to 2; values between 0 and 1 corresponds to very poor separability, and values between 1 and 1.5 suggest a poor separability, and high separability is associated with values in the range of 1.5 to 2. [27] allowed for precise selection of spectral bands that best discriminated between different damage classes. Only SI’s with J-M values above a predefined threshold (1.5) were included in the ML models.

2.4.2. Machine Learning Classification Methods

Traditional machine learning algorithms, such as Random Forest, Artificial Neural Networks, and Support Vector Machines, have consistently demonstrated high accuracy in modeling tree-level pest and disease infestations [28,29]. Their ability to automate detection and accurately discriminate subtle spectral changes associated with potential outbreaks makes them invaluable tools for forest insect pest and disease monitoring [28]. These algorithms excel at distinguishing healthy trees from affected ones, even in complex environments [13,30].

Machine learning algorithms were trained with 5LA (2019) in the Las Ovejas plantation, using 80% (n = 620) of a randomly selected holdout sample; final accuracy assessments were determined using the remaining 20% of the data (n = 156). To evaluate ML algorithms performance in estimating damage categories, a comparative analysis was undertaken.

Random Forest

Random Forest is a machine learning algorithm widely employed in forest research due to its robustness in addressing modeling challenges, particularly by mitigating overfitting and reducing model variance [7]. Because of its high accuracy and reliability, the RF classifier has become a key tool in remote sensing applications [29]. Random Forest models were trained using the randomForest package [23] in R software setting the number of trees, ntree, to 500. Internal evaluation through leave-one-out cross-validation assessed the model’s performance across different number of variables to randomly sample as candidates at each split (mtry) values. For each model, the mtry value that minimized the out-of-bag error was selected.

Support Vector Machines

Support Vector Machines are powerful machine learning algorithms adept at both classification and regression. For classification, SVM finds the optimal hyperplane that maximizes the margin between classes, ensuring robust decision boundaries. In regression, SVMs predict continuous values by fitting a function within a specified tolerance zone [13]. Built on the principle of structural risk minimization, SVMs balance model complexity and generalization. By maximizing the margin, they excel in handling high-dimensional data and mitigating overfitting. SVM assume a unique relationship between input features and the target variable, enabling the construction of accurate decision boundaries [13]. They have proven robust to outliers, dimensionality issues, and overfitting [3]. A polynomial function was used as the kernel function; it showed better behavior than lineal and sigmoidal kernel functions. The optimal cost and gamma parameters were tuned using svm and tune.svm functions from the e1071 R package.

Artificial Neural Networks

Artificial Neural Networks offer an approach to model complex, nonlinear phenomena in forest science. Typically composed of input, output, and one or more hidden layers, ANNs mimic human learning by establishing and strengthening connections between input and output data [3,31]. These connections enable data linkage without explicit programming. While various ANN architectures exist, including Radial Basis Function, Elman Recurrent, and Hopfield Neural Networks, Multilayer Perceptron Neural Networks (MLPNNs) with backpropagation have gained widespread popularity [31]. MLPNNs consist of interconnected layers of neurons, transforming input data into meaningful outputs. In this study, the MLPNN architecture comprised 35 input neurons, corresponding to the predictor variables by ‘monmlp” “neuralnet” R packages. The optimal number of hidden neurons was determined through training and validation, with a three-neuron configuration yielding the lowest error.

2.4.3. External Classification Accuracy Analysis

For external validation, the selected ML model was applied on a pixel-by-pixel basis within the Junín plantation. Model predictions were compared with a field assessment conducted by experts. We use the Kappa, overall accuracy, and user accuracy to measure the agreement between ML model predictions and field data. Kappa is a statistical measure used to assess the agreement between two raters or classifications, quantifying the extent to which classifications agree beyond what would be expected by chance [23]. A Kappa value of 1 indicates perfect agreement, while a value of 0 suggests that any agreement is due to chance alone; a negative Kappa value indicates worse-than-chance agreement. When evaluating classification models such as RF, SVM, and ANN, a comprehensive assessment requires going beyond overall accuracy [23]. While overall accuracy is a widely used metric that quantifies the proportion of correct predictions made by the model compared to the total number of observations, it provides only a basic understanding of model performance. User accuracy measures the reliability of the model’s class assignments, where a high user accuracy indicates that when the model assigns an observation to a class, it is typically correct. In contrast, producer accuracy assesses the model’s ability to identify all instances of a given class, with a high producer accuracy suggesting effective detection of most instances. By considering both user and producer accuracy, a more complete idea of the model’s performance can be obtained [23].

2.5. Temporal Variation in Spectral Indices

Changes in vegetation phenology are a key aspect of environmental change due to their sensitivity to climate change and significant impacts on ecosystem functions. Satellite-derived time series of vegetation indices can capture the seasonal dynamics of leaf development and are essential for spatially continuous observations of vegetation phenology, such as in forest plantations [15,32]. A time series analysis of spectral indices was conducted to identify and quantify changes in vegetation health over time induced by Sirex damage. Monthly image acquisitions were used to derive these indices, capturing variations in spectral reflectance across different months. To account for natural seasonal fluctuations in reflectance, seasonal adjustments and normalization techniques were applied, ensuring that the observed changes in spectral indices were primarily attributed to Sirex damage rather than seasonal effects [11]. Radiometric calibration is often essential for long-term image time series to account for sensor degradation. In this study, the short temporal extent minimized the impact of potential radiometric drift, rendering calibration unnecessary [11,15]. Despite variable image availability due to weather and satellite coverage, a fixed number of images was used for each smoothing. Snowy months with missing data were excluded.

3. Results

We generated and analyzed a comprehensive dataset of 35 composite orthomosaics collected annually from 2019 to 2023 in the Las Ovejas plantation (seven per year). Using a ML algorithm, we classified tree sample areas based on Sirex damage levels (i.e., proportion of infected trees) as no damage (without detectable Sirex damage) or Sirex-infested (with different damage levels), based on 32 spectral indices (Appendix B), and we developed a time-series of damage categories using selected indices. Additionally, we evaluated and validated the selected classification models in a new study area (Junín plantation) to assess their generalization and predictive accuracy.

3.1. Variable Importance Selection

According to importance inference and cross-validated permutation importance, the spectral indices revealed distinct patterns in their importance for classification tasks across the labeling approaches of Five-Labels, Four-Labels, Three-Labels, and Two-Labels Sirex-damage categories between the years 2019 and 2023. To confirm and remove non-informative feature selection, we employed the J-M separability index. For each image, we calculated the J-M value for all possible combinations of damage classes. By averaging these values, we obtained a single overall J-M value for each SI. This metric reflects the SI’s ability to differentiate among the three damage classes [7]. SI’s with J-M values exceeding 1.5 were selected as input features for the ML models.

In February 2019, the most influential indices for classifying Sirex damage severity classes were VARI, TGI, NDWI, GNDVI, and the green and NIR bands. Conversely, in February 2023, the indices shifted to VARI, TGI, NDWI, GNDVI, HUE, and the green and red bands.

3.2. Seleted Machine Learning Algorithms

Our results from the confusion matrices for the RF, SVM, and ANN indicate that classification accuracy varies across labeling approaches. A comparison of the overall accuracy for 5LA of the three artificial intelligence algorithms, RF, SVM, and ANN, yielded accuracy scores for the 2023 dataset of 88%, 70%, and 78%, and for the 2019 dataset, the scores were 57%, 47%, and 48%, respectively. In the 5LA (for RF, SVM, and ANN classification), we found that in 2023, the mild damage category exhibits the lowest accuracy compared to other categories, primarily due to the misclassification of moderate and mild cases in 2019. Notably, the no damage category achieves both user and producer accuracies exceeding 92%, 90%, and 89% for RF, SVM, and ANN, respectively. Given its superior performance, the Random Forest algorithm was selected to classify S. noctilio damage levels in this study. The confusion matrix detailing the classification results for 5LA is presented in

Table 1. Five-Labels approach RF-based classification results of 2023 and 2019.

2023		Incipient	Mild	Moderate	Severe	No Damage	User’s Accuracy	Overall Accuracy	Kappa
	Incipient	11,239	324	787	0	340	0.89	87.9%	0.79
	Mild	185	680	104	0	6	0.70
	Moderate	655	96	4867	0	23	0.86
	Severe	0	0	0	0	0	0
	No damage	28	3	7	0	1778	0.98
	Producer’s accuracy	0.93	0.6	0.84	0	0.83
2019		Incipient	Mild	Moderate	Severe	No Damage	User’s Accuracy	Overall Accuracy	Kappa
	Incipient	1803	1004	35	549	16	0.53
	Mild	1869	6110	465	2517	48	0.56
	Moderate	9	201	328	104	2	0.51	57%	0.42
	Severe	393	1573	418	2706	15	0.53
	No damage	0	3	0	0	931	1.00
	Producer’s accuracy	0.44	0.69	0.26	0.46	0.92

.

In the Four- and Three-Labels classification, the incipient category maintains consistent accuracy when comparing the 5LA across both years (Table 1). Additionally, there is a slight improvement in the no damage category’s accuracy, along with a substantial increase in overall model accuracy, indicating that the RF model has effectively enhanced its classification capabilities over time (

Table 2. Three- and Four-Labels approach RF-based classification results of 2023 and 2019.

2023		Moderate	Incipient	No Damage	User’s Accuracy	Overall Accuracy		Kappa
	Moderate	5946	898	28	0.87	89%		0.81
	Incipient	913	11181	358	0.90
	No damage	9	28	1761	0.98
	Producer’s accuracy	0.87	0.92	0.82
2019		Moderate	Incipient	No Damage	Severe	User’s Accuracy	Overall Accuracy	Kappa
	Moderate	7381	1992	31	2976	0.60	59%	0.45
	Incipient	1139	1556	2	608	0.47
	No damage	0	0	972	0	1.00
	Severe	1727	326	7	2382	0.54
	Producer’s accuracy	0.72	0.40	0.96	0.40

). Lastly, the Two-Labels (2LA) classification shows a marked improvement in accuracy across all categories compared to the previous classification approach (

Table 3. Two-Labels approach RF-based classification results of 2023 and 2019.

2023		Affected	No Damage	User’s Accuracy	Overall Accuracy	Kappa
	Affected	18,950	404	0.98	98%	0.89
	No damage	26	1743	0.99
	Producer’s accuracy	0.99	0.81
2019		Affected	No Damage	User’s Accuracy	Overall Accuracy	Kappa
	Affected	20,088	9	1.00	99%	0.99
	No damage	0	1003	1.00
	Producer’s accuracy	1.00	0.99

), with user accuracy rising from 87.8% to 98% in 2023 and from 57% to 99% in 2019, underscoring the improved performance of the Random Forest model in distinguishing between Sirex-affected and non-affected trees.

3.3. Random Forest Interpolation

The results obtained from the interpolation process using the selected Random Forest models indicate effective estimations of category values for every pixel within the study zone. The spatial pattern maps generated for the 2019 data (Figure 3) reveal clear delineations in the Five-Labels (5LA), Two-Labels, and Four-Labels classification approaches. In the assessment of interannual predictions, applying the Random Forest model trained on the 2019 labels to the 2023 plot raster dataset demonstrated reliable classification, particularly with the selected 4LA and 2LA classifiers (Figure 4). This approach enabled the classification into a severe category, which was absent in the 2023 ground truth, resulting in a high percentage of agreement between the predicted and observed field values. Figure 4 illustrates the spatial pattern maps generated by the selected Random Forest models for 2023 data, utilizing 5LA, 3LA, and 4LA (without severe class) classifications.

Table 4. Distribution of pixel classification percentages by category, approach (Five-Labels, Four-Labels, Three-Labels and Two-Labels approaches: 5LA, 4LA, 3LA, and 2LA, respectively) and year.

Year	Approach	Labels	Percentage
2019	2LA	Affected	99.71
		No damage	0.29
	4LA	Incipient	17.2
		Moderate	62.2
		Severe	20.1
		No damage	0.5
	5LA	Incipient	17.7
		Mild	55.2
		Moderate	3.1
		Severe	23.5
		No damage	0.5
2023	2LA	Affected	99.68
		No damage	0.32
	3LA	Incipient	58.13
		Moderate	41.35
		No damage	0.52
	5LA	Incipient	60.11
		Mild	7.84
		Moderate	31.52
		No damage	0.52

presents the calculated percentage of pixels covered by each category within each label approach. Additionally, the RF models’ predictions were validated in the Junín plantation (n = 140 plots) against expert-assigned labels, further supporting the robustness of the model’s predictions.

3.4. RF Interannual Extrapolation

To assess the quality of interannual predictions, a Random Forest model trained on 2019 labels was applied to the 2023 plot raster dataset (

Table 5. Four-Labels approach RF-based classification results for 2023 based on 2019 RF model.

	Incipient	Mild	Moderate	No Damage	User’s Accuracy
Incipient	2507	51	859	511	63.82
Mild	103	192	13	62	51.89
Moderate	216	29	631	327	52.45
No damage	100	3	145	422	62.99
Producer’s accuracy	85.7	69.8	38.3	31.9

and

Table 6. Two-Labels approach RF-based classification results for 2023 based on 2019 RF model.

	Affected	No Damage	User’s Accuracy
Affected	1714	1052	62.0
No damage	250	341	57.7
Producer’s accuracy	87.3	24.5

). Due to its superior model fit, a Two-Labels approach and a Four-Labels classifier were chosen (4LA enables classification into a severe class, which was absent from the 2023 ground truth). The evaluation was conducted by calculating the percentage of agreement between the predicted and observed field values.

3.5. Extrapolation of RF to Another Field of Study

The validation of the Random Forest 2023 model predictions in the Junín plantation (n = 140 plots) shows a strong correlation with the labels assigned by a field expert. The results indicated a high level of accuracy in the RF models for identifying tree attack caused by S. noctilio in nearby areas (

Table 7. Three-Labels approach RF-based (2023 RF model) classification results of the Junín plantation.

	Mild	Moderate	No Damage	User’s Accuracy
Mild	25	4	1	83.33
Moderate	10	44	0	81.48
No damage	8	12	35	63.64
Producer’s accuracy	58.1	73.3	97.2

).

3.6. Damage Categories Based on Spectral Signatures

The select spectral indices reveal distinct patterns in density distributions across Sirex-damage categories within the study area for February 2019 and February 2023 (Figure 5 and Figure 6). These density curves represent variations in index values across different levels of damage, from no damage to severe. The separation observed between the density curves for each damage level suggests that these indices can effectively differentiate between varying intensities of Sirex damage. However, some overlap between curves indicates potential challenges in discriminating certain damage categories. Additionally, the shape of these density curves reveals patterns in the distribution of index values; for instance, more symmetrical curves with lower variance may indicate a more consistent vegetation response within a specific damage category.

3.7. Temporal Variations in the Selected IS’s

Figure 7 illustrates the temporal trends of 2019 and 2023 spectral indices (VARI, TGI, NDWI, GNDVI, and HUE) and bands (green, red, and NIR) within the study area, categorized differently based on the Four-Labels Approach (incipient, moderate, no damage, and severe). The central tendency (or mean) of an index over time for a specific damage category is presented, and the shaded areas indicate the confidence interval of this trend, that is, the range of values within which the true value is expected to fall with a certain level of probability. Trend lines allow us to visualize how the indices evolve over time for each damage category [11]. This is useful for identifying seasonal patterns, long-term trends, or abrupt changes associated with specific events. The confidence intervals provide a measure of the uncertainty associated with the trend estimates. Narrower confidence intervals indicate a higher precision in the estimate, while wider intervals suggest greater variability in the data. By comparing trends and confidence intervals across different damage categories, we can assess the discriminatory power of the indices used to differentiate between various levels of pest infestation. Although some indices exhibit superior performance, the selection carried out by the RF algorithm ensures their complementarity. Changes in the slope or amplitude of the trend lines may indicate changes in tree health or environmental conditions. In this study, climatic variables that could affect these trends were not considered.

4. Discussion

Early detection and damage monitoring of forest pests are crucial for forest health programs [33,34]. The development of tools that improve the spatial and temporal quantification capabilities of trees affected by pests, with sensitive methods that identify their physiological changes, allows for the obtaining of data to allocate resources and implement targeted control measures more effectively. Furthermore, the applicability of these approaches to the management of plantations over large areas, integrated with other data generated through methods such as aerial monitoring, can provide valuable data in regional surveys for both companies and public forest health agencies. Accurate and timely forest health monitoring is indispensable for mitigating forest degradation and enhancing ecosystem resilience [7]. By detecting and quantifying the extent and severity of pest and pathogen infestations, forest managers can implement targeted interventions to minimize damage and reduce the risk of future outbreaks. Additionally, such monitoring enables the assessment of forest vulnerability and the evaluation of the efficacy of management strategies [7]. Our study demonstrates the effectiveness of integrating remote sensing and machine learning to detect and assess damage caused by S. noctilio in forest plantations, both at stand and large scales and over time. Similarly to other research [11], we identified the most effective spectral bands and spectral indices from Planetscope data to differentiate between various stages of S. noctilio infestation.

SI’s have proven to be valuable tools for monitoring plant health and stress. By mitigating the influence of environmental factors, such as geometry, background, and lighting conditions, SI’s enable accurate and reliable large-scale assessments of vegetation health [3]. This study compared the performance of RF, ANN, and SVM classifiers on equivalent training and test datasets of Sirex damage levels. In their work with Sirex and ML algorithms, Abdel-Rahman et al. (2014) [13] determined that RF and SVM were particularly effective due to their inherent ability to handle multidimensional and noisy hyperspectral data [13,35,36]. Their robustness to outliers and overfitting further enhances their suitability for this classification task [36].

To characterize spectral signatures associated with varying levels of S. noctilio damage in this study, we selected Random Forest classification models. RF demonstrated a higher accuracy in differentiating between no damage, incipient, moderate, and severe damaged categories compared to SVM and ANN. This finding is consistent with the results reported by Abdel-Rahman et al. (2014) [13] and Long et al. (2023) [35] regarding the superior classification performance of RF.The high accuracy of these models, validated across a new plantation area, provides a powerful tool for early detection and monitoring of spatio-temporal patterns of S. noctilio damage at both stand and large-area scales. The results indicate that the 2LA, focusing on no damage/affected tree detection, outperformed 5LA, 3LA, and 4LA, which aimed to classify specific Sirex damage categories. Our 2LA findings, consistent with the reserch of Gao et al. (2022) [3] on the red turpentine beetle, Dendroctonus valens, in Pinus trees, demonstrate the potential of SI’s to distinguish between early-infested and healthy trees (overal accuracy 89–98% and 84%, respectively).The RF model results obtained by [13,14] in their research of Sirex in South African pine plantations with three infection scales were comparable to our finding (2023 overall accuracy 79–81%). The high level of disagreement in 2019 between the 3LA, 4LA, and 5LA, as shown in the confusion matrices, highlights the challenges of accurately classifying pixels that belong to multiple classes, especially when considering spatial patterns in the observed and predicted data. Additionally, the model’s accuracy improved for 2023 data compared to 2019, suggesting better performance in detecting established infestations.

However, our study highlights the need for further refinement of the model to improve the classification of different damage severities. Time series analysis was found to be crucial in differentiating damage-related changes from natural variations [35], and prediction at a regional scale can be realized by using long time series remote sensing data [35]. Host water loss, although not always visually apparent, is a critical indicator of stress; remote sensing studies have revealed significant differences in spectral responses between canopy discoloration, characterized by chlorosis or reddening and insect-induced defoliation [12].

Our analysis revealed that specific spectral indices, such as VARI, GNDVI, TGI, and NDWI, were crucial for discriminating among the different S. noctilio damage categories. These indices serve as vital indicators of vegetation health, allowing us to detect subtle changes that may precede visible symptoms on trees. Spectral indices demonstrated sensitivity to changes in tree mortality caused by S. noctilio, capturing various aspects of plant stress and enabling a more comprehensive damage assessment. Despite the overall trends, the significant variability observed over time in index values can be attributed to fluctuations in S. noctilio population dynamics, climatic factors, and other uncontrolled variables. Recently, Villaverde (2020) [12] identified a significant relationship between Sirex infestation levels and various spectral indices within the same geographic region. By leveraging high-resolution satellite imagery from Landsat 8 and Sentinel-2, the study analyzed correlations between the Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), Modified Soil Index (MSI), and Soil-Adjusted Vegetation Index (SAVI) and the extent of insect-caused damage. According to Villaverde (2020) [12], the combination of different spectral indices provides a more comprehensive and robust understanding of tree stand status, offering deeper insight into the mechanisms underlying Sirex-induced damage. In our study, indices such as GNDVI and HUE are useful for detecting early signs of Sirex damage before symptoms become visible in aerial imagery or field surveys. Additionally, indices like VARI, GNDVI, NIR, and NDWI have proven to be effective indicators of overall tree health, helping to identify areas at higher risk. The decline in VARI values in Sirex-affected areas highlights its utility for monitoring damage progression.

GNDVI, with its strong correlation to Sirex damage, is particularly sensitive to changes in chlorophyll content, making it a valuable tool for assessing vegetation health. While NDWI shows a less direct relationship with Sirex damage, it provides valuable insights into plant water content, helping to identify water stress. The TGI, initially developed to correlate with chlorophyll content at the canopy level, exhibits its highest sensitivity to changes in Leaf Area Index [36]. This index’s strong correlation with chlorophyll content makes it a useful tool for detecting subtle variations in plant health directly related to green foliage losses caused by Sirex. According to Magalhães et al. (2024) [36], as the canopy becomes denser, indices strongly associated with green reflectance, such as VARI and TGI, accurately separate categories with less damage from those with more damage. Finally, indices such as TGI, NIR, green, red, and HUE offer supplementary information on various aspects of Sirex-induced damage, including reduced biomass, changes in pigment composition, and alterations in leaf structure (additional details can be found in

Table 8. Selected Remote Sensing indices for assessing vegetation damage.

Indices/Band	Concise Index Description	Observed Results	Interpretation
VARI = (Green − Red)/(Green + Red − Blue)	Visible Atmospherically Resistant Index highlights vegetation in the visible spectrum, mitigating illumination differences and atmospheric effects. It is ideal for detecting water stress, pests, and damage.	Evidence a marked seasonal pattern in all categories, with annual peaks and valleys. A clear differentiation is observed between the categories, with “No damage” being the one with the highest value and the least variability, while “Severe” presents the lowest value and the greatest dispersion. Presents an annual cyclical behavior, which suggests a strong influence of climatic factors.	Higher values indicate healthy vegetation; lower values suggest stress or damage. Useful for identifying areas at risk of damage
GNDVI = (Green − Red)/(Green + Red)	The Green Normalized Difference Vegetation Index is a measure of vegetation greenness and vigor. Higher values indicate a higher concentration of chlorophyll, suggesting healthy and vigorous vegetation	Although there is a seasonal pattern, a general downward trend in GNDVI values is evident for all categories throughout the study period. This may suggest an overall decline in tree health or alterations in environmental conditions	Although exhibiting seasonal patterns, the GNDVI values for all categories demonstrate a general declining trend throughout the study period. This may suggest an overall deterioration of tree health or alterations in environmental factors
NDWI = (Green − NIR)/(Green + NIR)	The Normalized Difference Water Index is a measure of water content in vegetation. Lower values indicate lower water content, suggesting water stress.	Provides a comprehensive view of temporal variations in tree water content, facilitating the identification of seasonal patterns and categorical differences. No clear trend of decreasing values is observed as damage increases.	It should be combined with other indices to obtain a more comprehensive view of vegetation health. Sensitive to changes in water content
TGI = (Green − Red)/(Green + Red + Blue)	The Triangular Vegetation Index is a measure of vegetation greenness, using red, green, and blue bands. Higher values indicate better plant health and higher chlorophyll content	A clear and consistent trend related to increasing damage severity is not evident in the data. There is significant interannual variability in the observed values.	The TGI serves as a proxy for tree health. Elevated TGI values are indicative of healthy, robust trees, whereas reduced values signal stress and potential decline. Trees experiencing mild to moderate stress exhibit intermediate TGI values.
NIR	The near-infrared region is sensitive to water and biomass content.	A decrease in reflectance is expected as damage severity increases	NIR serves as a proxy for tree health. Elevated NIR values are indicative of healthy, robust trees, whereas reduced values signal stress and potential decline. Trees experiencing mild to moderate stress exhibit intermediate NIR values.
Green band	Green spectral band. Utilized for quantifying and assessing vegetation health.	Decrease in values as damage increases.	It shows lower chlorophyll content in damaged trees; with considerable variability over time
Red band	Red spectral band. Crucial for vegetation analysis due to chlorophyll’s strong absorption in this region.	These bands allow differentiation between elements with diverse levels of affliction or state. Elevated metrics are indicative of a favorable condition, while reduced metrics suggest progressive degradation	Exhibits variability across different categories. “No damage” category displays the highest values, indicative of an optimal condition. In ‘Incipient’, ‘Moderate’, and ‘Severe’ categories, values decrease, suggesting progressive degradation. Notwithstanding these disparities, overall trends within each category remain relatively stable over time

).

In addition to identifying early signs of S. noctilio damage, our analysis also highlighted how these spectral indices fluctuated over time, reflecting the evolving dynamics of the infestation in pine plantations. Interestingly, we observed that the significance of these indices shifted between 2019 and 2023, likely reflecting changes in the dynamics of the Sirex population and its impact on tree mortality. As the infestation progressed, early signs of damage became increasingly pronounced, underscoring the growing impact of this invasive pest on plantation forest health. Furthermore, while the spectral indices of all damage categories exhibited similar temporal trends, indicating a generalized spectral response of vegetation to the infestation, inter-category variability was observed. Some damage categories showed higher or lower values in specific spectral bands, suggesting differential physiological responses to the Sirex infestation. By incorporating these temporal trends into our classification models, we achieved accurate mapping of damage levels across the study area, enabling visualization of the extent and severity of Sirex damage over time. This enhanced mapping capability is essential for guiding targeted pest management efforts, ensuring that interventions can be implemented promptly and effectively to mitigate further damage to forest resources.

Given that the accuracy and consistency of our approach largely depend on the field data used for training, future research should explore the integration of additional remote sensing data sources, such as hyperspectral imagery or LiDAR, which could provide greater spectral and structural detail to enhance the discrimination of subtle stress symptoms. In our study, we relied on training data obtained from aerial visual surveys of the damage categories generated by S. noctilio over large areas, which may introduce inherent biases. These biases can arise from subjective interpretations of Sirex damage levels, variations in visibility due to environmental conditions, and the limitations of aerial surveys in capturing fine-scale damage details. Additionally, classification results may be influenced by atmospheric conditions, sensor calibration, and variability in vegetation responses to stress factors. However, when evaluating the model’s performance in a new plantation affected by Sirex, where damage levels were quantified in smaller sampling units, its accuracy remained consistent. This suggests that our model can adapt to different assessment contexts and continues to be valuable for detecting Sirex tree mortality.

5. Conclusions

Our study highlights the potential of machine learning models in accurately classifying tree health status, distinguishing between healthy and infested trees, and assessing varying degrees of damage in forests plantation plots. Time series analysis of spectral indices revealed distinct temporal patterns associated with different infestation levels, facilitating the development of early warning systems. The generated maps provide valuable spatial information on infestation distribution and severity. These insights can guide targeted management strategies, such as selective treatments or the removal of infested trees, optimizing resource allocation and minimizing environmental impacts. The spectral indices employed in this study have proven useful. However, the development of more sensitive indices, based on a deep understanding of the spectral and biochemical properties of the plant and pathogen, represents a promising avenue for advancing early detection of S. noctilo infestations through remote sensing.

Although this study underscores the advantages of remote sensing and predictive machine learning models for forest health assessment, certain limitations must be addressed. Factors such as atmospheric variability, vegetation heterogeneity, and data quality remain challenges. Future research should focus on refining the methodology and incorporating additional data sources, including hyperspectral imagery and LiDAR. Such advancements could significantly enhance the accuracy and reliability of this approach, broadening its applicability for large-scale forest management.