Comparison of Artificial Intelligence Algorithms and Remote Sensing for Modeling Pine Bark Beetle Susceptibility in Honduras

Abstract

The pine bark beetle is a devastating forest pest, causing significant forest losses worldwide, including 25% of pine forests in Honduras. This study focuses on Dendroctonus frontalis and Ips spp., which have affected four of the seven native pine species in Honduras: Pinus oocarpa, P. caribaea, P. maximinoi, and P. tecunumanii. Artificial intelligence (AI) is an essential tool for developing susceptibility models. However, gaps remain in the evaluation and comparison of these algorithms when modeling susceptibility to bark beetle outbreaks in tropical conifer forests using Google Earth Engine (GEE). The objective of this study was to compare the effectiveness of three algorithms—random forest (RF), gradient boosting (GB), and maximum entropy (ME)—in constructing susceptibility models for pine bark beetles. Data from 5601 pest occurrence sites (2019–2023), 4000 absence samples, and a set of environmental covariates were used, with 70% for training and 30% for validation. Accuracies above 92% were obtained for RF and GB, and 85% for ME, along with robustness in the area under the curve (AUC) of up to 0.98. The models revealed seasonal variations in pest susceptibility. Overall, RF and GB outperformed ME, highlighting their effectiveness for implementation as adaptive approaches in a more effective forest monitoring system.

1. Introduction

Phenomena directly related to the effects of climate change, such as prolonged droughts, hurricanes, increased solar radiation, and drastic temperature shifts, are transforming not only ecosystems, but also the biology and ecology of pests [1,2,3]. These climatic conditions are intrinsically linked to the proliferation rate and dynamics of pests, promoting their expansion and adaptability [4,5].

In the forestry sector, pests have caused significant damage to both natural ecosystems and planted forests around the globe [6]. These disturbances have negatively impacted forest biodiversity, health, and resilience, leading to considerable environmental, economic, and productive damage [7,8]. The bark beetle, in particular, is one of the most devastating forest pests in coniferous forests worldwide [9]. In Europe, for instance, the volume of timber damaged by the bark beetle increased from 2.1 million m³ y⁻¹ in the 1970s to 14.5 million m³ y⁻¹ by 2010, with a projected increase to 17.9 million m³ y⁻¹ by 2030 [10]. In North America, a significant forest area was lost between 2000 and 2020: a recent outbreak in Alaska affected approximately 650,000 ha between 2016 and 2021; in British Columbia, Canada, over 8 million ha were affected during those 20 years, representing around 723 million m³ of timber, while in the United States, 6.6 million ha were affected between 1980 and 2000 alone [10,11]. This forest damage has also extended to Nicaragua, Mexico, Guatemala, and, primarily, Honduras [12,13,14].

In Honduras, between 2014 and 2016, pine beetles, primarily Dendroctonus frontalis and Ips spp., caused the greatest historical loss of coniferous forests in the country, affecting approximately one quarter (>500,000 ha) of Honduras’s pine forests [14,15,16]. This event coincided with an extreme drought and high temperatures recorded during that period. Specifically, 2015 was recorded as the hottest year in the last 136 years [17]. As a result, forests were subjected to severe stress, increasing their vulnerability to bark beetles and other forest pests [18]. The impact was concentrated mainly on four of the seven pine species found in Honduras (Pinus oocarpa, P. maximinoi, P. caribe, and P. tecunumanii) [19,20].

Due to the critical importance of coniferous forests, the losses caused by these beetles between 2014 and 2016 had profound repercussions throughout the country. In 2014, forest cover in Honduras accounted for 48% (5,398,137 ha) of the national territory, of which 36.3% (1,960,511 ha) consisted of coniferous forests, which were affected by the beetle starting that year [21]. From a productive standpoint, these forests are of paramount importance to Honduras as they support approximately 98% of the country’s forestry industry [22,23]. Additionally, they provide other wood-derived products such as charcoal, firewood, pulp, paper, and energy residues, as well as non-timber products such as resins and seeds, contributing more than 1% to the national gross domestic product (GDP) [24]. Furthermore, they offer essential ecosystem services, such as maintaining the hydrological cycle; soil conservation; water quality regulation across numerous watersheds; carbon (C) storage in forest biomass and soil; wind control; flood prevention; biodiversity conservation; tourism; landscaping, etc. [25,26,27,28].

For the reasons outlined above, it is essential to have effective strategies for forest protection through the use of innovative tools and advanced technologies, such as AI-based susceptibility models, for monitoring and assessing risks of pest attacks, thereby ensuring forest resilience and sustainability [29]. The implementation of early warning systems based on risk assessment models has gained relevance in recent years [30,31]. Since prevention and protection are more cost-effective compared to post-outbreak control, this helps to minimize the impact of these adverse events [32]. Moreover, these warning systems enable decision-makers to identify spatial changes in susceptibility and, consequently, develop more effective management, prevention, and control strategies [33].

Studies in various regions of the world have applied artificial intelligence (AI) algorithms to model spatial susceptibility to forest pest attacks. In Central and Eastern Europe, algorithms such as support vector regression (SVR), random forest (RF), and gradient boosting (GB) have been used to develop susceptibility models for bark beetle outbreaks, incorporating environmental covariates and training data. In these studies, RF and GB have proven to be key tools for scaling models to other regions due to their high accuracy and generalization capability [31]. In the southeastern United States, forest susceptibility to bark beetle infestations has been evaluated using algorithms such as GB and extreme gradient boosting (XGB). These models, combined with environmental characteristics and historical outbreak records, have demonstrated superior modeling capacity compared to conventional methods applied between 1979 and 2000 [34]. Comparisons between AI and traditional methods have also highlighted the effectiveness of RF in pest susceptibility modeling. In studies where it was contrasted with logistic regression (LR), RF consistently outperformed LR in terms of accuracy and robustness [35]. Moreover, the maximum entropy (ME) algorithm has been widely used to model the distribution of pest species, significantly contributing to the improvement of susceptibility models [36,37,38].

LR models used in Honduras to estimate the probability of pest occurrence have reported accuracy values below 70% [12]. To date, only the RF algorithm, applied through conventional software, has been used to characterize spatial patterns of beetle outbreaks [13]. However, current technological advances are evolving rapidly, with multiple AI algorithms, advanced remote sensing, and cloud processing offering more efficient tools, expanded capabilities, and greater methodological versatility to address this type of problem [39,40,41].

Thus, the Google Earth Engine (GEE) platform provides an efficient processing environment for handling and analyzing large-scale geospatial datasets [42,43], effectively addressing challenges related to processing, storage, and computation time that often limit conventional research [44] and enabling the use of multiple AI algorithms [45,46].

The advanced capability of GEE improves the performance of risk and forestry pests susceptibility models by the optimum use of machine learning (ML) algorithms, based on the derivation of covariate importance and with a proper hyperparameter configuration (key internal adjustments that optimize model performance) [47,48,49]. Additionally, they facilitate the creation of adaptive, user-friendly susceptibility monitoring systems [50]. These systems can be adjusted to provide annual models without temporal specificity or to offer models at shorter time intervals (seasonal, monthly, weekly) [51,52]. Consequently, access is broadened for researchers and professionals, facilitating decision-making in relevant organizations [53]. Moreover, GEE enables scripts to be converted into interactive applications that can be easily used by end users on any mobile device or integrated into other websites or monitoring, reporting, and verification (MRV) platforms, increasing their accessibility, utility, and demand [50,54]. However, although these algorithms have been used in other research contexts and computational platforms, their performance and comparison have not yet been thoroughly evaluated or explored in GEE, especially for the purpose of developing early warning systems and monitoring pest outbreaks in tropical conifer ecosystems.

In this context, our research seeks to reinforce preliminary efforts in modeling pine bark beetle susceptibility, integrating modern, advanced approaches to offer more robust and comprehensive models using cutting-edge tools such as GEE. Therefore, the objective of this study is to determine the efficacy and comparative performance of three advanced AI algorithms, RF, GB, and ME, using GEE as the key cloud processing tool, to develop an optimized geospatial model of pine bark beetle susceptibility in coniferous forest ecosystems in Honduras. We hypothesize that the use of AI algorithms available in GEE (RF, GB, ME) in combination with field samples and environmental covariates will enable the derivation of highly accurate models for the geospatial estimation of pine bark beetle susceptibility. Moreover, it will identify differences in modeling accuracies and patterns in the distribution of covariate importance, highlighting temporal scales (annual and monthly) as key elements to enhance early warning systems, particularly in tropical contexts where forest pests exhibit seasonal dynamics.

2. Materials and Methods

The methodology developed in this research is schematically presented in Figure 1, illustrating the steps followed to model the susceptibility to pine bark beetle attacks. This approach includes the use of pest absence and presence sites as training samples, the generation of exploratory covariates derived from remote sensing data, and the application of three algorithms to obtain susceptibility models and evaluate their accuracy and the behavior of covariate importance.

Definition of the susceptibility model. In the context of this study, a susceptibility model estimates the probability that a given location within the pine forest ecosystem is at risk to bark beetle infestations. This estimation is based on environmental and climatic covariates that influence bark beetle dynamics, including temperature, precipitation, water deficit, vegetation indices, and topographic characteristics such as elevation and slope [55,56] (Figure 1). This concept is modeled using a continuous approach, where values range from 0 (low susceptibility) to 1 (high susceptibility).

2.1. Study Area

The research focused on the entirety of coniferous forests located in the Republic of Honduras, covering 1,960,511 hectares (ha) distributed across the country (Figure 2). Therefore, the populations of Pinus spp. studied are distributed across the full range of geomorphological and climatic variation in Honduras, from 0 to 2849 m above sea level (m a.s.l.), which entails a wide range of precipitation and temperature. In the northern region, the 30-year average annual precipitation exceeds 2000 mm year⁻¹, reaching between 2400 and 3300 mm year⁻¹ in the department of Gracias a Dios [57,58,59]. Temperatures in the north range from 25 degrees Celsius (°C) to 29 °C, characterizing an equatorial tropical climate with a short dry season of four months [60,61]. In contrast, the southern zone is distinguished by an equatorial climate with a dry winter and monsoonal precipitation between 1500 and 2000 mm concentrated over six months, with temperatures exceeding 30 °C [62]. Finally, the central region has the lowest precipitation levels, with annual records between 1000 and 1300 mm and an equatorial climate with a dry summer [58,63,64].

2.2. Databases

2.2.1. Training Samples

The training dataset consists of two distinct categories. The first includes 5601 points documenting the date and geographic coordinates of bark beetle outbreak occurrences between 2019 and 2023, recorded by the Institute of Forest Conservation (ICF) (Appendix A, Figure A1). The distribution of these samples is primarily concentrated in areas different from those that experienced mass infestation events between 2014 and 2016. The second category comprises 4000 randomly generated points in locations without historical records of beetle infestation. These points were distributed based on an estimate of the available area free of pests, ensuring a minimum separation of one kilometer (km) from historically infested areas and each other. This restriction prevented the presence and absence points from falling within the same pixel, thereby reducing potential biases in modeling. Additionally, random dates were also assigned to these random points to facilitate monthly susceptibility modeling and maintain a balance in the number of samples. Further, a binary variable named “occurrence” was introduced to each dataset, assigning a value of 1 to points with confirmed pest presence (outbreaks) and 0 to those considered free of infestation (random points).

2.2.2. Remote Sensing Data or Environmental Covariates

The explanatory environmental variables listed in Table 1 were selected using a multicriteria analysis and considering scientific relevance (based on the literature) [31,66,67,68,69,70], data accessibility, statistical analysis, transparency, and replicability. These variables have been extensively documented in ecological and forestry studies for their role in influencing pine forest susceptibility to bark beetle infestations [67,68,69]. Additionally, a covariate selection process was carried out to optimize model performance. Initially, 11 variables documented in the literature were considered [66,67,68,69]. Then, through an iterative stepwise elimination approach in random forest (RF), those with lower importance in the model and less impact on accuracy were systematically discarded, following recommended approaches in ecological modeling and machine learning [71,72,73,74]. This process led to the exclusion of flow accumulation, distance to rivers, and distance to historical fires. Although in the KDE graph, evapotranspiration and NMDI showed lower visual separability compared to the other six variables, they were retained due to their contribution to model accuracy, as the algorithms capture nonlinear patterns and complex relationships between covariates [73,74].

Climatic factors (temperature, precipitation, and water deficit) regulate tree physiology and stress levels [75,76], while evapotranspiration reflects moisture availability [77]. Spectral indices (NDVI and NDMI) assess vegetation health [78], while elevation and slope, which shape microclimatic conditions and beetle dispersal [69], have been reported as important drivers of predisposition risk [79]. These investigations have highlighted the importance of climatic, spectral, and topographic covariates in building susceptibility models, which are also applied in this study [31,80,81,82]. These environmental covariates were obtained from various data collections available in GEE, standardized to a spatial resolution of 1000 m to ensure consistency in the scale of analysis. Additionally, the use of freely available remote sensing data ensures a replicable and transparent modeling approach, establishing a robust methodological foundation for future applications.

Table 1. Selected environmental covariates used in the pest susceptibility model.

No	Covariate	Collection	Observation
1	Land surface temperature (LST)	MODIS/061/MOD11A2	Provides 8-day averages of LST at 1 km resolution [83]
2	Precipitation	UCSB-CHG/CHIRPS/DAILY	0.05° resolution resampled to 1 km [84]
3	Evapotranspiration	MODIS/NTSG/MOD16A2/105	8-day averages at 1 km resolution [85]
4	Normalized difference vegetation index (NDVI)	MODIS/061/MOD09A1	MODIS multispectral images at 500 m resolution [86,87]
5	Normalized difference moisture index (NDMI)	MODIS/061/MOD09A1	Same collection as NDVI [88]
6	Water deficit (WD)	Precipitation—evapotranspiration	1 km resolution
7	Elevation (DEM)	WWF/HydroSHEDS/03CONDEM	Digital elevation model (DEM) with 100 m resolution [89]
8	Slope	WWF/HydroSHEDS/03CONDEM	Derived from elevation

Table 1. Selected environmental covariates used in the pest susceptibility model.

No	Covariate	Collection	Observation
1	Land surface temperature (LST)	MODIS/061/MOD11A2	Provides 8-day averages of LST at 1 km resolution [83]
2	Precipitation	UCSB-CHG/CHIRPS/DAILY	0.05° resolution resampled to 1 km [84]
3	Evapotranspiration	MODIS/NTSG/MOD16A2/105	8-day averages at 1 km resolution [85]
4	Normalized difference vegetation index (NDVI)	MODIS/061/MOD09A1	MODIS multispectral images at 500 m resolution [86,87]
5	Normalized difference moisture index (NDMI)	MODIS/061/MOD09A1	Same collection as NDVI [88]
6	Water deficit (WD)	Precipitation—evapotranspiration	1 km resolution
7	Elevation (DEM)	WWF/HydroSHEDS/03CONDEM	Digital elevation model (DEM) with 100 m resolution [89]
8	Slope	WWF/HydroSHEDS/03CONDEM	Derived from elevation

2.3. Remote Sensing Data Processing

The processing of remote sensing data was carried out on the GEE platform. The land surface temperature (LST) dataset required a scale conversion from Kelvin (K) to degrees Celsius (°C; Formula 1,

Table 2. Formulas used in the processing of covariates in GEE.

No.	Formula	Meanings/Descriptions
Formula 1	°C = K − 273.15	K = temperature in Kelvin °C = temperature in degrees Celsius
Formula 2	$T_m={\displaystyle \frac{{\sum }_{i=1}^N{T}_{d,1}}{N}}$	$T_m$ = average monthly temperature $T_{d,1}$ = 8-day average temperature N = number of observations in the month
Formula 3	$ETm={\sum }_{i_{=1}}^N{ET}_{d,i}$	$ETm$ = monthly evapotranspiration ${ET}_{d,i}$ = 8-day accumulated evapotranspiration N = number of observations in the month
Formula 4	$Pm={\sum }_{i_{=1}}^N{P}_{d,i}$	$Pm$ = monthly precipitation $P_{d,i}$ = daily precipitation N = number of days in the month
Formula 5	$NDVI={\displaystyle \frac{(NIR-RED)}{(NIR+RED)}}$	NDVI = normalized difference vegetation index NIR = band 2 (841–876 nm) RED = band 6 (620–670 nm)
Formula 6	$NDMI={\displaystyle \frac{(NIR-SWIR)}{(NIR+SWIR)}}$	NDMI = normalized difference moisture index NIR = band 2 (841–876 nm) SWIR = band 6 (1628–1652 nm)
Formula 7	$\theta ={\displaystyle \frac{\alpha \times \pi }{180}}$	$\theta$ = angle of the slope in radians $\alpha$ = angle of the slope in degrees π = 3.1416
Formula 8	P = tan(θ) × 100	P = slope in percentage $\theta$ = angle of the slope in radians tan = tangent

). Subsequently, a function was implemented to calculate the monthly mean temperature (Formula 2, Table 2). The same procedure was applied to the evapotranspiration data to obtain average monthly accumulation values (Formula 3, Table 2). For precipitation, initially recorded on an hourly basis, daily values were summed and then accumulated monthly, ending with the calculation of the monthly average (Formula 4, Table 2). The water deficit (WD) was estimated as the monthly difference between precipitation and evapotranspiration.

To determine the normalized difference vegetation index (NDVI) and normalized difference moisture index (NDMI), a cloud-free mosaic of MODIS multispectral images was initially generated, and the respective equations were applied (Formulas 5 and 6, Table 2). Elevation is a clip from the digital elevation model (DEM) in m a.s.l., while the slope in degrees was estimated using GEE’s ee.Terrain.slope algorithm based on the elevation model. To convert the slope from degrees to percentage, a transformation to radians was first applied (Formula 7, Table 2), followed by a conversion from radians to percentage (Formula 8, Table 2). These covariates were stored in the GEE cloud to facilitate and optimize processing time during the modeling phases.

2.4. Analysis of the Relationship Between the Covariates and Pest Occurrence

A Pearson correlation matrix was conducted to identify linear associations between the variables and pest presence, allowing for the detection of significant correlations and the evaluation of collinearity among the covariates, thereby optimizing the robustness of the susceptibility model. Additionally, a density distribution analysis was performed using KDE plots for the environmental covariates. The KDE provides a detailed visualization that allows the identification of pest distribution patterns based on each covariate. These analyses were crucial for interpreting pest activity patterns and gaining a deeper understanding of the interactions between environmental variables and the dynamics of pest attacks.

2.5. Artificial Intelligence or Machine Learning Algorithms Used

ML or AI algorithms are effectively applicable on the GEE platform, which offers a wide range of tools for classification, regression, and probability tasks [90,91]. To model susceptibility to Dendroctonus spp. and Ips spp. attacks, three algorithms were selected: RF, GB, and ME. These algorithms were chosen due to their availability in GEE, their robustness, and their ability to provide continuous probabilistic estimation (ranging from 0 to 1) derived from binary input data (presence/absence, or 0/1), which is crucial for modeling spatial susceptibility more accurately and with easy interpretability [92]. Other AI algorithms in GEE, such as support vector machine (SVM) and naïve Bayes (NB), also could construct continuous probability models. However, they present significant limitations. In the case of SVM, it requires transformations using techniques such as Platt scaling, which adds complexity to the model without guaranteeing better results [93,94]. On the other hand, NB assumes independence among explanatory variables, which is not always realistic in ecological or environmental studies [95,96]. Moreover, these three algorithms have proven to be highly effective in previous studies for modeling complex spatial patterns of pest distribution using covariates and training samples due to their robustness against noisy data and their ability to reduce overfitting, making them more reliable than simpler methods [31,33,34,35,36,37,38]. Finally, this study exclusively used the algorithms implemented in GEE, without including approaches based on deep neural networks, since they are not currently available on this platform, and their integration would require external processing in environments such as TensorFlow or PyTorch. A general description of each of the three algorithms used is presented below.

2.5.1. Random Forest

RF, proposed by Breiman (2001) [97], is a non-parametric ML approach that combines multiple decision trees through the random selection of features at each split [98]. The final estimates are generated through a voting procedure by each tree, integrating them by averaging; this provides high accuracy and good tolerance to outliers and noise while avoiding overfitting [99,100]. The RF algorithm evaluates model accuracy through the out-of-bag (OOB) error, providing a reliable estimate of error without the need for a separate test dataset [101]; however, in this study, an independent dataset was used for validation, thereby ensuring greater robustness [102,103].

2.5.2. Gradient Boosting

GB is an advanced, flexible, and non-parametric AI algorithm [104] that can be adapted to different loss functions to minimize errors [105,106]. Unlike other algorithms such as RF, which trains parallel models using data subsets and then combines them by averaging, GB uses a sequential iterative approach, building models based on the residuals of the previous estimates to correct its errors. This technique allows it to fit to the negative gradient of the loss function and obtain progressively more accurate estimates of the response variable [107,108,109,110].

2.5.3. Maximum Entropy

ME is an ML algorithm for ecological niche modeling (ENM) and species distribution based on the principle of maximum entropy [111,112]. ME relies on the environmental constraints of the covariates observed in occurrence records [113]. Despite its recent incorporation into GEE, it has demonstrated outstanding performance [82]. Unlike its desktop version or the R and Python packages, which rely solely on occurrence data [114], the implementation in GEE requires the inclusion of absence data for improved estimation.

2.6. Selection and Optimization of Hyperparameters

The selection and optimization of hyperparameters were conducted through a structured process in Python and R, ensuring consistency with the final implementation in Google Earth Engine (GEE). We began exploring hyperparameter selection in a Colab environment with Python, using the ‘scikit-learn’ package to fit preliminary models with an RF classifier and a GB classifier. A ‘grid search’ approach was applied to systematically explore a broad parameter space, combined with cross-validation [115,116]. At this stage, the approach enabled the selection of the best hyperparameter configuration based on model performance, as measured by the AUC metric of the receiver operating characteristic (ROC), which was key for hyperparameter selection (numberOfTrees, variablesPerSplit, minLeafPopulation, shrinkage, samplingRate, maxNodes) [117,118]. For the ME algorithm, we used the maxnet package in R (4.2.2 version) to obtain the best hyperparameters, given the similarity of its configuration with ME in GEE. Once the optimal hyperparameters were identified, a final evaluation was conducted using an independent validation set, which corresponded to 30% of the data that had been reserved from the beginning. With the insights gained, we transferred our approach to the GEE.

To optimize the selection of training samples, we excluded samples located in the edge pixels of the coniferous forest. This criterion was based on recognizing that peripheral pixels are sensitive to noise generated by the difference in adjacent forest cover [119,120]. Although the way covariates are used for modeling is not directly considered a hyperparameter, it was included in this section because choosing to use averages, minimums, or maximums of certain climatic covariates also significantly influences model results.

2.7. Importance of Covariates

To assess the relevance of covariates in modeling susceptibility to Dendroctonus spp. and Ips spp., we implemented the integrated functions explain and importance in the RF and GB classifiers, and explain and Contributions for ME within GEE. These functions provide scores that reflect the contribution of each variable to the model based on the sum of the impurity decreases induced by the covariate. The scale on which the algorithms generate importance scores varies; therefore, to enable comparison, the values were standardized to a scale of 0 to 100.

2.8. Spatial Susceptibility Modeling

A unified script was developed to model spatial susceptibility without temporal distinction, applying the three AI algorithms (RF, GB, ME) configured based on the best-selected hyperparameters. Then, to provide a continuous and adaptable monitoring tool, specific scripts were implemented per algorithm for monthly pest susceptibility estimation. This approach focused on utilizing dynamic covariates corresponding to the characteristics of each month and filtering outbreak samples by occurrence for each corresponding month, thus capturing the spatial and temporal variability of pest susceptibility and establishing a framework for the continuous updating and improvement of susceptibility models.

2.9. Thematic Accuracy Estimation

Thematic accuracy estimation of our models was conducted through a rigorous validation methodology. A 70:30 data partition was applied, with 70% of the dataset used for training and 30% for validation to assess model performance. Predictive accuracy was quantified using multiple statistical metrics. Overall accuracy (OA) measured the proportion of correctly classified susceptibility predictions [121], while sensitivity (recall or true positive rate (TPR)) evaluated the model’s ability to correctly identify areas susceptible to infestation, representing the true positive rate [122]. Specificity (true negative rate (TNR)) assessed the model’s ability to correctly classify non-susceptible areas [123], whereas the positive predictive value (PPV) determined the proportion of correctly predicted susceptible sites out of all classified as susceptible and, similarly, the negative predictive value (NPV) quantified the proportion of correctly predicted non-susceptible sites among all classified as non-susceptible [124]. F1 score, a harmonic mean of precision and recall, provided a balanced measure of sensitivity and specificity [123]. Additionally, the area under the receiver operating characteristic curve ROC-AUC served as a threshold-independent metric, evaluating the model’s ability to distinguish between susceptible and non-susceptible areas [125]. These metrics were computed from confusion matrices generated on the GEE platform. Furthermore, to enhance validation, the spatial patterns of susceptibility maps were compared against historical outbreak locations, ensuring consistency with past infestation dynamics.

In addition to the conventional 70:30 random split for model validation, we further assessed the predictive accuracy of our models using the leave-one-out cross-validation (LOOCV) approach [126]. LOOCV systematically partitions the dataset by excluding one specific year for evaluation while training the model on the remaining years. This ensures that the evaluation dataset is entirely independent from the training data in terms of temporal distribution, preventing data leakage and reducing overfitting risks. The following five experimental sets were constructed:

Training: 2020, 2021, 2022, 2023|Evaluation: 2019

Training: 2019, 2021, 2022, 2023|Evaluation: 2020

Training: 2019, 2020, 2022, 2023|Evaluation: 2021

Training: 2019, 2020, 2021, 2023|Evaluation: 2022

Training: 2019, 2020, 2021, 2022|Evaluation: 2023

3. Results

3.1. Relationship Between Environmental Covariates and the Pests

The Pearson correlation analysis identified significant relationships between environmental covariates and pest occurrence (Figure 3). A moderate inverse correlation was found with WD (−0.52) and precipitation (−0.53), while elevation showed a moderate direct correlation (0.42). Temperature exhibited a weak inverse correlation (−0.21), whereas slope and NDVI had weak direct correlations (0.25 and 0.17, respectively). Evapotranspiration and NDMI showed very weak inverse correlations with pests (−0.081 and −0.023, respectively). Although WD and precipitation displayed a high positive correlation (0.96), both covariates were retained in the model because they provide complementary information.

Through the analysis of environmental covariate distributions using KDE (Figure 4), clear differences between areas with and without pest activity were identified. While the Pearson correlation (Figure 3) highlights statistical relationships between the individual covariates and pest occurrence, KDE plots were used to show that higher pest occurrences were observed in mid-to-high elevation ranges (750–1500 m a.s.l.), which coincide with areas of greater host tree availability and suitable forest structure. Lower recorded temperatures corresponded to increased pest activity, reflecting infestation periods that follow prolonged tree stress rather than a direct relationship between temperature and infestation.

Water deficit and precipitation exhibited the strongest separability, indicating their key role in conditioning forests for infestation. In contrast, evapotranspiration and NDMI showed minimal separability, indicating that these covariates represent broader ecosystem moisture dynamics rather than direct physiological stress responses in host trees. These findings highlight the complex interplay of climatic, vegetation index, and topographic factors in determining forest susceptibility to bark beetle outbreaks.

3.2. Best Hyperparameters

The hyperparameters that contributed to the highest accuracies in the configuration of the algorithms, both for annual and monthly models, were obtained, and the following settings were used: for RF, numberOfTrees: 400, variablesPerSplit: 3, minLeafPopulation: 1, bagFraction: 1; for GB, numberOfTrees: 400, shrinkage: 0.05, samplingRate: 0.8, loss: LeastAbsoluteDeviation; and for ME, outputFormat: cloglog, hingeThreshold: 20, betaMultiplier: 0.1, betaHinge: −1.

The most effective settings for the covariates in the annual models included the maximum value of the monthly averages for temperature, as well as the monthly averages for precipitation and water deficit. For NDVI and NDMI, the monthly minimums provided the best modeling contribution. Non-dynamic covariates such as elevation and slope remained unchanged. As additional information, covariates such as water flow accumulation, proximity to the water network, and historical fires were initially included, but later excluded from the models as they did not provide significant effects. The prior removal of samples with pixel edge effects produced better accuracy levels than using the entire sample set.

3.3. Spatial Susceptibility Models

The models implemented with the RF, GB, and ME algorithms demonstrated efficiency in identifying areas with high susceptibility to infestation in Honduras. As illustrated in Figure 5, the central regions of the country, which experienced intense outbreaks between 2014 and 2016, were precisely highlighted by the models despite excluding samples from affected sites for that period during training. Visually, the maps reveal subtle differences between the algorithms. GB shows a tendency to identify areas with a history of pest outbreaks with higher values (darker reds), being slightly more sensitive compared to RF and more noticeably so compared to ME. The RF model exhibits patterns similar to GB but with a slightly lower intensity. In contrast, ME, although it adequately highlights historically affected areas, shows a lower intensity of susceptibility to beetle attacks (lighter color on the map, Figure 5).

In general terms, the high susceptibility estimated by the three models largely coincides with historically affected areas, particularly in the departments of Francisco Morazán, Comayagua, El Paraíso, Olancho, and Yoro (Figure 5), where the greatest forest losses due to this pest have been recorded. This correspondence suggests that the models accurately capture environmental relationships and relevant spatial patterns, supporting the suitability of the selected covariates to explain the potential distribution of the bark beetle and validating the robustness of the models. Furthermore, this reinforces the notion that the models do not simply replicate historical data but learn general susceptibility patterns.

This finding increases confidence that newly estimated high-susceptibility areas may indeed be at risk, even if they have not yet been affected. This underscores the importance of these models not only as a retrospective tool but also as a key input for proactive forest management and prevention strategies.

In contrast, susceptibility is considerably lower in the western part of the country, especially in Copán, Ocotepeque, and part of Lempira, as well as in the eastern region, in the department of Gracias a Dios (Figure 5). This trend aligns with the lower presence of historical outbreaks compared to the central departments, which can be attributed to overall less favorable environmental conditions for the development and expansion of pestы. However, these conditions should continue to be monitored because climate changes may generate effects on any of the environmental variables, which could lead to more intense bark beetle outbreaks in regions that were previously less affected.

Monthly models reveal significant differences and specific patterns. While the RF and ME models show similarity in how they model annual susceptibility, GB displays a tendency toward higher values. As shown in Figure 6, GB exhibits abrupt transitions from low to high susceptibility zones, shifting from green to red (low to high) without the smooth intermediate gradients characteristic of its annual model. Additionally, all three models highlight significant monthly variations in susceptibility values, with certain months, such as July, August, and September, showing an increase in the estimation of zones with higher susceptibility to pest outbreaks.

3.4. Importance of Environmental Covariates

The importance of covariates varied among the algorithms (Figure 7), although a trend of higher relevance for climatic variables was observed. This trend was more pronounced in RF and GB compared to ME. Notably, precipitation ranked as the most significant covariate in the RF and GB models, while WD held this position in ME, also occupying relevant positions in GB and RF. Temperature was the third most important covariate in RF and GB, but its relevance decreased, dropping to the seventh position in ME. Regarding topographic variables, elevation ranked second in RF, fourth in GB, and third in ME. It is important to highlight that the gradient of covariate importance decreases more gradually in RF, whereas more pronounced variations are observed in GB, and this effect is even more marked in ME, demonstrating significant diversity in the weighting of factors among the different susceptibility models.

3.5. Thematic Accuracy of the Models

The accuracy evaluation for the susceptibility models was conducted using two validation approaches to ensure a robust assessment of model performance. First, we applied stratified random sampling cross-validation (70:30 split), where 70% of the data were used for training and 30% for evaluation. The results of this approach are presented in

Table 3. Validation metrics for the thematic accuracy of the annual models using stratified random sampling cross-validation.

Metric	RF (%)	GB (%)	ME (%)
Overall accuracy (OA)	92.44	92.08	85.01
Sensitivity (recall or TPR)	94.87	94.25	86.51
Specificity (TNR)	89.11	89.11	82.95
Positive predictive value (PPV)	92.27	92.22	87.42
Negative predictive value (NPV)	92.70	91.88	81.78
F1 score	93.55	93.22	86.96

. Additionally, we implemented the leave-one-out cross-validation (LOOCV) method to further analyze the models’ generalization capabilities across different years. The results of this approach are presented in

Table 4. Validation metrics for the thematic accuracy of the annual models using leave-one-out cross-validation (LOOCV).

IA	Random Forest (%)						Gradient Boosting (%)						Maximun Entropy (%)
Year	Acc.	Sen.	Spe.	PPV	NPV	F1	Acc.	Sen.	Spe.	PPV	NPV	F1	Acc.	Sen.	Spe.	PPV	NPV	F1
2019	89.88	77.07	97.71	95.38	87.44	85.25	90.72	81.28	95.63	90.63	90.76	85.70	84.24	68.87	94.29	88.75	82.25	77.55
2020	87.39	74.56	96.99	94.88	83.60	83.50	83.69	69.14	95.70	93.00	78.97	79.32	71.33	54.35	93.84	92.13	60.80	68.37
2021	91.74	91.60	91.99	95.38	85.86	93.45	89.50	91.40	86.41	91.63	86.06	91.51	84.32	85.92	81.47	89.25	76.36	87.55
2022	93.08	90.19	95.17	93.13	93.04	91.64	91.29	89.26	92.69	89.38	92.61	89.32	86.97	81.85	90.90	87.38	86.68	84.52
2023	91.95	92.39	91.28	94.13	88.81	93.25	90.03	92.57	86.56	90.38	89.53	91.46	85.67	88.65	81.58	86.88	83.94	87.75
µ	90.81	85.16	94.63	94.58	87.75	89.42	89.05	84.73	91.40	91.00	87.59	87.46	82.51	75.93	88.42	88.88	78.01	81.15

Note: Acc.: accuracy, Sen.: sensitivity (also known as recall), Spe.: specificity, PPV: positive predictive value, NPV: negative predictive value. F1: F1 score, µ: average.

. In the first approach (stratified random sampling cross-validation), the accuracy evaluation for the susceptibility models showed that RF and GB significantly outperformed ME, with an OA of 92.44% and 92.08% for RF and GB, respectively, compared to 85.01% for ME (Table 3). The RF algorithm led in sensitivity (94.87%) and NPV (92.70%). In most accuracy metrics, RF and GB were similar, while ME was significantly lower across all metrics. The F1 score, which measures the balance between precision and sensitivity, was highest for RF (93.55%), but GB showed a nearly similar result (93.22%), compared to ME (86.96%).

The LOOCV approach results are summarized in Table 4. They show a slight decrease in overall accuracy (OA) compared to the stratified random sampling cross-validation approach.

RF and GB maintained high OA values, while ME performed significantly lower. Sensitivity (TPR) remained high for RF and GB, confirming their ability to detect susceptible areas. TNR and NPV showed a slight decline, indicating potential challenges in distinguishing non-susceptible areas under varying temporal conditions. The F1 score remained high for RF and GB, reinforcing their stability across different time periods.

The discrimination capacity of the models measured by the AUC, was notably high, with RF reaching 0.98, followed by GB with 0.95, and ME with 0.90, as shown in Figure 8.

Monthly accuracy analyses reveal an irregular distribution pattern. Fluctuations in models’ effectiveness are observed (Figure 9 and Table 4), reflecting the influence of seasonal variations on pest susceptibility modeling. Although all three algorithms maintain robust performance in most months, this is not consistent throughout the year. In the case of ME, the differences between monthly and annual accuracy are smaller compared to RF and GB, where these differences are more pronounced, suggesting that ME might be less sensitive to sample size in most months. However, ME exhibited a particularly low accuracy in November, which seems to be related to the reduced number of samples available that month (Figure A6). Nonetheless, this effect was not observed in December, despite both months having a similar sample size. This could be attributed to ME prioritizing a limited number of covariates, assigning less importance to the rest. Therefore, if the most relevant covariates in November exhibited lower variability or a weaker relationship with pest dynamics, this might have had a more significant impact on ME’s performance for that specific month. In contrast, RF and GB distribute importance more evenly across multiple covariates, demonstrating greater resilience to such scenarios.

After observing the monthly fluctuations in model accuracy, it can be noted that, on average, all three algorithms showed a slight reduction in accuracy compared to the annual model. Nevertheless, these values are robust and reliable (

Table 5. Monthly comparison of the RF, GB, and ME performance in accuracy and estimation metrics.

AI	Random Forest (%)						Gradient Boosting (%)						Maximum Entropy (%)
Month	Acc.	Sen.	Spe.	PPV	NPV	F1	Acc.	Sen.	Spe.	PPV	NPV	F1	Acc.	Sen.	Spe.	PPV	NPV	F1
J	82	84	80	84	80	84	85	86	85	88	83	87	81	79	84	86	76	82
F	82	86	77	83	81	84	79	84	73	80	78	82	75	71	80	82	68	76
M	79	82	73	84	71	83	80	83	76	86	72	84	80	77	86	91	69	83
A	88	93	79	89	86	91	90	95	81	90	89	92	86	87	85	91	78	89
M	87	94	71	88	83	91	86	93	71	88	82	90	85	86	82	92	72	89
J	84	96	64	82	91	89	86	96	69	84	91	90	83	89	73	85	80	87
J	85	94	67	84	86	89	84	93	69	84	84	88	81	91	64	82	79	86
A	87	93	76	86	88	89	87	94	75	85	89	90	83	83	82	88	76	85
S	92	94	88	91	93	92	94	94	94	95	93	95	86	86	87	88	84	87
O	89	88	89	88	90	88	87	87	86	85	89	86	85	80	90	87	84	84
N	83	80	84	70	90	74	81	80	82	67	90	73	82	64	90	76	84	69
D	90	84	93	86	92	85	94	90	96	92	95	91	86	76	92	83	88	79
µ	86	89	79	84	86	87	86	90	80	85	86	87	83	81	83	86	78	83

Note: Acc.: accuracy, Sen.: sensitivity (also known as recall), Spe.: specificity, PPV: positive predictive value, NPV: negative predictive value. F1: F1 score, µ: average.

) to support decision-making in pest management.

4. Discussion

The results obtained with GEE demonstrate that its capabilities to execute advanced AI algorithms, such as RF, GB, and ME, are notably efficient and comparable to those of more complex programming software, such as R and Python. This highlights the value of GEE as a viable and accessible alternative for performing bark beetle susceptibility modeling in less time, taking advantage of its simplicity and processing capacity, as noted in [51]. The annual models of RF and GB achieved outstanding accuracies (92.44% and 92.08%), surpassing ME, which reached 85.01%. Although this last accuracy is high, it was lower than that of the other two algorithms (RF and GB) despite ME being widely used for species distribution modeling in conventional software. This may be due to the recent incorporation of this algorithm into GEE [82]. However, these trends have also been observed in studies on pest distribution in central Chile where RF and GB performed better than ME (85% and 83.6% versus 80.1%) [92]. Consequently, the annual accuracies of our models were also significantly higher than those obtained with logistic models applied in Honduras, with 72% reported in [12]. Additionally, the robustness of these models is evidenced by their AUC, with RF and GB reaching values of 0.98 and 0.95, respectively, surpassing previous research that exclusively used RF in the R version, such as [13], where an AUC of 0.94 in the dry season and 0.87 in the rainy season was obtained. In the case of ME, the AUC of 0.90 shows comparable effectiveness to [82] and is higher than the AUC of 0.87 [127], which also demonstrates its high potential in GEE.

Moreover, the comparison between the stratified random sampling cross-validation (70:30 split) and the LOOCV revealed that although RF and GB maintained strong predictive performance in both cases, the LOOCV approach resulted in slightly lower accuracy scores. This decrease suggests that random partitioning may lead to overfitting due to temporal data similarities, reinforcing the need for validation techniques that account for temporal variability. Despite this, RF and GB continued to outperform ME, demonstrating their stability and reliability even when evaluated under different temporal conditions. The selection of the best covariates and hyperparameter tuning were essential for building more accurate models, as also emphasized in [48]. Additionally, following refs. [119,127] in the careful selection of training samples, we discarded samples from the areas affected by Dendroctonus spp. and Ips spp. between 2014 and 2016, as well as the samples spectrally contaminated by edge effects between different land uses, mitigating possible biases and improving model accuracy in unseen areas. Notably, despite not using samples from the 2014–2016 outbreak, the models accurately predicted these areas as high-risk zones. This highlights their ability to accurately reflect historical realities, reaffirming their potential utility for concentrating on prevention and mitigation strategies in the future. This comprehensive approach demonstrates that success in modeling depends not only on the quantity of samples or covariates, but also on the quality of each adjustment and methodological decision.

The analyses also revealed that covariates with a higher Pearson correlation and greater separability in KDE distribution analyses were also the most important for reducing errors in the models. These findings align with a similar study [68], which demonstrated that correlations between environmental variables were key to identifying climatic patterns of Dendroctonus spp. in Canada. This emphasizes the strategic selection of covariates by also paying attention to density distributions and correlations, even modeling non-linear phenomena. This approach not only enriches our understanding of the interactions between environmental and biological factors, but also offers an applicable criterion for probabilistic modeling of other phenomena.

The distribution of covariate importance showed differentiated patterns among the models, which could be related to modeling accuracy. RF, the most accurate model, assigned relatively balanced importance values across covariates, while GB presented a moderately more skewed distribution. In contrast, ME, the least accurate of the three, exhibited an abrupt distribution, with few covariates considered more important and assigning low importance to most others. This highlights that each algorithm prioritizes covariates differently, influencing their outcomes. The temperature covariate showed a weak Pearson correlation with the pest (−0.23), influencing the importance derived by the algorithms, ranking third in RF and GB, and seventh in ME. This differs from studies such as [67], where temperature played a more important role in pest modeling. This discrepancy can be explained by the conditions in Honduras, where extreme temperatures exceeding 35 °C during the dry season do not favor the development of this pest (Appendix C, Figure A4 and Figure A5), as this insect thrives within a more specific temperature range. Thermal threshold was preliminarily determined for D. frontalis in experimental conditions as being optimal between 20 to 30 °C for the insect’s whole lifecycle [128]. In contrast, humidity, represented by precipitation, plays a more prominent role, as during the rainy season, both conditions (temperature and humidity) are favorable for the beetle’s development (Appendix B, Figure A2 and Figure A3). Consistent with this, the density distribution analyses showed that temperature in sites with pest presence peaks around 30 °C, with a normal distribution range between 25 °C and 35 °C. This pattern aligns with the development rate of D. frontalis as a function of temperature described in [129]. Beetle populations decrease as they approach the temperature extremes of this optimal range; according to refs. [128,130], egg hatching in some pine beetles is nonexistent above 34 °C or 35 °C. Additionally, it was observed that at these temperatures, beetles become photo-tactic, which could explain the lower susceptibility during months with higher tempera-tures [131].

The intensity of the monthly models fluctuates considerably between different months and algorithms (Figure 6), which aligns with studies such as [132], which highlights the multitemporal variability in bark beetle presence. GB stands out for showing greater intensity in high-risk areas, such as sites affected between 2014 and 2016, while RF and ME present smoother models, which are more aligned with their general annual models. Months such as July, August, and September stand out for showing higher susceptibility, in accordance with bark beetle trap counts reported by ICF over three consecutive years (2021, 2022, and 2023) [133]. These findings reinforce the crucial influence of seasonal and climatic factors on pest dynamics and underscore the need for adaptive and detailed approaches in forest management. Although annual models provide better accuracy, monthly models offer greater detail about seasonal patterns in such a kind of pest affection.

The methods developed in this study, which combine advanced AI algorithms within the cloud computing infrastructure of GEE, along with the generation of covariates from multiple remote sensing sources, have high potential for scaling to other regions and ecosystems worldwide where probabilistic modeling is required. This includes applications in susceptibility and risk assessment for other pests, invasive or beneficial species, and is relevant to forestry, agriculture, and the mapping and distribution of ecological niches. For example, AI-based approaches have been successfully applied in studies on agricultural pest management [134], forest fire monitoring [135], and habitat modeling for flora and fauna species [136], demonstrating their broad range of applicability and replicability. This modeling potential could guide management strategies in areas where pests such as the ambrosia beetle (Xyleborus spp.) threaten tree health [137]. Likewise, these methods could be applied in agricultural systems to model the spread of crop pests, such as the fall armyworm (Spodoptera frugiperda), which has caused significant losses in maize and other crops worldwide [138]. However, it is important to acknowledge that these methods require substantial training datasets and careful selection of covariates, which may not always be available in regions where systematic records of such adverse phenomena are lacking.

5. Conclusions

This study demonstrates the effectiveness of AI algorithms in GEE for modeling pine bark beetle susceptibility. RF and GB showed higher patterns and accuracies than ME, suggesting that either one, or a combination of both, could be suitable for more robust and comprehensive models. GB stands out in monthly modeling due to its ability to highlight risk intensities, which is crucial for raising public awareness. These approaches are valuable for MRV systems proposed by the United Nations Framework Convention on Climate Change (UNFCCC), contributing to the Reducing Emissions from Deforestation and Forest Degradation (REDD+) objectives through adaptable and updatable models.

The differentiated assignment of importance values to covariates by the algorithms highlights the critical role of covariate selection and prioritization in model performance. This underscores the relevance of a meticulous selection, both in the quantity and modeling capacity of covariates, to achieve an appropriate balance between accuracy, interpretability, and the temporal and spatial efficiency of the models.

In this study, precipitation-driven humidity was identified as a key covariate, surpassing temperature in importance. This contrasts with studies from temperate regions where temperature is predominant, highlighting the need for a holistic approach in geospatial modeling that adapts models to specific local or regional environmental conditions and different temporal scales.

Annual models demonstrated better overall precision. However, monthly models provided more detailed and relevant information for early warning systems, particularly by highlighting specific temporal risk patterns. GB excelled in monthly modeling by more clearly identifying the most critical sites, which is valuable for forest planning and short-term decision-making to raise environmental awareness. This multiscale approach allows for the integration of more recent and significant temporal details, complementing annual trends and enhancing the ability to respond to pest outbreaks.