Utilizing LightGBM to Explore the Characterization of PM_2.5 Emission Patterns from Broadleaf Tree Combustion in Northeastern China

Abstract

PM_2.5 emissions significantly impact atmospheric environments and human health in the context of forest fires. However, research on PM_2.5 emissions from forest fires remains insufficient. This study systematically investigated PM_2.5 emission characteristics from broadleaf tree combustion through controlled experiments examining three key factors: species variation (Acer tegmentosum [AT], Acer ukurunduense [AU], Acer pictum [AP], Tilia amurensis [TA], Phellodendron amurense [PA], Ulmus davidiana [UD], Ulmus laciniata [UL], Prunus padus [PP], Prunus maackii [PM]), moisture content (0%–20%), and phenological stages (budding [A], growing [B], defoliation [C]). The results demonstrated: (1) Significant interspecies differences, with UL showing the lowest, and PM the highest emissions; (2) A unimodal moisture—emission relationship peaking at 15% moisture content across most species, while AT, UL and PM exhibited unique linear responses; (3) Distinct phenological patterns, including triphasic fluctuations during the growing and defoliation phases. The LightGBM model effectively predicted emissions (R² = 0.97), identifying species (36.2% importance) and moisture content (21.6%) as dominant factors. These findings provide critical data for wildfire emission modeling and highlight the need for species-specific parameters in air quality forecasts.

1. Introduction

1.1. Background

As China experiences swift economic and social advancement, ecological and environmental challenges are increasingly coming to the fore. Among these, atmospheric pollution has emerged as a significant factor that impacts public health and hampers the sustainable development of the society. PM_2.5 refers to fine particulate matter in the atmosphere with a diameter of 2.5 μm [1], has become a key indicator of atmospheric pollution control due to its strong adsorptive properties, long-distance transport capacity, and far-reaching health impacts. Forest resources play a vital role in sustaining the global environment and supporting human society. Beyond enhancing the quality of the atmospheric environment, forests assume a pivotal role in the global carbon cycle, soil characteristics, and climate modulation. In recent years, the aggravation of global warming along with extreme drought conditions [2], has led to an upward trend in fires. The increase in fires, in turn, aggravates global warming [3]. Fire ranks among the most significant disturbances within global plant ecosystems [4]. Forest fires require three elements to occur: forest fuels, an accelerant (oxygen), and a certain temperature, with fuels being the only human-controllable factor [5]. The burning of any vegetated land, including farmland, grasslands, and forests-is termed open biomass burning (OBB) [6]. OBB produces pollutants that contribute to warming by converting incident solar radiation into heat. Forest fires include above-ground fires, which are surface fires and crown fires, where the primary fuel is trees, shrubs, and herbaceous plants above the surface, and below-ground fires, where the primary fuel is humus or peat [7].

1.2. Related Work

Over the recent past, numerous scholars have conducted research on forest fires as a calamitous event, influenced by a confluence of natural and human-induced factors. Certain studies have employed remote sensing data to evaluate fire patterns, encompassing intensity, seasonality, frequency, and fire return intervals [8]. Some have analyzed the frequency of wildfires and areas burned, as well as climatic patterns [9] and atmospheric conditions [10]. Studies have been conducted to examine the dynamics within indigenous forests [11]. Others have analyzed the impacts of wildfires on Mediterranean-type ecosystems, focusing on soil respiration and the composition of bacterial microbiota [12]. A recent study utilized advanced machine learning techniques—specifically, logistic regression, random forest, support vector machine (SVM), and XGBoost algorithms—to predict vegetation fire occurrences [13], and utilized neural networks for the prediction of forest fires [14]. Since plant flammability directly affects fire severity (as demonstrated in [15]), its assessment is critical for forest management. Another study analyzed the impact of forest fires on terrestrial chemical elements, analyzed the effects of forest fires on ground chemical elements [16]. A further study analyzed the dynamic changes in fuel flammability after trees were treated with different intensities [5]. Moreover, a research endeavor employed machine learning algorithms in conjunction with geospatial techniques to evaluate the post-fire recovery of vegetation within forest ecosystems [17]. The substantial PM_2.5 emissions from forest fires during combustion have drawn significant research attention. Some studies have analyzed the influencing factors of PM_2.5 emissions [18]. Several studies have investigated the correlation between chronic PM_2.5 exposure and its adverse health outcomes, including mortality risk [19,20]. There have been predictions of the daily PM_2.5 levels in Australia utilizing the Random Forest Model [21]. Other researchers have applied Response Surface Methodology (RSM) to quantitatively demonstrate the significant interactive effects of moisture content and fuel load on PM_2.5 emission levels. PM_2.5 emissions occur during combustion processes [22]. Forest fires release large amounts of smoke into the atmosphere, accounting for 42% of particulate emissions from biomass burning. Biomass burning emissions are predominantly composed of particulate matter, with PM_2.5 constituting over 90% of the total particulate emissions [23]. High concentrations of PM_2.5 affect people of different ages in different degrees and in different ways [24], PM_2.5 exposure not only causes immediate respiratory effects but also induces chronic cardiorespiratory impairment and hepatic damage [25,26,27,28].

While prior research has demonstrated the relationship between forest fires and vegetation flammability, along with forecasting PM_2.5 emission dynamics and associated drivers, it frequently neglects the temporal variations in intrinsic tree characteristics as contributing factors.

1.3. Research Gap

At present, domestic and international studies on PM_2.5 emission from forest fires mainly focus on coniferous species [29,30], and relatively few studies on broadleaf trees. Therefore, exploring the regular characteristics of PM_2.5 emission from broadleaf tree combustion can help to enrich and improve the research system in the field of forest fire and air pollution. Understanding the characteristics of PM_2.5 emission from broadleaf trees can help governmental departments to formulate targeted forest fire prevention and control policies and air pollution prevention measures, and to reduce the impacts of fires on the environment and human health. The study of the pattern of PM_2.5 emission from the burning of broad-leaved trees can provide theoretical support for the protection and restoration of forest ecosystems and help realize the sustainable use of forest resources.

As an indispensable component of forest ecosystems, broad-leaved trees are widely distributed and diverse, are instrumental in sustaining ecological equilibrium and offering ecological services. However, due to climate change, anthropogenic activities, and other factors, forest fires occur frequently [31], and the problem of PM_2.5 emission from the burning of broadleaf trees has become increasingly serious. The current understanding of PM_2.5 emissions from the combustion of broadleaf trees is insufficiently developed, thereby constraining our comprehension of the extent to which forest fires contribute to air pollution. This limitation also impacts the formulation and execution of pertinent preventative and control strategies. In this study, we conduct an analysis and process the data using laboratory data and the lightGBM (Light Gradient Boosting Machine) algorithm to derive pertinent conclusions.

1.4. Purpose

The purpose of this paper is to explore the regular characteristics of PM_2.5 emission from the combustion of broadleaf trees, and the main research includes: analyzing the combustion characteristics of different broadleaf trees and exploring their effects on PM_2.5 emission; through laboratory simulation combustion tests, studying the effects of moisture content, fuel load, tree parts, period and other factors on PM_2.5 emission from the combustion of broadleaf trees; combining with the lightGBM to carry out data analysis and prediction modeling were carried out to explore the effects of different characteristics on the regularity of PM_2.5 emission from broadleaf tree combustion and its correlations.

LightGBM, an algorithm grounded in Gradient Boosting Decision Tree (GBDT) methodology, exhibits superior training speed and diminished memory utilization [32]. Additionally, it demonstrates commendable accuracy, rendering it particularly apt for tasks entailing extensive datasets and high precision requirements [33,34]. Its application extends to the realms of forest fire analysis and prediction [35,36]. There are studies on forest fire susceptibility modeling based on the LightGBM method to generate accurate fire susceptibility maps [37]; the prediction of forest fire occurrence probability and the zoning of high-risk forest fire areas [38]; there are studies on the incorporation of spatial effects and temporal features into the LightGBM model to estimate the ground—level particulate matter [39]; and there is a study that, after comparing several models [40] show that concluded that the LightGBM model is the best single model for predicting PM_2.5 using historical PM_2.5 concentration data [41].

This study employed a methodology that integrated field sampling, laboratory-based simulated combustion experiments, and LightGBM data analysis to investigate the patterns and characteristics of PM_2.5 emissions resulting from the combustion of broadleaf trees. The specific technical methodologies are as follows: compile pertinent research results at home and abroad, and sort out the prevailing research status of PM_2.5 emissions from broadleaf tree combustion; select exemplary broadleaf tree species and scrutinize their combustion characteristics; design laboratory simulated combustion experiments to monitor PM_2.5 emissions from broadleaf tree combustion across various conditions; and employ lightGBM to analyze meteorological data, thereby investigating the influence of distinct characteristics on PM_2.5 emissions from broadleaf tree combustion. This paper aims to elucidate the regular characteristics of PM_2.5 emission from broadleaf trees, and provide the following theoretical support for forest fire prevention and control and air pollution prevention and control in China. It seeks to clarify the potential of PM_2.5 emission from the combustion of different broadleaf trees, thereby establishing a foundation for the selection and strategic deployment of tree species. Furthermore, the study proposes specific measures for forest fire prevention to mitigate the risk of PM_2.5 pollution. Additionally, it aims to provide a scientific foundation for the governmental entities in formulating policies for the prevention and control of air pollution. The research will offer a scholarly basis for the governmental departments to develop air pollution control policies.

2. Materials and Methods

2.1. General Introduction to the Study

The study site of this paper is the Liangshui Experimental Forestry Farm (Liangshui National Nature Reserve) of Northeast Forestry University, which is located on the eastern slope of the Dali Dailing Branch, the southern section of the Xiaoxinganling Mountain Range in the eastern mountains of northeastern China, with a total area of 12,133 ha, and the geographic coordinates of longitude from 128°47′8″ to 128°57′19″, latitude between 47°6′49″ and 47°16′10″ (Figure 1). The zonal vegetation of Liangshui Experimental The forest is a temperate mixed coniferous and broad-leaved forest, in which Korean pine serves as the dominant tree species [42], which has greater typicality and representativeness. The species composition of the original Korean pine forest is dominated by Korean pine, accompanied by a variety of temperate broadleaf species, such as Tilia, Betula, Ulmus laciniata, etc., up to more than 20 species.

Liangshui Experimental Forestry Farm is located on the eastern edge of the Eurasian continent. Deeply influenced by the oceanic climate, it exhibits distinct characteristics of a temperate continental monsoon climate [43]. Due to its elevated latitude, the region experiences an annual mean temperature of −0.3 °C. The mean maximum temperature reaches 7.5 °C, while the mean minimum temperature dips to −6.6 °C. Annual precipitation amounts to 676 mm [44]. The climate is predominantly characterized by long, cold, and dry winters and short, wet, cool, and rainy summers.

The prevailing topographical pattern indicates elevated terrains in the northern, eastern, and western regions, with a decline towards the central and southwestern areas. These regions encompass portions of the core and experimental zones situated within a basin. The elevation spans from 280 m to 707.3 m, and the average elevation of the whole area is approximately 400 m, featuring relative height differences ranging from 80 to 300 m [45]. The area falls within the category of low hills, characterized by rounded hilltops and predominantly asymmetrical slopes. The average gradient lies between 10 to 15 degrees, while certain segments exhibit steep inclines exceeding 20 to 40 degrees.

2.2. Sample Collection and Preparation

We selected nine broadleaf tree species in the Liangshui National Nature Reserve as study materials, including Acer tegmentosum (AT), Acer ukurunduense (AU), Acer pictum (AP), Tilia amurensis (TA), Phellodendron amurense (PA), Ulmus davidiana (UD), Ulmus laciniata (UL), Prunus padus (PP), Prunus maackii (PM) (

Table 1. Predominant tree species inhabiting Liangshui Experimental Forestry Farm.

Species of Trees	Taxonomic Families and Genera
Acer tegmentosum	Aceraceae
Acer ukurunduense	Aceraceae
Acer pictum	Aceraceae
Tilia amurensis	Tiliaceae
Phellodendron amurense	Rutaceae
Ulmus davidiana	Ulmaceae
Ulmus laciniata	Ulmaceae
Prunus padus	Rosaceae
Prunus maackii	Rosaceae

). All broadleaf tree species were carefully selected from mature individuals situated on slopes with the same aspect. During the budding period (early April), the growing period (late July), and the defoliation period (early October), samples were systematically collected from three distinct tree parts: branches, bark, and leaves. Initially, sampling was carried out on the slope situated to the south of Liangshui National Nature Reserve, where the slope gradient ranged from 20 to 23 degrees. To determine the maturity of the trees, we considered multiple essential factors, with the diameter at breast height (DBH), the vertical position of the lowest living branch, the dimension of bark thickness, and the amount of stem curvature being the main ones [46]. Upon confirming the species of the sampled trees, we proceeded to excise the branches using high pruning shears. Subsequently, we carefully separated the branches, bark, and leaves using scissors. Then, for each tree species, we combined the similar parts (branches with branches, bark with bark, and leaves with leaves) and discarded any damaged or unsuitable samples within each group. These samples were then promptly placed into airtight plastic bags to prevent dehydration and meticulously labeled with permanent markers. It is pertinent to note that during the budding period in April, only branches and bark were available for sampling, as the trees had not yet developed leaves.

Subsequently, the samples underwent laboratory drying, and the moisture content was determined. Utilizing a drying oven set at 105 °C, the samples were dehydrated until they reached a state of absolute dryness, after which they were portioned into plastic bags, duly labeled, and preserved in a cool environment for the subsequent calculation of their moisture content.

The moisture content of fuel is defined by Equation (1) [47]:

$$FMC={\displaystyle \frac{W_F-W_D}{W_F}}\times 100\%,$$

(1)

In the formula, FMC represents the moisture content of fuel, and W_F and W_D denote the fresh and dry weights (kg) of fuel.

During indoor simulated combustion experiments, PM_2.5 concentrations were released and comparatively analyzed from various broadleaf species, including branches, bark, and leaves, at different times and moisture contents.

Following extensive experimentation during the preliminary phase, it was established that fuels exhibiting a moisture content exceeding 20% experience heightened difficulty in ignition and tend to undergo incomplete combustion [29]. We wanted to explore how different moisture contents affect PM_2.5 emissions from fuel combustion. For this purpose, we set up five different moisture contents with 0%, 5%, 10%, 15% and 20%, for combustion experiments. Based on the known moisture content of the combustible material needed in the experiment, the amount of water to be added was calculated according to Equation (1). The water was quickly sprayed on the surface of the combustible material with a sprayer and the combustible material sealed in a plastic bag for 24 hs so that could completely absorb the water.

2.3. Determination of PM_2.5 Compositions

The collected samples were first dried and weighed, followed by simulated combustion experiments carried out in a self-designed device (Figure 2). These experiments were conducted within the fume hood of the laboratory at Northeast Forestry University, utilizing an experimental setup primarily composed of a custom-built combustion bed, a particulate matter sampler and a PM_2.5 concentration monitor.

The combustion experiments were designed to simulate smoke emissions from a fire occurring in a natural environment. Different combinations of fuel beds were laid according to the experimentally established fuel load gradient levels, and three trials were conducted for each bed type. The fuel beds were constructed using a 30 cm × 30 cm wire mesh to minimize heat loss during combustion. The experimental beds were leveled to maintain a strictly horizontal orientation. Initial experiments indicated that the smoke generated from the combustion of an excessively large fuel load was not fully absorbable. Consequently, three distinct fuel loads were established, weighing 1 g, 2 g, and 3 g respectively. Prior to each combustion experiment, the samples were meticulously weighed and uniformly distributed upon the combustion bed. The experiment employed an alcohol lamp as an ignition source to guarantee a uniform combustion state in each iteration. A total of 3240 combustion experiments were conducted. The experiments were designed to cover three periods, nine tree species, with each species having samples from three different tree parts. Additionally, the study considered five moisture contents, three fuel loads, and each experimental condition was replicated three times. The gravimetric method, as referenced in literature [48] was used to measure the concentration of PM_2.5, which could be calculated from the mass difference of the filter membrane and the standard sampling volume. PM_2.5 was collected using a JCH-6120 med a flow ambient particulate sampler manufactured by Qingdao Juchuang Company (Qingdao, China), with a sampling flow rate of 60–130 L/min. The atmospheric particulate matter collector functions by drawing a specific volume of air through a filter membrane of known mass at a steady flow rate. This process enables the suspended particles present in the air to be trapped and retained within the filter membrane. Determined from the increment of the filter membrane’s mass and the volume of traversed air, the mass concentration value of total suspended particles in the atmosphere provides crucial data for environmental studies. The duration for sampling was established at 20 min. The indoor PM_2.5 concentration was continuously monitored using the GQ-058 PM_2.5 air quality monitor, which is manufactured by Shenzhen Jishun An Technology Co., Ltd., located in Shenzhen, China. The basic operational principle of the PM_2.5 concentration monitoring apparatus is based on the phenomenon of light scattering. This occurs when a beam of infrared light illuminates particles as they traverse the detection zone. The intensity of the scattering observed at an angle perpendicular to the light path correlates with the diameter of the particles. Through enumeration, it is possible to ascertain the real-time particle volume concentration. Subsequently, employing empirical conversion formulas and calibration techniques, a mass concentration that aligns with official units can be derived. At the end of each combustion experiment, when the indoor PM_2.5 concentration reaches the same level, the collection of PM_2.5 particles is terminated, and the sampling volume under standard conditions is then recorded.

PM_2.5 was collected with a glass fiber filter membrane. Subsequently, the samples were transferred to a controlled-environment chamber kept at a steady temperature of 25 °C and a relative humidity of 50% for 24 h [49]. To measure the weight of the filter membranes, an electronic balance (Model: FA2004N, with an accuracy of 0.0001 g, Shanghai, China) was employed. At the conclusion of the experiment, the filter membranes with adsorbed PM_2.5 particles were re-placed in a constant temperature and humidity chamber under identical conditions to eliminate any potential impact on the PM_2.5 measurement results. Additionally, PM_2.5 samples were collected from the combustion chamber under non-burned sample conditions to serve as a blank control.

The PM_2.5 mass concentration is represented by Equation (2):

$$P={\displaystyle \frac{W_1-W_2}{V}}\times 10^6,$$

(2)

where P represents the PM_2.5 mass concentration (µg·m⁻³); W₁ represents the weight of the filter membrane prior to sampling, measured in grams (g), while W₂ indicates the weight of the filter membrane after the sampling process, also in grams (g). The variable V denotes the standard volume, with the unit being cubic meters (m³).

In this study, pollutant emissions are quantified by means of emission factors (EFs) as specified in Equation (3). Emission factors are defined as the mass of a pollutant released per unit mass or unit energy of the fuel that is combusted. The EFs for PM_2.5 for each combustion experiment were calculated using the following formula. The emission factor of PM_2.5 can be expressed and calculated using Equation (3):

$${EF}_{{PM}_{2.5}}={\displaystyle \frac{m_{{PM}_{2.5}}}{m_{burned}}}={\displaystyle \frac{C_{{PM}_{2.5}}\times V}{m_{burned}}},$$

(3)

where ${EF}_{{PM}_{2.5}}$ represents the emission factor for PM_2.5, with the unit being micrograms per kilogram (µg.kg⁻¹); $m_{{PM}_{2.5}}$ is the mass of PM_2.5 collected, which is calculated as $C_{{PM}_{2.5}}$ × V; $C_{{PM}_{2.5}}$ is the average concentration of PM_2.5, measured in micrograms per cubic meter (µg.m⁻³); V is the standard volume, expressed in cubic meters (m³); and $m_{burned}$ denotes the mass of the fuel burned, and its unit is grams (g).

3. Results

3.1. Analysis of Total PM_2.5 Emission from Combustible Combustion

In this paper, the total amount of PM_2.5 emitted from the combustion of each tree part under different moisture content states of the fuel was compared. Figure 3 comprises five panels (A–E), each depicting the total PM_2.5 emissions from different tree parts at varying moisture content levels tested in the experiments. The percentage values in the figures represent the proportional contribution of each component. From Figure 3, we can see Acer tegmentosum (AT), Acer ukurunduense (AU), Acer pictum (AP), Tilia amurensis (TA), Phellodendron amurense (PA), Ulmus davidiana (UD), Ulmus laciniata (UL), Prunus padus (PP), Prunus maackii (PM) of PM_2.5 emission rules. Upon comprehensive examination, it was determined that there are considerable variations in the emission of PM_2.5 resulting from the combustion of various broadleaf trees, as well as notable differences in PM_2.5 emissions from distinct sections of the same species. The species with higher PM_2.5 emission were Prunus maackii (PM), Acer ukurunduense (AU), Acer pictum (AP), Phellodendron amurense (PA); several other tree species released less PM_2.5. The total amount of PM_2.5 emission by each species was lowest at 0% moisture content, and the total amount of PM_2.5 emission by each species was highest at 15% moisture content. At equivalent moisture levels, Ulmus laciniata (UL) exhibited the lowest total PM_2.5 emissions, whereas Prunus maackii (PM) displayed the highest. For branches of the same species, the total PM_2.5 emissions were notably lower, generally constituting less than 30% of the overall emissions; conversely, leaves emitted the greatest quantity of PM_2.5, typically exceeding 40% of the total emissions [3].

As evidenced by Figure 3, the PM_2.5 emission factor does not exhibit a singular variation in response to increasing moisture content, necessitating an individual analysis. The emission factors for Acer ukurunduense (AU) peaked at a moisture content of 10% and subsequently diminished; the emission factor for Tilia amurensis (TA), Phellodendron amurense (PA), Ulmus davidiana (UD), and Prunus padus (PP) attained their highest values at a moisture content of 15% before declining. This indicates that various tree species combust to achieve the peak concentration of PM_2.5 at distinct moisture levels. Conversely, the PM_2.5 emission factors for Acer tegmentosum (AT), Ulmus laciniata (UL), and Prunus maackii (PM) exhibited an ascending trend in correlation with increasing moisture content, implying that these species adhere to a linear progression in moisture content ranging from 0% to 20%. The remaining tree species in the study also demonstrated their own unique patterns. The moisture content of the remaining species exhibited a gradual decline subsequent to attaining the peak concentration of PM_2.5 emitted by themselves, indicating that the maximum value lies within the range of 0%–20%. The precise value necessitates further experimentation to ascertain and refine the moisture content levels.

3.2. Patterns of PM_2.5 Emission from Tree Branches in Fuel Load Combustion

Analysis of Figure 4 reveals distinct patterns in PM_2.5 mass concentration across different phenological phases. The budding (A) and defoliation (C) periods exhibited significantly higher variability in PM_2.5 concentrations (range: 0–50,000 μg/m³), suggesting greater particulate emission dynamics during these transitional growth stages. In contrast, the growing phase (B) showed markedly lower and more stable PM_2.5 levels (range: 0–25,000 μg/m³), indicating relatively consistent emission patterns during active vegetation periods.

The experimental data demonstrate a consistent positive correlation between fuel load (1 g, 2 g, 3 g) and PM_2.5 mass concentration across all observation periods, with particulate emissions increasing proportionally to fuel quantity.

Analysis of Figure 4 reveals three key moisture-dependent emission patterns: (1) During the budding period (A), PM_2.5 concentrations increased progressively (0%–20% moisture), peaking at 15%–20%; (2) The growth period (B) showed an initial decline (0%–5%), subsequent rise (5%–10%), and stabilization (10%–20%) of PM_2.5 emission; (3) The leaf fall period (C) exhibited similar but more pronounced (30%–40% greater amplitude) triphasic fluctuations.

3.3. Patterns of PM_2.5 Emission from Bark in Fuel Load Combustion

Analysis of the three phenological periods including budding (A), growing (B), and defoliation (C) revealed significant variations in bark PM_2.5 emissions. The budding and defoliation periods exhibited substantial fluctuations in PM_2.5 mass concentrations (range: 0–50,000 μg/m³), indicating higher emission volatility during these transitional phases. In contrast, the growing period demonstrated more stable emission patterns with significantly lower concentration ranges (0–25,000 μg/m³).

Figure 5 demonstrates a consistent positive correlation between fuel load (1 g, 2 g, 3 g) and PM_2.5 mass concentration across all three experimental periods, with particulate emissions increasing proportionally to fuel quantity.

Figure 5 demonstrates distinct moisture-dependent emission patterns across phenological periods. During the budding period (A), all tree species displayed progressively increasing PM_2.5 concentrations with rising moisture content (0% to 20%). The growing period (B) revealed a triphasic pattern: initial decrease (0%–5% moisture), followed by an increase (5%–10%), and subsequent decrease (10%–20%). This trend persisted during the deciduous period (C) for most species (80%).

3.4. Patterns of PM_2.5 Emission from Leaves in Combustible Combustion

Figure 6 reveals consistent PM_2.5 emission patterns from leaves across key phenological stages, with mass concentrations maintaining similar fluctuation ranges (0–40,000 μg/m³) during both the growing (B) and defoliation (C) periods.

The experimental data demonstrate a consistent positive correlation between fuel load (1 g, 2 g, 3 g) and PM_2.5 mass concentration across both observation periods, with particulate emissions scaling proportionally to fuel quantity.

Figure 6 demonstrates distinct moisture-dependent PM_2.5 emission patterns across phenological periods: (1) Budding period (A) shows a consistent positive correlation between moisture content (0%–20%) and PM_2.5 concentrations for all species; (2) Growing period (B) exhibits a triphasic trend—initial decrease (0%–5% moisture), subsequent increase (5%–10%), and final decrease (10%–20%); (3) Deciduous period (C) replicates this pattern for 80% of species.

In conclusion, our study reveals that PM_2.5 emissions from tree components (branches, bark, and leaves) demonstrate significant temporal variations across phenological stages (A–C), modulated by both fuel load and moisture content. Key findings include: (1) Emission stability decreases progressively from budding (A) to defoliation (C); (2) Fuel loads show a consistent positive correlation with PM_2.5 concentrations; (3) Moisture content exhibits complex, non-linear impacts that vary by phenological stage. These results underscore the multifactorial nature of biomass burning emissions, necessitating integrated analysis approaches for comprehensive pattern elucidation. Future work should employ multivariate regression to quantify relative factor contributions.

3.5. LightGBM Analysis

LightGBM is an open-source framework for implementing Gradient Boosting Decision Tree (GBDT) [50,51]. A decision tree is top-down tree-structured algorithm that splits to select the optimal partitioning attributes through impurity measures such as information gain to obtain optimal classification results. LightGBM introduces gradient-based one-side sampling (GOSS) and exclusive feature bundling (EFB) [52] to efficiently process big data. GOSS focuses on instances with more significant gradients, ensuring that the model focuses more on harder-to-fit data. GOSS focuses on instances with more significant gradients, ensuring that the model focuses more on harder-to-fit data points. Furthermore, EFB diminishes the feature count by aggregating mutually exclusive features. Additionally, LightGBM employs a histogram-based algorithm to discretize the data and constructs the tree using an incremental leaf-wise growth methodology, thereby facilitating quicker convergence and superior performance. LightGBM is primarily utilized to address the issue of diminished training speed inherent in traditional GBDT when dealing with large-scale datasets.

The essence of the LightGBM algorithm lies in the iterative training of a weak classifier to achieve an optimal model.

3.5.1. Flowchart

Initially, the LightGBM regression model (Figure 7) is where the independent variable X encompasseds the period, site, fuel load, moisture content, and tree species, while the dependent variable Y representeds the mass concentration of PM_2.5; Subsequently, the feature importance is as computed using the established LightGBM model. Thereafter, the constructed LightGBM regression model was applied to both training and the test datasets, and to obtain the outcomes results were obtained of model evaluation. Unlike traditional models that yield a definitive equation, LightGBM typically assesses the model based on the prediction accuracy of the test data. In this paper, the LightGBM regression model is evaluated using metrics such as MSE, RMSE, MAE, MAPE, R² indicators.

The model evaluation metrics include MSE (Mean Square Error) quantifying the average squared deviation between predicted and actual values, with lower values indicating better accuracy; RMSE (Root Mean Square Error), as the square root of MSE, reflectings the actual magnitude of errors while maintaining the same interpretation principle; MAE (Mean Absolute Error) evaluating the model performance by calculating the mean absolute deviation; MAPE (Mean Absolute Percentage Error), expressed in percentage form, is particularly useful for relative error assessment; R² (Coefficient of Determination) measure the model’s explanatory power by comparing predictions with the data mean, ranging from 0 to 1 where values closer to 1 denote better fit. Together these metrics form a comprehensive evaluation framework providing multidimensional references for model optimization.

3.5.2. Detailed Description

The modeling process consisted of three key steps:

(1)

Data preprocessing: The dataset of 3240 samples was split into training (70%) and testing (30%) sets using sklearn’s train_test_split function (random_state = 42 for reproducibility)

(2)

Hyperparameter optimization: implemented through randomized search

(3)

Feature encoding: Numerical features underwent outlier treatment, while categorical features were transformed via one-hot encoding

The parameter configuration of the LightGBM regression model is shown in

Table 2. Model parameters.

Parameter Name	Parameter Value
Training time	0.421 s
Data segmentation	0.7
Data shuffling	Yes
Cross validation	5
Base learner	gbdt
Number of base learners	300
Learning rate	0.15
L1 regular term	0.5
L2 regular term	0
Sample sampling rate	0.5
Tree feature sampling rate	1
Node split threshold	0
Minimum weight of samples in leaf nodes	0.001
Maximum tree depth	11
Minimum number of samples for leaf nodes	20

. Key settings include: a training time of 0.421 s, 7:3 data splitting ratio (70% training set and 30% test set) with data shuffling enabled; GBDT (Gradient Boosting Decision Tree) was selected as the base learner with 300 base learners (i.e., iteration rounds) and a learning rate of 0.15. For regularization, L1 (Lasso) was set to 0.5, while L2 (Ridge) was 0; Sampling strategy parameters including sample selection rate and feature sampling rate were configured; For tree structure control, key parameters such as node split threshold, minimum leaf sample weight (min_child_weight), maximum tree depth and minimum samples in leaf nodes were specified. This parameter combination ensures modeling efficiency while effectively controlling overfitting risks through regularization and sampling strategies. These parameters are determined by cross-validation and hyperparameter tuning to ensure the stability of the model. Table 2 provides a detailed reference for reproducing this experiment.

Figure 8 presents the feature importance analysis from the LightGBM regression model, highlighting significant variations in individual feature contributions to the prediction performance. Notably, the feature importance of tree species, at 36.2%, underscores its paramount role in the model’s predictive capabilities, serving as the central factor in explaining variations in the target variables. This was closely followed by moisture content with 21.60% importance, showing that this characteristic has a significant impact on the model and its variation is able to predict the results. The significance of temporal and locational characteristics was determined to be 18.40% and 17.10%, respectively. Although these values were lower than those for tree species and moisture content, they remained significant predictors that should not be overlooked. Their contribution to the model’s predictive capacity was notably substantial, highlighting the impact of temporal and locational factors on the target variable. In comparison, the contribution of fuel load was found to be merely 6.7%, the smallest among the listed characteristics. This suggests that, within the current model and dataset, the direct influence of fuel load on the predictive outcomes is relatively minor.

The evaluation metrics presented in

Table 3. Results of model evaluation.

	MSE	RMSE	MAE	MAPE	R2
training set	548,867.201	740.8557	537.1264	6.1105	0.988
cross-validation set	6,460,880.09	2536.101	1777.268	16.176	0.912
test set	1,331,578.99	1153.940	852.2080	9.3805	0.973

quantitatively demonstrate the predictive performance of the LightGBM model across three distinct datasets: the cross-validation set (reflecting model generalization), training set (indicating fitting accuracy), and test set (representing real-world prediction capability). Notably, the cross-validation metrics provide real-time feedback for hyperparameter tuning, ensuring the development of a robust and stable predictive model. The evaluation outcomes of the GBM regression model reveal the model’s efficacy across various datasets. Within Table 3, the metrics of MSE, RMSE, MAE, and MAPE are explicitly presented, each with a mutual correlation. Given the substantial magnitude of the MSE, RMSE, and MAE figures, differentiation among them is challenging. Consequently, attention is directed towards the MAPE values, which are 6.11%, 16.1%, and 9.38% for the training, cross-validation, and test sets, respectively. This error is comparatively minor. The coefficient of determination (R²) attains values of 0.988, 0.912, and 0.973 within the training set, cross-validation set, and test set, respectively, signifying the model’s robust fit to the data, thus indicating a relatively satisfactory model performance.

From Figure 9, it can be seen that the fitting situation of the predicted have real values had a high degree of overlap, which indicates that LightGBM has a better prediction of the test data with a higher accuracy.

4. Discussion

4.1. Variability in PM_2.5 Emissions Among Broadleaf Species

(1)

Our comparative analysis revealed substantial interspecies variation in PM_2.5 emission characteristics among the studied broadleaf species (p < 0.05). Ulmus laciniata (UL) consistently demonstrated the lowest emission rates across all experimental conditions, while Prunus maackii (PM) exhibited the highest emission outputs. These differences likely stem from variations in leaf morphology and biochemical composition [31,53].

(2)

At the tree parts level, we observed distinct emission patterns: foliar components accounted for over 40% of total PM_2.5 emissions, whereas branch emissions remained consistently below 30% of the total output. This disparity may be attributed to the greater surface-area-to-volume ratio and higher volatile content in leaves compared to woody tissues [54].

4.2. Moisture-Dependent Emission Dynamics

(1)

The moisture-PM_2.5 emission relationship followed a characteristic unimodal pattern, with minimum emissions at 0% moisture content and peak emissions consistently occurring at 15% moisture across all study species. This optimal moisture range likely represents a balance between sufficient water content to facilitate combustion while avoiding excessive moisture that would suppress burning efficiency [55].

(2)

Notably, Acer tegmentosum (AT), Ulmus laciniata (UL), and Prunus maackii (PM) displayed a unique linear response to increasing moisture content (0%–20%), contrasting with the unimodal pattern observed in other species. This divergence suggests fundamental differences in combustion physiology that warrant further investigation.

4.3. Phenological Influences on Emission Patterns

(1)

The budding period (A) exhibited a straightforward positive correlation between moisture content and PM_2.5 emissions. In contrast, both the growing period (B) and defoliation period (C) displayed complex triphasic (“decline-rise-decline”) emission patterns. These phase-specific responses likely reflect seasonal changes in plant physiology [55], including variations in stomatal conductance [56] and secondary metabolite production [57].

(2)

During the defoliation period (C), we observed a 30%–50% amplification in emission variability compared to other phenological stages. This enhanced variability may result from the combined effects of leaf senescence processes [58] and changes in tissue water distribution [59] during autumn.

5. Conclusions

5.1. Key Findings

(1)

This study establishes that PM_2.5 emissions from broadleaf tree combustion are governed by a complex interplay of species-specific traits, moisture content, and phenological stage. Our results demonstrate that leaves with foliar components are consistently the dominant emission source.

(2)

The identification of 15% moisture content as the peak emission condition across multiple species provides crucial data for wildfire emission modeling and forest management strategies. This finding has particular relevance for predicting air quality impacts during periods of moderate drought conditions.

5.2. Practical Implications

(1)

Our results suggest that wildfire mitigation efforts should prioritize monitoring and management during the defoliation phase (C), particularly when moisture conditions approach the 15% threshold. This combination of factors represents the highest risk scenario for substantial PM_2.5 release.

(2)

The developed LightGBM model, with its demonstrated predictive accuracy for within-range conditions (R² = 0.97), offers a valuable tool for regional air quality forecasting during fire seasons. However, users should exercise caution when extrapolating beyond the studied moisture range (0%–20%).

5.3. Future Research Directions

(1)

Subsequent studies should incorporate detailed characterization of leaf surface properties and biochemical composition to better explain the observed interspecies variation in emission patterns.

(2)

We recommend expanding the experimental scope to include a wider range of moisture conditions (particularly, 20%–30%) to improve model generalizability. Parallel studies should investigate the chemical composition of emitted particulates to assess health and climate impacts.