An Interpretable Ensemble Learning Framework Based on Remote Sensing for Ecological–Geological Environment Evaluation: The Case of Laos

Highlights

What are the main findings?

• An interpretable stacking ensemble framework integrating seven heterogeneous classifiers significantly enhanced ecological–geological quality prediction in Laos, outperforming all individual models (accuracy = 91.14%, recall = 96.71%, F1 = 93.62%, AUC = 95.05%) while ensuring transparency through SHAP-based interpretation of variable contributions.
• Applied at the national scale, the framework produced high-resolution ecological–geological zoning maps that identified stable plains in the northeast and southeast and highlighted vulnerable zones around metropolitan areas and tectonically active mountainous regions.

What is the implication of the main finding?

• The results provide decision-ready evidence for ecological governance, risk identification, and sustainable development in Laos, demonstrating clear regional management value.
• Beyond this case study, the framework establishes a transferable, data-driven paradigm for ecological–geological assessment in tropical mountainous and data-scarce regions, offering methodological guidance with global relevance.

Abstract

As a critical ecological security barrier in the Indo-China Peninsula, the Lao People’s Democratic Republic (Lao PDR) is increasingly threatened by forest degradation, frequent geological hazards, and intensified anthropogenic disturbances. To address the urgent need for a scientific evaluation of eco-geological environmental quality, this study develops a comprehensive assessment framework integrating multi-source remote sensing imagery, geological maps, and socio-economic datasets. A total of ten indicators were selected across four dimensions—geology, topography, ecology, and human activity. A stacking ensemble learning model was constructed by combining seven heterogeneous base classifiers—AdaBoost, KNN, Gradient Boosting, Random Forest, SVC, MLP, and XGBoost—with a logistic regression meta-learner. Model interpretability was enhanced using SHAP values to quantify the contribution of each input variable. The stacking model outperformed all individual models, achieving an accuracy of 91.14%, an F1 score of 93.62%, and an AUC of 95.05%. NDVI, GDP, and slope were identified as the most influential factors: vegetation coverage showed a strong positive relationship with environmental quality, while economic development intensity and steep terrain were associated with degradation. Spatial zoning results indicate that high-quality eco-geological zones are concentrated in the low-disturbance plains of the northeast and southeast, whereas vulnerable areas are primarily distributed around the Vientiane metropolitan region and tectonically active mountainous zones. This study offers a robust and interpretable methodological approach to support ecological diagnosis, zonal management, and sustainable development in tropical mountainous regions.

1. Introduction

Located in the heart of the Indo-China Peninsula, the Lao People’s Democratic Republic (Lao PDR) is part of the Indo-Burma biodiversity hotspot. The country features high ecological heterogeneity, complex geological structures, and a fragile environmental foundation that is extremely sensitive to external disturbance. Approximately 78% of its territory consists of mountainous and hilly terrain, with pronounced elevation gradients and a high susceptibility to geological hazards. These characteristics make Lao PDR’s ecological systems highly sensitive to external disturbances [1]. In recent years, infrastructure development and agricultural expansion have accelerated, leading to intensified deforestation, soil erosion, and landslides. According to recent reports, the country’s annual forest loss rate has reached 1.5%, with clear evidence of ecological degradation, reduced carbon sink capacity, disrupted hydrological processes, and increasing land degradation [2,3]. As a key partner in the Belt and Road Initiative, Lao PDR’s ecological security plays a critical role in maintaining regional ecological stability across Southeast Asia [4]. Therefore, developing a robust and reproducible eco-geological assessment mechanism is both scientifically necessary and regionally urgent for environmental risk identification, sustainable resource management, and policy-making [5]. Given the increasing demand for science-based environmental management in transnational initiatives such as the Belt and Road Initiative, this framework also holds practical relevance for regional cooperation and sustainable development planning across Southeast Asia.

In this context, establishing a reliable and scalable eco-geological assessment framework is urgently needed. Such a framework enables the scientific diagnosis of environmental degradation patterns, supports the zoning of ecologically vulnerable areas, and informs early warning and decision-making systems. It is particularly valuable in tropical mountainous regions like Lao PDR, where the interplay between fragile ecology and active tectonics creates complex risks.

Ecological and geological environment assessment refers to the integrated evaluation of ecological status and geological conditions, serving as a vital tool for assessing regional carrying capacity and providing early warnings of potential environmental risks. Traditional methods—such as Principal Component Analysis (PCA), the Analytic Hierarchy Process (AHP), the Entropy Weight Method, and Fuzzy Comprehensive Evaluation (FCE)—are widely applied in ecological studies [6,7,8,9]. These approaches rely on index construction, weight assignment, and composite calculation to classify environmental quality levels. Although methodologically mature and operationally simple, traditional models are limited by assumptions of linearity and subjectivity in weight determination. They struggle to capture complex non-linear interactions among geological, ecological, and anthropogenic factors and are often inadequate in processing high-dimensional, heterogeneous datasets such as remote sensing imagery and geological maps [10,11]. Recent research has highlighted the need for spatially explicit, data-driven methods in large-scale environmental assessments—for example, efforts to predict the historical and future dynamics of soil organic carbon across China using gridded geospatial models [12]. In tropical mountainous regions with intricate landforms and diverse formation mechanisms, these limitations hinder the ability to accurately identify ecological risk zones or extract dominant controlling factors [13,14].

By contrast, machine learning (ML) algorithms have recently demonstrated strong modeling capabilities and adaptability in ecological assessment tasks, particularly under high-dimensional, multi-source, and non-linear data scenarios. Unlike conventional linear models, ML techniques—such as Random Forests (RFs), Support Vector Machines (SVMs), and neural networks—can autonomously learn complex patterns and feature interactions from data without prior assumptions [15,16,17]. These methods have shown excellent generalization performance in applications including land use monitoring, ecosystem service evaluation, and geological hazard prediction [18,19]. Recent studies have also explored the role of ensemble learning and spectral feature engineering in ecological monitoring, such as predicting soil organic matter using fractional-order derivatives and customized vegetation indices [20] and monitoring saline vegetation via hyperspectral classification and Gaussian mixture models [21]. Moreover, ensemble learning approaches, which integrate multiple base learners, have emerged as powerful tools for improving model robustness and generalization [22,23]. In remote sensing and environmental mapping specifically, stacked generalization (stacking) has repeatedly demonstrated gains in accuracy and stability—for example, in national-scale forest change mapping [24], advancing flood susceptibility modeling [25], and landslide susceptibility assessment in complex terrains [26].

With the advancement of interpretability techniques such as SHapley Additive exPlanations (SHAP), the “black-box” nature of ML models has been increasingly addressed. SHAP enables the quantification of each variable’s contribution, direction, and marginal effect, thereby supporting more transparent and interpretable ecological modeling [27]. This aligns with the broader movement toward explainable artificial intelligence (XAI), which promotes responsible and accountable data-driven modeling [25,28]. Recent SHAP-based studies have further demonstrated its capacity to explain vegetation–climate interactions and to support soil salinity mapping from multi-temporal Sentinel-2 data [29,30].

Building upon these developments, this study critically reviews existing eco-geological assessment methodologies and develops a comprehensive index system consisting of ten indicators. These indicators were derived from multi-source datasets, including remote sensing products such as Normalized-Difference Vegetation Index (NDVI), Normalized-Difference Water Index (NDWI), and land use type, as well as geological maps and socio-economic statistics. Seven base learners—AdaBoost, KNN, Gradient Boosting, Random Forest, SVC, MLP, and XGBoost—were integrated into a stacking ensemble model with a logistic regression meta-learner. SHAP was applied to analyze model interpretability and identify key controlling factors of eco-geological quality [31,32,33]. The final output is a high-accuracy, interpretable, and transferable assessment framework tailored to the complex ecological conditions of tropical mountainous regions. This framework provides scientific support for the identification of vulnerable ecological zones, ecological governance, and informed policy-making in Lao PDR and similar ecological hotspots. Our design choice is consistent with prior remote sensing studies showing that stacking can synthesize heterogeneous learners to improve mapping accuracy and generalization [34].

To address persistent limitations in ecological–geological quality assessment, this study presents an interpretable ensemble learning framework specifically designed for tropical mountainous regions. The framework introduces three key innovations. First, it establishes a spatially explicit evaluation system that integrates remote sensing-derived vegetation and moisture indices (e.g., NDVI, NDWI), geological distance metrics, topographic gradients, and anthropogenic disturbance indicators within a unified 2 km × 2 km spatial grid. This design ensures consistent spatial resolution and facilitates detailed characterization of heterogeneous landscapes. Second, the framework employs a stacking ensemble model that integrates diverse base classifiers (e.g., RF, SVM, XGBoost), thereby enhancing predictive robustness and minimizing the biases and structural assumptions associated with conventional methods such as the AHP, PCA, and entropy-based weighting. Third, it incorporates SHAP (SHapley Additive exPlanations) to provide transparent quantitative interpretations of model predictions by attributing importance to individual predictors. This interpretability bridges the gap between black-box machine learning outputs and actionable ecological insights. Overall, the proposed framework offers a scalable, transferable, and fully data-driven solution for environmental quality assessment in ecologically fragile and data-scarce regions.

2. Study Area and Data

2.1. Study Area

Lao People’s Democratic Republic (Lao PDR) is located between 13°54′N and 22°05′N latitude and 100°10′E and 107°30′E longitude (Figure 1) [35]. As the only landlocked country in the Indo-China Peninsula, it borders China to the north, Cambodia to the south, Vietnam to the east, Myanmar to the northwest, and Thailand to the west. The country covers an area of approximately 240,900 km² [35]. Geographically, more than 80% of Lao PDR is dominated by mountainous and plateau terrain, with the remaining areas comprising hills and lowland plains [36].

The climate is classified as tropical to subtropical monsoon, with an annual average temperature of around 25 °C and an average annual precipitation of approximately 1800 mm [36]. The year is divided into two distinct seasons: the rainy season (May to October), which accounts for the majority of the annual rainfall and is characterized by high humidity, and the dry season (November to April), which is dominated by dry northeasterly monsoons and contributes only about 13% of the total annual precipitation [36].

Lao PDR currently has a forest coverage rate of approximately 53%, but this has been declining in recent years due to agricultural expansion, logging, and timber trade [37]. According to the World Bank [38], the country’s total population reached approximately 7.95 million in 2024, with a per capita GDP of about USD 2648 and a population density of 30 people per square kilometer. These figures reflect significant spatial disparities between socio-economic development and ecological conditions across the country.

2.2. Construction of the Evaluation Indicator System

The scientific selection of indicators and the rational design of the evaluation system are critical prerequisites for assessing the eco-geological environment. Most existing studies follow four fundamental principles—comprehensiveness, scientific validity, relevance, and data availability—to ensure that the assessment results are systematic, operational, and accurate [39].

Grounded in eco-geological theory and supported by previously collected geological maps, tectonic data, and remote sensing imagery of Lao PDR (Table 1), we reviewed prior studies [40] and identified four primary categories of influencing factors:

• Geological conditions: Geological features fundamentally determine the stability and carrying capacity of ecosystems. Relevant factors include stratigraphic structure, lithology, fault systems, geological hazard risks, and groundwater distribution [41,42].
• Topographic conditions: Topography serves as a foundational coupling interface between geological and ecological processes. It plays a central role in controlling hydrological processes, vegetation distribution, and the suitability of human development. The most commonly used indicators—elevation, slope, and aspect—are widely adopted in the analysis of mountain ecosystem stability [43].
• Ecological conditions: Ecological indicators reflect the environmental baseline and recovery potential of a region. These include vegetation status, surface water distribution, and land use type. NDVI, NDWI, and land use type are commonly used metrics for ecological quality monitoring and are especially suitable for remote sensing-based spatial modeling [44,45].
• Anthropogenic conditions: Human activities exert cumulative and spatially heterogeneous impacts on eco-geological systems. These are typically represented by population density, GDP, and land development intensity [46,47,48]. In areas with rapid population growth and intensive resource exploitation, anthropogenic factors are often the primary drivers of ecological degradation.

Based on eco-geological theory and existing research on ecosystem evaluation, and considering indicator availability, spatial representativeness, and ecological sensitivity, a total of ten secondary indicators were selected to construct the final comprehensive evaluation system. The indicator selection was informed by widely applied variables in recent studies on geological hazard assessment [45], ecosystem service monitoring [44], and environmental carrying capacity modeling [49], with an emphasis on quantifiability and the feasibility of remote sensing extraction [40,43].

This indicator system aims to integrate theoretical rigor with practical applicability, enabling the effective fusion of multi-source data and supporting spatial modeling. It also provides a robust foundation for subsequent machine learning analysis and SHAP-based interpretability assessments.

Table 1. Evaluation Indicators for Comprehensive Ecological and Geological Environmental Assessment in Lao PDR.

Primary Category	Secondary Indicator	Indicator Description
Geological Conditions	Fault Distance	Describes the density of fault points within the region, indicating the intensity of tectonic activity.
	Lithology	Represents the distribution of lithology units as fundamental geological structures, indicating lithology variations.
Topographic Conditions	Elevation	Influences climate, hydrology, and vegetation distribution, shaping the regional ecological pattern.
	Slope	Describes the steepness of the terrain, affecting soil erosion and geological hazard risks.
	Aspect	Determines sunlight duration and intensity, influencing vegetation growth and soil moisture.
Ecological Conditions	NDWI	Reflects water resource abundance and affects hydrological cycles and ecosystem stability.
	NDVI	Measures surface vegetation coverage, reflecting the overall health of the ecosystem.
	Land Use Type	Represents land cover categories derived from satellite imagery (e.g., forest, cropland, urban), reflecting patterns of human land use and ecological change.
Anthropogenic Conditions	Population Density	Reflects the concentration of human settlement, indicating resource consumption and environmental pressure.
	GDP	Describes economic productivity, reflecting resource use, environmental pollution, and ecological change.

Table 1. Evaluation Indicators for Comprehensive Ecological and Geological Environmental Assessment in Lao PDR.

Primary Category	Secondary Indicator	Indicator Description
Geological Conditions	Fault Distance	Describes the density of fault points within the region, indicating the intensity of tectonic activity.
	Lithology	Represents the distribution of lithology units as fundamental geological structures, indicating lithology variations.
Topographic Conditions	Elevation	Influences climate, hydrology, and vegetation distribution, shaping the regional ecological pattern.
	Slope	Describes the steepness of the terrain, affecting soil erosion and geological hazard risks.
	Aspect	Determines sunlight duration and intensity, influencing vegetation growth and soil moisture.
Ecological Conditions	NDWI	Reflects water resource abundance and affects hydrological cycles and ecosystem stability.
	NDVI	Measures surface vegetation coverage, reflecting the overall health of the ecosystem.
	Land Use Type	Represents land cover categories derived from satellite imagery (e.g., forest, cropland, urban), reflecting patterns of human land use and ecological change.
Anthropogenic Conditions	Population Density	Reflects the concentration of human settlement, indicating resource consumption and environmental pressure.
	GDP	Describes economic productivity, reflecting resource use, environmental pollution, and ecological change.

2.3. Data Sources

The geological data used for evaluating geological conditions were derived from the China–ASEAN Geoscience Information Big Data Platform, with a map scale of 1:1,000,000. The indicators related to topographic, ecological, and anthropogenic conditions were extracted through the interpretation of remote sensing imagery. Specifically, 26 scenes of Landsat 8 imagery acquired between November 2023 and February 2024 (the dry season) were selected to minimize cloud cover and reduce the effects of vegetation phenology on land cover classification. In addition, 40 scenes of ASTER GDEM data were used to extract topographic variables. Detailed information on data sources and spatial resolutions is provided in

Table 2. Data Sources and Resolution Information for Topographic Indicators Used in the Environmental Assessment.

Primary Indicator	Secondary Indicator	Resolution	Data Source	Link
Topographic Conditions	Elevation	30 m	ASTER GDEM	http://lpdaac.usgs.gov/
	Slope	30 m	ASTER GDEM	http://lpdaac.usgs.gov/
	Aspect	30 m	ASTER GDEM	http://lpdaac.usgs.gov/
Ecological Conditions	NDWI	30 m	Landsat 8	https://earthexplorer.usgs.gov/
	NDVI	30 m	Landsat 8	https://earthexplorer.usgs.gov/
	Land Use Type	30 m	Landsat 8	https://earthexplorer.usgs.gov/
Anthropogenic Conditions	Population Density	100 m	WorldPop	https://www.worldpop.org/
	GDP	National	WorkBank	https://data.worldbank.org/

.

2.4. Indicator Extraction and Analysis

Quantitative extraction of evaluation indicators was achieved through remote sensing imagery. In this study, Google Earth Engine (GEE), ENVI5.6, and ArcGIS Pro 3.2 were employed as the primary platforms for remote sensing interpretation. Preprocessing steps included the selection of high-quality imagery with cloud coverage less than 5% and no visible haze, striping, or sensor-induced anomalies. GEE was used in the initial stage for previewing image availability, screening cloud-free scenes, and inspecting spatial coverage. All subsequent radiometric and atmospheric correction (FLAASH), as well as image fusion to 30 m resolution, was performed manually using ENVI to ensure spatial consistency and completeness across the study area. ArcGIS was used for seamless mosaicking and clipping based on administrative boundaries of Laos. Vegetation and water indices, such as NDVI and NDWI, were calculated in ENVI and exported as GeoTIFF layers [44].

Topographic factors including slope, aspect, and elevation were derived from ASTER GDEM data at a 30 m resolution [28]. Land use type was based on the 2020 GlobeLand30 product [50], which provides 30 m global land cover categories derived from satellite imagery. The entire preprocessing pipeline was modularly designed: GEE was used for initial filtering and visualization; ENVI handled image correction, pan-sharpening, and index calculation; and ArcGIS supported layer mosaicking, reprojection, clipping, and raster alignment (Figure 2). All layers were standardized in terms of coordinate system and spatial resolution, with vector-to-raster conversions applied as needed to ensure a unified data structure for subsequent modeling.

All spatial layers were reprojected to WGS 84/UTM Zone 48N (EPSG:32648) using ArcGIS tools to ensure cross-platform consistency. Bilinear interpolation was applied to continuous raster layers during reprojection to preserve gradient smoothness, while nearest-neighbor resampling was used for categorical data. To facilitate model training, all indicator values were normalized to a (0, 1) range using min–max scaling to ensure comparability across heterogeneous variables. Slope, aspect, and elevation layers were derived from ASTER GDEM (30 m) using ArcGIS’s Spatial Analyst tools, while NDVI and NDWI were calculated from Landsat 8 OLI bands (Band 5—NIR, Band 4—Red, and Band 3—Green) within ENVI using standard index formulas. Each raster was clipped to the Laos boundary and aligned to a common grid prior to interpolation. Following interpolation (IDW), all layers were converted to tabular format through zonal aggregation over the 2 km fishnet grid. This modular pipeline guarantees reproducibility and supports integration of multi-source indicators into a unified, modeling-ready dataset.

After completing the extraction and standardization of all eco-geological evaluation factors, spatial interpolation and visualization were conducted using ArcGIS to further reveal the spatial variability of each indicator and its potential geo-ecological significance across the study area. Ten standardized indicators were interpolated and mapped to provide an intuitive spatial representation of their distribution patterns (Figure 3).

Figure 3 illustrates the spatial distribution patterns of ten eco-geological evaluation indicators across Lao PDR, reflecting the regional heterogeneity in four dimensions: natural landforms, geological conditions, ecological status, and human activity.

From the perspective of topographic indicators, the elevation and slope layers jointly reveal the pronounced undulating terrain of Lao PDR. The northern regions (e.g., Phongsaly and Luang Prabang) and eastern areas (e.g., Xayaboury and Sekong) are dominated by mountainous and hilly landscapes, where both elevation and slope values are high—slope angles often exceed 20°, forming typical steep-slope environments with relatively low geological stability. The aspect layer shows a spatially mixed pattern, with southeast-, south-, and west-facing slopes prevailing. This suggests the influence of east–west-oriented tectonic deformation and terrain incision. Aspect plays a critical role in hydrothermal distribution and vegetation growth; notably, south-facing slopes often overlap with high-NDVI zones, likely due to increased solar radiation and stronger vegetation regeneration capacity.

Ecological indicators also display clear spatial differentiation. NDVI (Normalized-Difference Vegetation Index) is generally higher in the northeastern, southeastern, and northern border regions (e.g., Phongsaly and Houaphanh), indicating robust forest cover and a stable ecological foundation. By contrast, the southwestern urban corridor and central agricultural areas exhibit low NDVI values, reflecting intense anthropogenic disturbances and elevated ecological stress. The NDWI (Normalized-Difference Water Index) layer highlights high values in the central and southern floodplain areas, particularly along the Mekong River and its tributaries, indicating abundant surface water resources that are vital for sustaining regional ecosystems.

In terms of geological indicators, the fault distance layer reflects the intensity of tectonic activity. Dense fault zones are observed in the southwest, north, and central high-slope areas, marking them as geologically sensitive zones. By contrast, the eastern and certain lowland regions exhibit sparse fault distribution, indicating relatively stable tectonic settings. Lithological distance shows a more scattered pattern; northern and northeastern areas are widely proximal to lithological units with high rock hardness and rugged terrain, implying greater difficulty of ecological restoration.

Anthropogenic indicators such as GDP and population density clearly delineate the core zones of socio-economic activity. The Vientiane metropolitan area, Pakse, and Savannakhet exhibit GDP values exceeding USD 200,000/km², indicating high urbanization levels and dense artificial land use. Population density follows a similar trend: southwestern areas and urban peripheries commonly exceed 40 persons/km², in stark contrast to mountainous regions with sparse distributions below 20 persons/km². The strong spatial coupling between human activity intensity and ecological risk makes anthropogenic disturbance a key driver of environmental degradation.

In summary, the eco-geological environment of Lao PDR demonstrates a spatial pattern characterized by “high in the north, low in the south; dense in the east, sparse in the west; strong mountains, sparse population; clustered valleys, concentrated settlements.” The topographic and geological context defines the ecological baseline, while anthropogenic pressures compound ecological risk distribution.

3. Methods

This study proposes an ensemble learning framework for assessing ecological–geological quality by integrating multi-source geospatial data. The workflow consists of four key stages: (1) data preprocessing and indicator extraction from remote sensing, geological, and socio-economic sources; (2) construction of a gridded ecological–geological dataset and stratified sample labeling; (3) development of a stacking ensemble model combining seven base classifiers with logistic regression as a meta-learner; and (4) model interpretation using SHAP values and spatial zoning based on predicted probabilities. This approach is designed to capture complex nonlinear interactions while providing interpretable and transferable predictions suitable for regional environmental planning.

3.1. Base Classification Models

This study employed seven representative classification algorithms as base learners: AdaBoost, KNN, Gradient Boosting (GB), Random Forest (RF), Support Vector Classifier (SVC), Multi-Layer Perceptron (MLP), and Extreme Gradient Boosting (XGBoost). AdaBoost enhances overall model performance by iteratively combining weighted weak learners, making it suitable for moderate-sized, nonlinearly separable datasets [51]. KNN classifies samples based on the labels of nearest neighbors without explicit training, serving as a simple and effective baseline method [52]. GB improves prediction accuracy by fitting residuals iteratively and excels in modeling complex nonlinear relationships [53]. RF, built on the Bagging framework, constructs an ensemble of decision trees with randomized feature selection and bootstrap sampling, offering strong generalization and robustness [4]. SVC constructs an optimal hyperplane in high-dimensional space using kernel functions, ideal for tasks with clear decision boundaries and limited samples [54]. MLP, a typical feedforward neural network, captures complex nonlinear interactions among eco-geological indicators through nonlinear activation and backpropagation [55]. XGBoost extends GB by incorporating regularization, pruning, and parallel computing, significantly improving computational efficiency and overfitting resistance, and has been widely applied in ecological modeling [7].

3.2. Stacking

Stacking is a two-layer ensemble learning framework initially proposed by Wolpert in 1992 [56]. Its core idea is to integrate multiple base learners with complementary strengths by introducing a meta-learner that learns from the outputs of the base models to generate improved predictions. Unlike Bagging and Boosting, which emphasize model homogeneity and resampling mechanisms, stacking focuses on combining heterogeneous models and is well suited for complex heterogeneous multi-source data structures [57].

In practice, a single model often struggles to fully capture the multi-level patterns inherent in ecological–geological datasets. For instance, tree-based models such as Random Forest (RF) and Gradient Boosting (GB) are known for their robustness to noise and their ability to handle high-dimensional, nonlinear feature spaces. Support Vector Machines (SVMs) are effective for margin-based classification and perform well on smaller datasets, although their scalability can become a limitation when kernel functions are applied to large training sets. Neural networks offer powerful non-linear modeling capabilities but require careful hyperparameter tuning and high-quality training samples. K-Nearest Neighbors (KNN), although intuitive and simple, tends to overfit in high-dimensional spaces due to the “curse of dimensionality” [58]. Therefore, the stacking strategy effectively mitigates these limitations by integrating structurally diverse models.

In this study, the stacking model consists of two hierarchical layers:

A.

Base Learner Layer This layer includes seven classifiers with diverse theoretical foundations: AdaBoost, K-Nearest Neighbors (KNN), Gradient Boosting (GB), Random Forest (RF), Support Vector Classifier (SVC), Multi-Layer Perceptron (MLP), and Extreme Gradient Boosting (XGBoost). These models represent various paradigms, including tree-based methods, kernel machines, distance-based learning, and deep learning. Together, they capture complex internal structures in the eco-geological feature space from multiple perspectives.

B.

Meta-Learner Layer Logistic regression is employed as the meta-learner, aiming to minimize the following loss function:

$$\begin{array}{c}\mathcal{L}=-{\displaystyle \frac{1}{N}}{\displaystyle \sum _{i=1}^N}\left[y_i\log ({\widehat{y}}_i)+(1-y_i)\log (1-{\widehat{y}}_i)\right]\\ \mathrm{Subject} \mathrm{to} {\widehat{y}}_i=\sigma \left({\displaystyle \sum _{j=1}^M{w}_j\cdot h_j\left({\mathbf{x}}_i\right)}\right)\end{array}$$

(1)

$\sigma \left(\cdot \right)$ denotes the Sigmoid function, and $h_j\left({\mathbf{x}}_i\right)$ is the predicted probability of the j-th base model for sample ${\mathbf{x}}_i$. $M=7$ denotes the number of base models.

Model Training Procedure:

Data Splitting and Base Model Training: The dataset $D={\left\{\left(x_i,\hspace{0.33em}{y}_i\right)\right\}}_{i=1}^N$ is split into a training set ${\mathcal{D}}_{\mathrm{train}}$ and a test set ${\mathcal{D}}_{\mathrm{test}}$ using a 7:3 ratio. In the training phase, the seven base models are independently trained to output predicted labels ${\widehat{y}}_i^{(j)}$ and corresponding probabilities $p_i^{(j)}\in [0,1]$, where j = 1, …, 7.

Construction of Meta-Learner Training Data: To avoid overfitting, 5-fold cross-validation is applied to the training set. For each sample, a cross-validated prediction probability vector is obtained:

$$z_i=\left[p_i^{(1)},p_i^{(2)},\dots ,p_i^{(7)}\right]$$

(2)

Each vector ${\mathbf{z}}_i$ is then paired with its true label $y_i$, forming the new meta-training dataset:

$${\mathcal{D}}_{meta}={\left\{\left({\mathbf{z}}_i,y_i\right)\right\}}_{i=1}^{N_{train}}$$

(3)

Meta-Learner Training and Fusion: A logistic regression model is trained on ${\mathcal{D}}_{meta}$ to learn the mapping from base model probabilities to final prediction outcomes. The resulting ensemble model is expressed as

$${\widehat{y}}_i=\sigma \left({\mathbf{w}}^{\mathrm{T}}{\mathbf{z}}_i+b\right)$$

(4)

Test Set Prediction and Evaluation: On the test set ${\mathcal{D}}_{test}$, the prediction probability vector ${\mathbf{z}}_i^{test}$ composed of all base model outputs is obtained and fed into the trained logistic regression model to produce the final prediction probability and label.

Full Data Prediction and Interpretability Analysis: In the final experiment, the prediction probabilities $p_i$ for all samples generated by all base models are recorded. These are used as the input for SHAP analysis to further reveal the influence mechanisms of key variables on eco-geological environment quality prediction and support spatial visualization of the results (Figure 4).

3.3. SHAP

SHAP (SHapley Additive exPlanations) is a model-agnostic interpretability method based on additive feature attributions [27]. In this approach, all input features are considered “contributors,” and the method estimates the marginal contribution of each feature by measuring the change in the model’s output when the feature is added. This allows for the evaluation of each feature’s importance in prediction.

The SHAP value for a feature is computed by averaging its marginal contributions across all possible feature orderings. The final SHAP value reflects the extent to which each feature influences the model’s output. The equations are as follows [58]:

$$V_i=V_{i\left(base\right)}+\sum _{j=1}^{S}shap\left(x_{i,j}\right)$$

(5)

$$shap(x_{i,j})=\sum _M{\displaystyle \frac{S!}{\left|M\right|!\left(S-\left|M\right|-1\right)}}\left(V_M\cup x_{i,j}-V_{\left(M\right)}\right)$$

(6)

where $V_i$ is the predicted output for sample i, $V_{i\left(base\right)}$ denotes the expected value of the model output for sample i, $shap\left(x_{i,j}\right)$ represents the SHAP value of feature j for sample i, $S$ is the full set of input features, M is a subset of features excluding j, and $V_M\cup x_{i,j}$ and $V_{\left(M\right)}$ represent the model outputs with and without feature j, respectively.

Recent advances have further integrated SHAP into ensemble interpretation workflows, enabling visual exploration of variable effects [59] and improving model transparency in environmental prediction [14].

3.4. Data Structuring and Sample Selection Strategy

To enable machine learning modeling, all spatial indicators were standardized and converted into tabular format by gridding the study area into 2 km × 2 km cells, which resulted in 40,874 spatial units. Specifically, we used the “Create Fishnet” tool in ArcGIS to divide the entire study area into a regular grid, using the Laos national boundary as a mask. Each grid cell was assigned a unique ID and centroid coordinate, serving as the basic spatial unit for indicator aggregation and sample labeling.

To prepare input features for modeling, all indicator layers were interpolated into continuous raster surfaces using Inverse Distance Weighting (IDW), then aggregated to the 2 km grid using ArcGIS’s “Zonal Statistics as Table” tool. Mean values were extracted for continuous variables (e.g., NDVI, slope), while majority rule was applied for categorical data (e.g., land cover type), ensuring fidelity to original distributions. The final tabular dataset includes all normalized indicators and retains full traceability from raw imagery to structured machine learning inputs.

This grid cell size balances eco-geological heterogeneity and computational efficiency while mitigating overfitting risks from excessively fine grids. The 2 km unit provides a suitable compromise between data granularity and generalization, aligning well with the resolution of key remote sensing inputs (e.g., NDVI and NDWI at 30 m) and supporting reliable spatial aggregation. It also facilitates regionally meaningful interpretations for ecological planning, particularly in heterogeneous mountainous landscapes. By contrast, finer grids (e.g., 250 m) would significantly increase data volume and modeling complexity, whereas coarser grids (e.g., >5 km) may obscure localized ecological degradation or hazard features [44,60].

It is important to clarify that the 2 km × 2 km fishnet was used only as a common analysis unit to summarize variables and to structure training/validation samples; it does not change the native resolution of the input rasters. Remote sensing indices (e.g., NDVI, NDWI) were computed on the original 30 m imagery and kept at 30 m. We then computed zonal statistics within each 2 km cell from the underlying 30 m pixels and used these summaries as model inputs. This aggregation step organizes fine-resolution information into consistent regional units without any resampling or interpolation of the 30 m data. The same grid also provides a uniform spatial key to attach non-remote-sensing predictors (e.g., census-based GDP and population density), ensuring that all variables are spatially aligned for modeling.

To construct labeled samples for model training and validation, we adopted a stratified random sampling strategy based on the surface coverage of two clearly defined ecological categories: good-quality areas and degraded areas. The definition of these categories was guided by ecological and geological thresholds supported by published literature and regional statistics.

Degraded areas were defined as follows: NDVI < 0.3, indicating sparse or stressed vegetation [61,62,63]; NDWI < 0, suggesting arid or dry conditions; slope > 20°, associated with elevated risk of erosion or geo-hazards [64]; and within 1 km of mapped faults, reflecting geologically fragile zones [65].

Based on these criteria, areas belonging to the first group were labeled as positive samples, while those satisfying the second group were labeled as negative samples. These binary labels served as training targets in the supervised learning framework, as shown in Figure 5.

While positive and negative samples were defined based on specific ecological and geological thresholds, these thresholds were chosen to reflect well-established degradation conditions rather than extreme outliers. To reduce potential sampling bias, we ensured that the selected samples were distributed across a wide range of eco-geological zones with varying terrain, land cover types, and socio-economic conditions. Although not all input features were used as selection criteria, the spatial heterogeneity of the sampling framework helps preserve variability across unused indicators. This design increases the likelihood that the model remains sensitive to a broad spectrum of input variables during training and avoids overfitting to specific feature combinations.

These thresholds were further confirmed by analyzing the histogram and cumulative distribution of each indicator across the national dataset to ensure that they captured meaningful ecological distinctions while minimizing overlap. A stratified random sampling approach was applied, in which samples from each ecological category (good quality and degraded) were drawn proportionally to the surface area covered by that category across the entire study area. This ensures that both dominant and minority ecological zones are adequately represented in the training set. The dataset was then randomly split into 70% for training and 30% for testing. To ensure reproducibility and fair comparison, a fixed random seed (random_state = 42) was applied during all data splits.

To ensure adequate spatial representativeness, the labeled samples were manually selected across diverse ecological–geological settings, including mountainous, valley, and urban regions. Sampling was guided by stratified spatial principles based on terrain, land cover, and geological zoning to capture variability across different environmental conditions. Although the total number of labeled samples was limited (n ≈ 100), the selection was guided by stratified spatial principles and ecological diversity, ensuring coverage across representative mountainous, valley, and lowland environments. The number of positive and negative samples was approximately balanced (54 and 46, respectively), helping mitigate class imbalance during training. Model performance was evaluated using 10 random stratified splits to enhance robustness. Additionally, ensemble learning was adopted to further stabilize predictions under limited data conditions, especially for base models such as SVM and MLP, which are sensitive to sample scarcity.

To further enhance reproducibility and transparency, the ecological label assignment process was spatially structured by linking each labeled sample to its corresponding 2 km grid centroid. The labeled samples were organized into a tabular dataset containing spatial coordinates, grid ID, indicator values, and ecological category. All samples were referenced under a consistent coordinate system and fully aligned with the gridded multi-source indicator layers. This structured approach ensures clear traceability between the spatial units and classification labels, thereby supporting reproducibility and future comparative analyses.

3.5. Model Training, Validation, and Ensemble Integration

Based on the labeled dataset described in Section 3.4, all supervised learning models were trained using a consistent training–testing split (70/30, stratified) with a fixed random seed (random_state = 42) to ensure reproducibility and balanced class representation. Prior to modeling, input features were standardized using StandardScaler, and categorical variables such as land use type and aspect were one-hot encoded as needed.

A stacking ensemble model was subsequently constructed by integrating the outputs of seven base learners. The predictions from these base models served as inputs for a logistic regression meta-learner. This ensemble was trained on the same training set and evaluated on the testing set to ensure consistency in performance comparison. All models were implemented using the Scikit-learn (v1.3) and XGBoost (v1.7) libraries in Python.3.9. Unless otherwise specified, default hyperparameters were used for both base classifiers and the logistic regression meta-learner to ensure comparability and avoid overfitting due to excessive tuning.

4. Results

4.1. Model Performance Evaluation

To systematically evaluate the performance of ecological–geological classification models in Laos, seven machine learning models were tested, including Random Forest (RF), Gradient Boosting (GB), Support Vector Classifier (SVC), K-Nearest Neighbors (KNN), Multi-Layer Perceptron (MLP), Extreme Gradient Boosting (XGBoost), and AdaBoost. Models were assessed using accuracy, precision, recall, F1 score, and the Area Under the ROC Curve (AUC). The reported values represent the mean ± standard deviation of each metric over ten stratified random runs, ensuring robust and fair comparisons across models.

Table 3. Performance Evaluation of Classification Models.

Model	Accuracy	Precision	Recall	F1	ROC_AUC
AdaBoost	0.8943 ± 0.0487	0.8943 ± 0.0487	0.8943 ± 0.0487	0.8943 ± 0.0487	0.8943 ± 0.0487
KNN	0.8200 ± 0.0468	0.8200 ± 0.0468	0.8200 ± 0.0468	0.8200 ± 0.0468	0.8200 ± 0.0468
GB	0.9029 ± 0.0525	0.9029 ± 0.0525	0.9029 ± 0.0525	0.9029 ± 0.0525	0.9029 ± 0.0525
RF	0.8886 ± 0.0475	0.8886 ± 0.0475	0.8886 ± 0.0475	0.8886 ± 0.0475	0.8886 ± 0.0475
SVC	0.8743 ± 0.0590	0.8743 ± 0.0590	0.8743 ± 0.0590	0.8743 ± 0.0590	0.8743 ± 0.0590
MLP	0.8857 ± 0.0433	0.8857 ± 0.0433	0.8857 ± 0.0433	0.8857 ± 0.0433	0.8857 ± 0.0433
XGB	0.8971 ± 0.0780	0.8971 ± 0.0780	0.8971 ± 0.0780	0.8971 ± 0.0780	0.8971 ± 0.0780
Stacking	0.8971 ± 0.0630	0.8971 ± 0.0630	0.8971 ± 0.0630	0.8971 ± 0.0630	0.8971 ± 0.0630

summarizes the comparative results. The stacking model consistently outperformed all base models, achieving the highest accuracy (91.14%), precision (90.96%), recall (96.71%), F1 score (93.62%), and AUC (95.05%). Gradient Boosting closely followed, exhibiting slightly lower but competitive metrics. Conversely, the KNN model demonstrated notably lower performance, reflecting limitations in handling high-dimensional ecological–geological datasets.

The superior performance of the stacking model demonstrates its robustness in capturing complex eco-geological interactions, affirming its suitability for heterogeneous environments.

4.2. Model Interpretability Analysis

SHAP analysis was employed to interpret the contributions of variables influencing the ecological–geological quality predictions. Figure 6 presents the SHAP summary plot, clearly illustrating that NDVI, GDP, and slope were the three most influential indicators.

NDVI exhibited the strongest positive impact, with higher values corresponding to higher ecological–geological quality, indicating robust vegetation as a crucial stabilizing factor. GDP displayed a predominantly negative influence, confirming that intensive economic activities and urbanization adversely impact ecological conditions. Slope showed a negative relationship, emphasizing the role of steep terrains in increasing ecological fragility due to associated geological hazards and erosion risks. Other indicators, including NDWI, lithology, and fault distances, provided secondary but meaningful insights into regional ecological conditions.

The SHAP results significantly enhance the interpretability of the stacking model, clearly delineating the ecological and geological mechanisms driving spatial heterogeneity across Laos.

4.3. Regional Ecological–Geological Quality Zoning

To translate model-generated probabilities into ecological–geological quality zones, we applied an equal-interval classification scheme. The predicted probabilities from the stacking model (ranging from 0 to 1) were divided into five classes of width 0.2, corresponding to five ecological quality levels: Poor (I), Relatively Poor (II), Moderate (III), Good (IV), and Excellent (V) (

Table 4. Classification Thresholds for Ecological–Geological Quality. Classification thresholds were derived using equal-width intervals applied to the predicted probability values (range 0–1). This approach supports standardized spatial interpretation and facilitates zoning across heterogeneous regions.

Grade	Index Value Range (%)	Description
Poor (I)	0–20	Severe degradation; ecosystem functions are extremely weak or severely damaged.
Relatively Poor (II)	20–40	Lower ecological quality; moderate to strong human disturbances present.
Moderate (III)	40–60	Average ecological condition; basic recovery potential exists.
Good (IV)	60–80	Relatively stable ecosystems with well-functioning services.
Excellent (V)	80–100	Intact and well-functioning ecosystems with excellent ecological quality.

).

Equal-interval classification is a widely accepted method in remote sensing and thematic cartography, particularly for mapping continuous variables such as indices, probabilities, or suitability scores. It offers a transparent, repeatable framework that avoids subjective thresholding and supports clear spatial interpretation across heterogeneous regions. Previous studies have demonstrated its effectiveness for classifying geospatial data where interpretability and consistency are prioritized [66,67].

While these thresholds are not calibrated against field ecological measurements, they provide a practical zoning framework for regional assessment and policy support. Future efforts may explore expert-informed or data-driven approaches, such as natural breaks or fuzzy classification, to refine class boundaries when more in situ validation becomes available.

It is important to note that the five-grade classification was generated based on full-coverage prediction results from the trained stacking model, not from the training samples themselves. The initial labeled samples (good-quality and degraded areas) were only used for supervised model development. The final zonation reflects model-inferred ecological–geological quality across the entire study area, including regions not used in training, thus supporting independent spatial assessment and decision-making.

The final spatial distribution of ecological–geological quality is illustrated in Figure 7.

Poor- and Relatively Poor-quality zones were primarily concentrated around densely populated urban areas (e.g., Vientiane) and geologically sensitive mountainous regions (e.g., Savannakhet). Urban areas showed significant anthropogenic disturbance, while mountainous regions displayed pronounced geological instability.

Moderate zones typically formed transitional belts between high-quality plains and low-quality mountainous or urban areas, reflecting intermediate levels of human activity and geological stability. Good ecological zones were predominantly identified in gently sloped plains and hilly regions, characterized by adequate vegetation cover and moderate human impacts.

Excellent zones were mainly located in low-disturbance northeastern and southeastern plains, marked by stable geological settings, mild slopes, dense vegetation coverage, and minimal anthropogenic pressures. These zones represent benchmarks for ecological conservation and sustainable development.

In summary, the stacking ensemble model effectively delineated spatial patterns of ecological–geological quality in Laos, capturing intricate interactions between ecological conditions, geological factors, and human activities. These spatial insights provide a robust foundation for targeted ecological management and policy formulation.

5. Discussion

5.1. Effectiveness of the Stacking Framework

The superior performance of the stacking ensemble model across all evaluation metrics (accuracy: 91.14%, precision: 90.96%, recall: 96.71%, F1 score: 93.62%, AUC: 95.05%) demonstrates the value of integrating heterogeneous classifiers for eco-geological assessment. Unlike single models, the stacking framework combines diverse algorithmic strengths to reduce individual biases and improve generalization. Tree-based models (RF, GB, XGBoost) capture complex nonlinear relationships, kernel-based SVC provides strong margin-based classification, and MLP captures deeper interactions. The logistic regression meta-learner integrates its outputs into a stable decision boundary.

Although the performance gain over strong single models such as XGBoost and GB is numerically modest, stacking consistently achieves top scores across all metrics, particularly in recall and AUC. This consistency is critical in ecological evaluation, where both false negatives and model robustness are major concerns. By fusing classifiers with different inductive biases, stacking enhances the model’s ability to generalize across heterogeneous eco-geological landscapes and small-sample regimes. Furthermore, stacking improves the reliability of interpretation through aggregated SHAP outputs, offering more stable and representative insights into feature contributions. These benefits justify the adoption of stacking beyond mere accuracy comparison, as it supports more resilient and interpretable assessments in complex spatial environments.

5.2. Interpretation of Key Drivers

The SHAP interpretability analysis revealed critical insights into the primary drivers influencing ecological–geological quality, notably, NDVI, GDP, and slope. The substantial positive relationship between NDVI and ecological quality strongly indicates vegetation’s pivotal role in maintaining ecosystem stability, preventing soil erosion, and supporting hydrological processes, which are particularly vital within Laos’s steep, monsoon-driven topography. This observation corroborates ecological theories emphasizing vegetation as a foundational stabilizing factor. Conversely, GDP exhibited a clear inverse relationship with ecological quality, highlighting significant ecological degradation associated with economic intensification and urban expansion, particularly in metropolitan areas such as Vientiane. This finding underscores the widely recognized tension between development and ecological preservation. Furthermore, the negative correlation between slope steepness and ecological quality is consistent with geomorphological theories, linking steep terrains to increased susceptibility to erosion, geological instability, and reduced vegetation resilience. Together, these findings provide robust empirical validation of theoretical frameworks addressing ecological stability within mountainous landscapes.

5.3. Spatial Patterns and Regional Insights

The ecological–geological spatial zonation results delineated by the stacking model identified three distinct ecological–geological categories: urban–industrial degradation zones, mountainous hazard-prone regions, and ecologically stable plains. The observed spatial distribution clearly reflects interactions between topographic, anthropogenic, and ecological factors. Urban–industrial degradation zones near Vientiane reflect substantial ecological pressure from human activity. Mountainous regions, such as those surrounding Savannakhet Province, were identified as high risk due to their geological instability, steep slopes, and sparse vegetation cover, which are often associated with dense fault networks. Conversely, high-quality ecological zones, predominantly found in the northeastern and southeastern plains, were characterized by gentle terrain, dense vegetation cover, and minimal human disturbance, which are indicative of optimal ecological conditions. This spatial distribution pattern aligns well with documented ecological patterns in tectonically active mountainous regions (e.g., the Himalayas, Yunnan), confirming the importance of integrated geomorphological, ecological, and anthropogenic considerations in regional ecological management strategies.

5.4. Methodological Contributions, Limitations, and Prospects

In contrast to traditional ecological assessment approaches such as PCA and the AHP, the stacking ensemble model offers a significant methodological advancement by eliminating subjectivity in weight assignment and linear assumption constraints. Its data-driven, flexible approach effectively accommodates complex interactions among ecological indicators. The use of SHAP further enhances interpretability, improving the transparency of the modeling process and facilitating informed decision-making among policy-makers and stakeholders. This approach’s adaptability and scalability also provide substantial potential for application across other mountainous or data-limited regions, thereby broadening its applicability.

It is acknowledged that the labeled training dataset is relatively small. Nonetheless, methods such as Random Forest (RF), Gradient Boosting (GB), Support Vector Classifier (SVC), and K-Nearest Neighbors (KNN) are known to perform reliably under limited-sample conditions. To build the stacking ensemble, we applied 5-fold cross-validation to generate meta-level features, making efficient use of the available samples while minimizing overfitting. Several recent remote sensing studies confirm the effectiveness of this approach in data-constrained scenarios. For example, Tahmouresi et al. successfully applied a stacking framework to downscale soil moisture in the Urmia Basin, achieving an R² of 0.97 with only a moderate sample size, and used SHAP for interpretability [68]. Additionally, the potential influence of spatial autocorrelation among neighboring grid cells was not explicitly addressed, which may lead to overestimation of the model’s performance due to spatial leakage between the training and testing samples. Although we applied random cross-validation in this study, future research should adopt more spatially aware validation strategies, such as spatial blocking or leave-location-out cross-validation, to rigorously control for autocorrelation effects and assess true model generalization. Recent literature has emphasized the importance of such techniques in geospatial modeling tasks where spatial dependence is prevalent [69].

Our findings are consistent with a growing body of literature emphasizing the effectiveness of stacking ensemble models in environmental classification and susceptibility mapping. For instance, El Bouzekraoui et al. [70] applied a stacking ensemble to model gully erosion susceptibility in a mountainous semi-arid watershed, achieving higher prediction accuracy and spatial reliability in complex terrain. Similarly, Yao [71] demonstrated that stacking and blending techniques can enhance flash flood potential assessment in a Chinese river basin, providing greater robustness than single-model predictions. More recently, Guo [72] developed a hybrid model that integrates high-resolution remote sensing with stacking ensemble learning for detailed landslide susceptibility mapping, highlighting improved spatial precision in mountainous landscapes. Collectively, these studies support our approach by confirming that ensemble frameworks can effectively generalize across diverse environmental and geographical contexts.

However, several limitations warrant acknowledgment. Firstly, the ecological quality thresholds were subjectively determined based on the literature, which may influence generalizability and introduce biases. Secondly, this study’s reliance on single-period remote sensing imagery constrains insights into temporal ecological dynamics. Additionally, the potential influence of spatial autocorrelation among neighboring grid cells was not explicitly addressed, potentially affecting prediction accuracy. In addition, some remote sensing indicators used in this study—such as NDVI and NDWI—are affected by seasonal variability and atmospheric conditions. Although cloud-free imagery was selected as much as possible, residual cloud contamination and vegetation phenology may still introduce uncertainty into index values. These limitations suggest that incorporating multi-temporal imagery and applying strict cloud masking could further enhance the reliability of future assessments.

Future research should focus on integrating multi-temporal remote sensing datasets to capture ecological dynamics and trends comprehensively. Improvements in sampling strategies, such as combining expert judgment and automated methods, could significantly reduce classification biases. Investigating spatial autocorrelation explicitly through spatial statistical modeling will further strengthen model robustness and accuracy. Addressing these aspects will significantly enhance the methodological rigor and broaden the applicability of ecological–geological assessment frameworks.

6. Conclusions

Using Laos as a case study, this research proposes a comprehensive framework for evaluating ecological–geological quality in tropical mountainous regions. A total of ten evaluation indicators were constructed across four primary dimensions: geology, topography, ecology, and human activity. These indicators were derived through the integration of remote sensing interpretation, geological map processing, and socio-economic data mining, resulting in a structured, high-resolution input dataset for modeling.

In terms of methodology, we systematically compared the performance of eight classification models, including AdaBoost, KNN, Gradient Boosting, Random Forest, SVC, MLP, XGBoost, and stacking. A stacking ensemble model was then constructed using the seven base models as base learners and logistic regression as the meta-learner. This model significantly improved predictive accuracy and robustness. The experimental results demonstrated that the stacking model outperformed all others in terms of key metrics such as accuracy, F1 score, and AUC, exhibiting strong generalization and error control capabilities. By effectively integrating the advantages of diverse base models, the stacking model achieved high-precision classification and spatial prediction of ecological–geological quality.

To enhance model interpretability, the SHAP method was applied, which revealed that the interaction between ecological fundamentals, geomorphological features, and human activities plays a dominant role in shaping the spatial patterns of ecological–geological quality. Based on model predictions, a spatial zoning analysis was conducted, which delineated ecological high-quality zones and vulnerable areas across Laos. The results indicate that the spatial heterogeneity of the ecological–geological environment is primarily driven by the coupling of natural succession processes, geological structural background, and anthropogenic disturbances.

Accordingly, the integrated framework developed in this study—comprising indicator system construction, ensemble modeling, factor interpretation, and spatial zoning—provides a feasible pathway for the quantitative assessment of ecological environments under complex geological conditions. This framework is not only applicable to tropical mountainous regions, disaster-prone zones, and ecologically sensitive areas but also offers theoretical support and methodological guidance for policy-making and spatial planning practices.

In addition to its application in Laos, the proposed framework can be readily adapted to other mountainous regions facing similar ecological pressures, particularly in areas with limited field data but rich remote sensing coverage. Compared with traditional evaluation methods, the ensemble learning model shows superior generalization, flexibility in variable integration, and high-resolution prediction capacity. The inclusion of SHAP interpretation further enhances transparency, making it suitable for practical decision-making and policy support in diverse geographical contexts.

Nevertheless, the modeling process still faces certain limitations, including sample label bias and the absence of temporal geological data. Future research should focus on the integration of multi-temporal remote sensing data, simulation of disaster evolution processes, and assessment of uncertainty propagation to further expand the model’s applicability and interpretive capacity.