MuRaF-LULC: A Systematic Multivariate Random Forest Framework for Annual Land-Use and Land-Cover Mapping and Long-Term Change Detection

Abstract

Land-use and land-cover (LULC) change is one of the most pervasive drivers of socioenvironmental transformation worldwide. Given its impacts on ecosystems and climate, the systematic analysis of LULC dynamics remains a central objective of land-change science. Despite major advances in Earth observation capabilities, robust, flexible, and scalable algorithms for long-term monitoring remain unevenly adopted, particularly in remote, forested tropical regions. This study introduces the Multivariate Random Forest Land-Use and Land-Cover (MuRaF-LULC) framework, a supervised and generalizable framework that produces annual, multi-class LULC maps from Landsat time series, with interannual change derived through year-to-year comparisons. A key methodological component of the framework is its predictor-selection strategy, in which variable-importance rankings are used to identify an optimized subset of predictors prior to final model training. MuRaF-LULC was implemented in Google Earth Engine (GEE) and evaluated in Guatemala’s Maya Biosphere Reserve (MBR) for the 2018–2024 period using probability-based sampling and uncertainty-aware accuracy assessment and area estimation. Results show that MuRaF-LULC generates robust annual LULC classifications across multiple years (overall accuracy = 0.90–0.92) and reliable estimates of agropecuario expansion (the dominant transition in the study area) when change is assessed over longer temporal windows where transitions signals stabilize and for which the framework is best suited (producer’s accuracy = 0.97 ± 0.03; user’s accuracy = 0.69 ± 0.05). By prioritizing consistent annual, multiclass LULC trajectories, MuRaF-LULC complements breakpoint- and disturbance-oriented approaches commonly used in land-change studies. Implemented in publicly available, well-documented GEE scripts, MuRaF-LULC facilitates policy-relevant LULC assessment by remote sensing practitioners in governmental and private organizations, where reproducibility, clarity, and ease of deployment are as important as methodological sophistication.

1. Introduction

Land-use and land-cover (LULC) changes have produced the largest relative negative impact on terrestrial and freshwater ecosystems since 1970 [1]. The primary driver of these changes is the expansion of croplands and pastures for livestock production, which together occupy more than one-third of the Earth’s terrestrial surface [2] and are responsible for 86% of global deforestation [3], with proportions reaching up to 99% in tropical regions [4]. These LULC pressures have already reduced local terrestrial biodiversity worldwide, resulting in within-site species richness declines of 13.6% globally and up to 75% in the worst-affected habitats [5]. LULC changes are also a major contributor to climate change. About 25% of global emissions originate from land clearing, crop production, and fertilization, with animal-based food systems contributing 75% of that total [1]. Given the magnitude of these changes, detecting and monitoring their trajectories across spatiotemporal scales, along with their human–environment implications, is a long-standing central objective of global environmental research and a foundational motivation for the emergence of land change science [6,7,8,9]. The increasing availability of consistent, repeated Earth observations from satellite missions, particularly Landsat, with a more than 50-year, freely accessible record [10], has expanded our ability to detect both abrupt and gradual landscape transformations [11,12]. Despite these advances, robust, flexible, and scalable algorithms designed for long-term monitoring remain unevenly adopted in monitoring workflows, particularly in remote, forested tropical regions.

The best-known Landsat change detection algorithms have historically given priority to vegetation disturbance detection over the generation of multi-class, temporally consistent LULC maps. These canonical algorithms rely either on temporal segmentation approaches—e.g., Landsat-based detection of Trends in Disturbance and Recovery (LandTrendr) [13], Breaks For Additive Seasonal and Trend (BFAST) [14], Continuous Monitoring of Forest Disturbance Algorithm (CMFDA) [15], Continuous Change Detection and Classification—Spectral Mixture Analysis (CCDC-SMA) [16], Continuous Degradation Detection (CODED) [17]—or, less commonly, on multivariate machine-learning frameworks, such as the Multi-variate Time-series Disturbance Detection (MTDD) [18]. In contrast, LULC mapping has typically relied on general-purpose classifiers, such as decision trees, Random Forests, or Support Vector Machines, that are not designed a priori to ensure consistency across years. A notable exception is Continuous Change Detection and Classification (CCDC) [19], a temporal segmentation-based algorithm that explicitly integrates LULC classifications.

Temporal segmentation-based algorithms differ primarily in how they model pixel-level time series of specific spectral bands or indices and in how they define change. LandTrendr detects and characterizes vegetation loss by modeling spectral trajectories as a sequence of linear segments, where breakpoints along the fitted trajectories are interpreted as disturbance and subsequent recovery [13]. BFAST decomposes spectral time series into trend, seasonal, and residual components to detect structural breaks in either trend or seasonality; these breaks are often interpreted as LULC change events, but multi-class LULC maps are not produced [14]. CMFDA, CCDC, and CCDC-SMA share a common modeling framework in which harmonic or seasonal–trend models are fit to pixel-level reflectance time series, and change is flagged when observed values deviate persistently from model predictions. Unlike CMFDA, designed to identify a single forest disturbance event through the time series, CCDC supports multiple change segments by refitting models after each detected change and by assigning LULC labels to the resulting temporal segments. CCDC-SMA extends this latter approach by incorporating spectral mixture analysis, which tracks subpixel fraction trajectories to improve sensitivity and produce forest degradation maps [15,16,19]. CODED also leverages spectral mixture by detecting statistically significant changes in Normalized Difference Fraction Index (NDFI) trajectories to map degradation and deforestation [17]. MTDD departs from segmentation-based approaches by training a machine-learning classifier with multivariate metrics from spectral trajectories [18], enabling degradation and deforestation detection without explicit temporal segmentation but remaining focused on identifying disturbance events rather than multi-class LULC transitions.

To bridge the gap between disturbance-centric change detection and temporally consistent LULC mapping, this study introduces the Multivariate Random Forest Land-Use and Land-Cover (MuRaF-LULC) algorithm, a supervised, generalizable machine-learning framework designed for detecting and mapping LULC change from Landsat time series. Compared to segmentation-based approaches such as CCDC, where LULC labels are typically assigned at the segment level and may not change until a new breakpoint is detected, MuRaF-LULC enables flexibility to produce annual LULC maps and to analyze year-to-year LULC dynamics. MuRaF-LULC is built around multivariate image stacks that integrate spectral bands, spectral indices, and ancillary variables to model LULC transitions. A key component of the framework is its systematic candidate variable selection, in which variable importance rankings derived from Random Forest training [20] are used to identify an optimal and stable subset of predictors for classification. Using this predictor set, MuRaF-LULC applies models trained in a reference year consistently to target years, enabling temporally coherent LULC mapping across multiple spatial and temporal scales. The algorithm is implemented in the Google Earth Engine (GEE) cloud-computing environment, and the full source code is made publicly available to support reproducibility and transferability.

Tropical forested regions present some of the most challenging conditions for long-term LULC monitoring due to persistent environmental constraints such as consistent cloud cover and complex human land-change dynamics linked to governance, land tenure, and economic pressures, which together produce spatially fragmented and temporally abrupt transitions [21]. The Maya Biosphere Reserve (MBR) in northern Guatemala exemplifies these challenges [22] and, therefore, provides a rigorous testbed for evaluating the performance of LULC change detection algorithms. The objectives of this study are to (i) develop and implement the MuRaF-LULC algorithm as a generalizable framework for multi-class, annual LULC mapping, using the MBR as a complex tropical study area, and (ii) evaluate MuRaF-LULC performance by validating both single-date LULC classifications and multi-temporal LULC change maps. The following sections describe the study area, data sources, and methodological workflow used to implement MuRaF-LULC.

2. Study Area

This study uses Guatemala’s Maya Biosphere Reserve (MBR) (Figure 1) as a demonstration area for evaluating MuRaF-LULC. The MBR is a 21,602 km² protected area that has experienced extensive deforestation associated with complex human dynamics [23,24]. These dynamics generate diverse, nonlinear LULC change trajectories that are observed across tropical forest frontiers worldwide [25,26,27].

The MBR forms part of one of the largest tropical forest complexes remaining in the Americas and includes savannas and wetland systems [28,29]. Previous studies have shown that forest-to-pasture conversion, occurring under diverse and uneven governance conditions, is the dominant LULC transition in the region, with minimal conversion of forest to agriculture [22,30]. In Latin America, agropecuario land use encompasses both farming and cattle ranching; accordingly, the LULC types present within the MBR include agropecuario, forest, savanna, wetland, and water. Built-up areas occupy a very limited spatial extent and contribute negligibly to overall LULC change. This study examines LULC dynamics in the MBR over the period 2018–2024, which captures recent patterns of forest conversion under contemporary governance and land-use conditions.

The reserve is administratively divided into the multiple-use zone (40% of the reserve), the core protected zone (36%), and the buffer zone (24%). The core protected zone comprises National Parks and Biotopes, which are the most ecologically and culturally significant areas of the reserve. Sierra del Lacandón and Laguna del Tigre National Parks, located in the western portion of the reserve, have experienced higher rates of deforestation than other areas of the MBR [22]. From a remote sensing perspective, the MBR is characterized by a strong seasonal signal in vegetation condition associated with alternating wet and dry periods, which introduces intra-annual spectral variability in optical imagery used to generate annual LULC maps. Topography across the MBR is predominantly low relief, which limits topographic illumination effects. However, elevation and terrain-related gradients remain relevant for capturing hydrological conditions and landscape context associated with different land-cover types.

3. Materials and Methods

This section describes the implementation and validation of MuRaF-LULC, a multivariate Random Forest framework—that generates annual LULC maps and detects LULC transitions on a pixel-by-pixel basis using Landsat data. The workflow was implemented using the GEE JavaScript API via the Code Editor interface (https://code.earthengine.google.com/).

The MuRaF-LULC workflow consists of: (1) collection of spatially and thematically representative reference observations for a selected training year; (2) generation of annual cloud-free surface-reflectance composites; (3) derivation of candidate input variables (including spectral bands, spectral indices, topography, and geographic location) that serve as predictors; (4) training of an initial Random Forest model to quantify variable-importance; (5) Top-ranked variable selection based on variable-importance ranking to define an optimized predictor subset; (6) training of the final Random Forest model using the optimized predictor subset; (7) year-by-year pixel-wise classification to produce consistent annual LULC maps, and (8) pixel-wise comparison of consecutive years to derive LULC change maps. The code used to implement MuRaF-LULC is publicly available as a set of GEE scripts corresponding to the data-preparation, model-training and classification, and change-detection components of the workflow (Supplementary Material).

The model performance is evaluated through a validation strategy that addresses both single-date LULC classification and multi-year LULC change detection using a stratified sampling design. Accuracy is assessed using confusion matrices and accuracy metrics, as described in Section 3.3.

3.1. Candidate Input Variables

MuRaF-LULC employs a set of candidate input variables (Table 1) designated to capture spectral characteristics and environmental context associated with LULC types, with particular relevance to tropical forest frontiers.

The shortwave infrared (SWIR) region is particularly sensitive to vegetation and soil moisture conditions and exhibits strong contrast between vegetated and non-vegetated surfaces, which facilitates discrimination among vegetation types and other LULC classes. To further characterize vegetation conditions, several widely used spectral indices are integrated. The Normalized Difference Vegetation Index (NDVI) [31] detects vegetation presence, density, and vigor across a range of canopy conditions. The Normalized Difference Fraction Index (NDFI) [32], a spectral unmixing–based metric sensitive to canopy cover, has demonstrated strong performance in detecting forest loss in tropical forests (e.g., [17,33]). Vegetation moisture dynamics are characterized using the Normalized Difference Moisture Index (NDMI) [34], which is effective in distinguishing vegetation types with naturally different water content. The Soil-Adjusted Vegetation Index (SAVI) [35] is incorporated to improve vegetation detection in areas with sparse or heterogeneous cover, where soil background effects can reduce the performance of indices such as NDVI. The Normalized Difference Water Index (NDWI) [36] supports the delineation of open water bodies. In addition to spectral bands and indices, MuRaF-LULC incorporates ancillary environmental variables to provide contextual information relevant to LULC patterns. Elevation is included to account for altitudinal gradients that influence vegetation zonation and LULC suitability, while geographic location (latitude and longitude) supports the modeling of broad spatial gradients and regional patterns associated with climate, physiography, and human activity.

Together, these variables form a set of predictors that support the discrimination of LULC classes by combining spectral information with environmental context.

Table 1. MuRaF-LULC candidate input variables.

Acronym	Full Name	Formulation/Description	Data Source
SWIR1640	Shortwave Infrared1	Spectral region from 1.55 to 1.75 μm	Atmospherically corrected surface reflectance from Landsat Thematic Mapper (TM), Landsat Enhanced Thematic Mapper + (ETM+), and/or Landsat Operational Land Imager (OLI) (Courtesy of the U.S. Geological Survey)
SWIR2130	Shortwave Infrared2	Spectral region from 2.08 to 2.35 μm
NDVI	Normalized Difference Vegetation Index	(NIR − Red)/(NIR + Red)
NDFI	Normalized Difference Fraction Index	(GVshade − (NPV + Soil))/(GVshade + (NPV + Soil)), where ° GVshade = GV/1 − Shade ° GV = Green Vegetation ° NPV = Non-Photosynthetic Vegetation
NDMI1640	Normalized Difference Moisture Index1	(NIR − SWIR1640)/(NIR + SWIR1640)
NDMI2130	Normalized Difference Moisture Index2	(NIR − SWIR2130)/(NIR + SWIR2130)
SAVI	Soil-Adjusted Vegetation Index	1.5 ×(NIR − Red)/(NIR + Red + 0.5)
NDWI	Normalized Difference Water Index	(Green − NIR)/(Green + NIR)
—	Elevation	Elevation in meters above sea level	Copernicus Digital Elevation Model (DEM) [37]
—	Location	Geographic coordinates (latitude and longitude) in decimal degrees	EPSG:4326 Geographic Coordinate System (GCS) based on the WGS 84 datum [38]

Table 1. MuRaF-LULC candidate input variables.

Acronym	Full Name	Formulation/Description	Data Source
SWIR1640	Shortwave Infrared1	Spectral region from 1.55 to 1.75 μm	Atmospherically corrected surface reflectance from Landsat Thematic Mapper (TM), Landsat Enhanced Thematic Mapper + (ETM+), and/or Landsat Operational Land Imager (OLI) (Courtesy of the U.S. Geological Survey)
SWIR2130	Shortwave Infrared2	Spectral region from 2.08 to 2.35 μm
NDVI	Normalized Difference Vegetation Index	(NIR − Red)/(NIR + Red)
NDFI	Normalized Difference Fraction Index	(GVshade − (NPV + Soil))/(GVshade + (NPV + Soil)), where ° GVshade = GV/1 − Shade ° GV = Green Vegetation ° NPV = Non-Photosynthetic Vegetation
NDMI1640	Normalized Difference Moisture Index1	(NIR − SWIR1640)/(NIR + SWIR1640)
NDMI2130	Normalized Difference Moisture Index2	(NIR − SWIR2130)/(NIR + SWIR2130)
SAVI	Soil-Adjusted Vegetation Index	1.5 ×(NIR − Red)/(NIR + Red + 0.5)
NDWI	Normalized Difference Water Index	(Green − NIR)/(Green + NIR)
—	Elevation	Elevation in meters above sea level	Copernicus Digital Elevation Model (DEM) [37]
—	Location	Geographic coordinates (latitude and longitude) in decimal degrees	EPSG:4326 Geographic Coordinate System (GCS) based on the WGS 84 datum [38]

3.2. MuRaF-LULC Framework

The MuRaF-LULC framework is organized as a sequence of eight methodological steps that together enable the generation of annual LULC maps and the detection of interannual LULC transitions. The steps described below correspond directly to the operational implementation of the algorithm in GEE, with the exception of reference data collection, which is performed independently:

(1)

Reference observation collection. MuRaF-LULC begins with the compilation of spatially and thematically representative reference observations for a selected training year. Each observation is assigned a LULC class label and represents conditions that are stable and clearly identifiable at the spatial resolution of Landsat imagery. For the study area examined here, 10,000 reference observations were collected for the year 2020. The observations were obtained through visual interpretation of true-color composites derived from Planet-Norwegian International Climate and Forest Initiative (NICFI) (4.77 m spatial resolution) and PlanetScope imagery (3 m). The collection process was supported by the Collect Earth Online (CEO) platform [39] and Planet imagery access.

(2)

Annual Landsat compositing. For each year of analysis, MuRaF-LULC constructs an annual Landsat surface-reflectance composite using the LandTrendr library [13,40]. Cloud, cloud-shadow, snow, and water pixels are masked, and annual pixel values for blue, green, red, NIR, SWIR₁₆₄₀, and SWIR₂₁₃₀ are summarized using a medoid approach, which selects the observation closest to the median reflectance across all valid observations for the year. As a result, all valid Landsat observations within each calendar year contribute to the annual composites, which ensures consistent temporal representation across years without reliance on fixed acquisition dates. For the MBR application presented here, annual composites were generated for the 2018–2024 period, with selected two-year intervals used in subsequent classification and validation steps.

(3)

Candidate input variable generation. From each annual Landsat composite, MuRaF-LULC derives a multi-band image containing the candidate input variables listed in Table 1. These variables include two spectral bands (SWIR₁₆₄₀ and SWIR₂₁₃₀); six spectral indices related to vegetation condition, moisture, and canopy cover (NDVI, NDFI, NDMI₁₆₄₀, NDMI₂₁₃₀, and SAVI); a water sensitive index (NDWI); and ancillary environmental variables (elevation and geographic coordinates). The resulting multi-band image constitutes the candidate predictor stack used for model training and classification.

(4)

Initial Random Forest training and variable-importance analysis. An initial Random Forest classifier is trained using the reference observations described in step (1) and the full set of candidate input variables generated in step (3) for the reference year. The classifier consists of 500 decision trees, each built from a bootstrap sample of the training data. Following standard Random Forest parameter settings, the number of variables considered at each split was set to the square root of the total number of predictors, and trees were grown to full depth without explicit node-size or depth constraints. During the training, Random Forests compute impurity-based variable-importance metrics that quantify the contribution of each predictor to class discrimination across all trees in the Random Forest ensemble [41,42]. When a predictor is used to split a node within a decision tree, the split partitions the data into child nodes that are more homogeneous with respect to LULC class membership than the parent node. The improvement achieved by a split is quantified as a reduction in node impurity, which reflects enhanced class separation. For each predictor, these impurity reductions are accumulated across all splits and all trees in the ensemble, which yields an importance score that represents the predictor’s overall contribution to separating LULC classes. This initial model is used exclusively for variable-importance assessment and is not used for final LULC classification.

(5)

Top-ranked variable selection. Based on the variable-importance rankings obtained in step (4), MuRaF-LULC performs feature selection to define an optimized subset of predictors. This selection is implemented using one of two user-defined criteria: (i) a top-K approach, in which the K highest-ranked variables are retained, or (ii) a cumulative-importance approach, in which variables are retained until a user-defined proportion of the total importance is reached (e.g., 90–95%). Variable-importance scores are retrieved from the metadata of the trained Random Forest classifier in GEE, where impurity-based importance metrics are computed automatically during model training and stored as part of the classifier diagnostics. This selection strategy reduces predictor redundancy and limits the inclusion of weakly informative or correlated variables, while preserving those that contribute most strongly to class separability. Feature selection prior to final model training has been shown to improve classification stability, reduce overfitting, and enhance computational efficiency in Random Forest–based remote sensing applications [41,42]. These criteria are provided as configurable options within the MuRaF-LULC framework. In this study, the cumulative-importance approach was used with a threshold of 90%.

(6)

Final Random Forest training. A final Random Forest classifier is trained using the optimized predictor subset identified in Step (5) and the reference observations sampled from the selected training year described in Step (1). The classifier is trained with 500 decision trees. This final model constitutes the operational MuRaF-LULC classifier and is applied consistently to all target years.

(7)

Year-by-year LULC classification. The final Random Forest classifier is applied independently to each target year after restricting the multiband image generated in step (3) to the optimized predictor subset. LULC class labels are assigned to individual pixels using majority voting across the decision trees in the ensemble. This process produces temporally consistent annual LULC maps for the entire study period.

(8)

LULC change-map derivation. After annual LULC maps are generated, MuRaF-LULC derives change maps through pixel-wise comparisons of class labels between consecutive years. No additional temporal consistency rules, persistence constraints, or post-classification filtering are applied. Change detection is implemented with respect to a user-defined target class, allowing the identification of specific interannual transitions (e.g., loss or gain of a given LULC type). The resulting change maps explicitly capture the spatial and temporal patterns of LULC transitions across the study period. In this study, change maps focus on gain and loss of the agropecuario class, which is the dominant LULC transition in the study area.

3.3. Validation Strategy

Because MuRaF-LULC produces both annual LULC classifications and multi-year LULC change maps, validation was conducted separately for single-date classifications and for interannual change detection within Guatemala’s MBR. The assessment covers the period 2018–2024, with single-date LULC classifications validated for the years 2018, 2020, 2022, and 2024, and change detection evaluated at two-year intervals (2018–2020, 2020–2022, and 2022–2024), as well as for a cumulative 2018–2024 period. Two-year intervals were adopted as a strategy to improve validation efficiency while maintaining consistency across single-date classifications and change detection.

Both validation components are based on independently interpreted reference observations and follow stratified sampling designs adjusted to the spatial distribution of the mapped classes, in accordance with best practices for map accuracy assessment and area estimation [43,44]. Single-date classifications were validated using standard confusion matrices expressed in sample counts, whereas change maps, where unbiased area estimation is essential, were assessed through area-adjusted accuracy estimators.

3.3.1. Collection of Reference Observations

Reference observations used for the validation of both single-date and change-detection maps were obtained through visual interpretation of true- and false-color composites derived from PlanetScope imagery (3 m spatial resolution). The collection process was supported by the Planet Insights platform accessed through an academic research license, which facilitated systematic inspection of sample locations and temporal navigation of imagery. Planet imagery was used exclusively for reference interpretation and was not employed as input to the classification process.

Once the required number of samples was determined according to the validation designs described in Section 3.3.2 and Section 3.3.3, reference observations were collected independently for each validation target. Single-date reference labels were assigned for the years 2018, 2020, 2022, and 2024, corresponding to the annual LULC maps. Change-map reference labels were assigned for the interannual intervals 2018–2020, 2020–2022, and 2022–2024, as well as for a cumulative 2018–2024 period. For both single-date and change-detection interpretation, reference labels were determined by examining the full temporal context of each sample location within the target year(s). In cases where two or more LULC conditions were observed within a single year, the class corresponding to the longest persistence during that year was assigned as the reference label. This rule minimized ambiguity associated with transitional or short lived land-cover states.

3.3.2. Validation of Single-Date LULC Maps

The accuracy of the annual MuRaF-LULC maps (2018, 2020, 2022, and 2024) was assessed using independent sets of reference observations derived from stratified random samplings. Accuracy was quantified using conventional confusion matrices expressed in sample counts, from which overall accuracy, user’s accuracy, and producer’s accuracy were computed for each LULC class. This approach was selected because the primary objective of the single-date assessment was to evaluate the thematic reliability of each annual classification rather than to produce design-based area estimates, consistent with recommended best practices [43]. Each map was assigned approximately 280 samples, with a minimum of 20 samples for small classes (water, wetland, and savanna) and larger samples sizes for dominant classes (forest and agropecuario). Total sample sizes differ slightly across years because class proportions vary between classifications. Reference locations were reused across years when possible (with reference labels independently reassigned for each year) to reduce interpreter burden and facilitate comparability among maps, while maintaining a stratified coverage.

3.3.3. Validation of Multi-Year LULC Change Maps

Change detection accuracy was evaluated for four temporal intervals: 2018–2020, 2020–2022, 2022–2024, and a cumulative 2018–2024 period. Because the dominant LULC dynamic in the MBR is the conversion of land to agropecuario use, MuRaF-LULC generated change maps with the following mutually exclusive classes: (1) agropecuario loss, (2) agropecuario gain, (3) stable agropecuario, and (4) stable non-agropecuario. Unbiased estimation of the areas of these changes, particularly agropecuario gain, is essential for land-change analysis and cannot be derived from error matrices expressed only in sample counts, unless strata weights are incorporated. Accordingly, consistent with [43,45], this validation followed a stratified random sampling framework in which both accuracy metrics and proportions of area were estimated from the error matrix using design-based estimators that explicitly account for the sampling design and provide unbiased estimators with associated standard errors.

Sample sizes and stratified allocation were determined using the AREA2 sampling design tool, a GEE application that implements the stratified random sampling and estimation framework described by [43,45]. For every change map, sampling parameters were defined as follows: target class was agropecuario gain; the anticipated user’s accuracy for the target class was set to 0.80; the target standard error for the estimated area of the target class was set to 0.005 (expressed as a proportion of the total map area); and the anticipated proportion of true target-class pixels present in other strata was specified as 0.03 for agropecuario loss, 0.05 for stable agropecuario, and 0.01 for stable non-agropecuario.

Based on AREA2 outputs, which provide recommended total sample sizes and proportional allocation across strata, an additional design constraint was applied to ensure robust estimation of the target class. Specifically, agropecuario gain was assigned a minimum of 75 samples to ensure adequate precision of both area and accuracy estimates. The remaining samples were allocated proportionally to the mapped area of the other strata. To avoid undersampling rare classes, a minimum of 30 samples per stratum was enforced; when this constraint was applied, additional samples were reassigned from the largest stratum. Final sample sizes were adjusted within the range recommended by AREA2 to maintain comparable totals across change intervals, resulting in approximately 220 samples per map (

Table 2. Proportional allocation of sample size across change maps based on AREA2.

Change Interval	Agropecuario Loss	Agropecuario Gain	Stable Agropecuario	Stable Non-Agropecuario
2018–2020	30	75	37	78
2020–2022	30	75	33	83
2022–2024	30	75	34	81
2018–2024	30	75	32	83

).

Area and accuracy metrics were derived using design-based estimators that explicitly account for the stratified random sampling design. Following [43,45], error matrices were expressed as area proportions by weighting stratum-level sample outcomes by mapped stratum area, which yields unbiased estimates of class area and overall accuracy with associated standard errors. User’s and producer’s accuracies were computed as ratio estimators from the same weighted error matrix, with uncertainty quantified using standard error propagation. All computations were implemented in R 4.3.2.

4. Results

This section presents the outputs and validation results of the MuRaF-LULC framework for Guatemala’s MBR over the period 2018–2024. Results are organized into (i) representative single-date classifications, multi-year change maps, and associated model diagnostics; (ii) thematic accuracy of the annual LULC maps assessed using confusion matrices; and (iii) uncertainty-aware, area-adjusted estimates of agropecuario change and the corresponding accuracy metrics across multiple temporal intervals.

4.1. MuRaF-LULC Outputs

As described in Section 3.2, MuRaF-LULC generated outputs using an optimized subset of predictors selected to account for 90% of the cumulative variable-importance. Variable-importance diagnostics derived from the initial Random Forest model indicate that geographic location and topography were the most influential predictors, followed by spectral variables, including SWIR, NDWI, NDMI, and NDVI (

Table 3. Predictor variable importance derived from the initial MuRaF-LULC Random Forest model. Importance scores are unitless and reflect the relative contribution of each predictor to the classification performance.

Predictor Variable	Rank	Importance Score	Retained in the Final Model
Latitude	1	4539	Yes
Longitude	2	4301	Yes
DEM	3	4184	Yes
SWIR1640	4	3450	Yes
SWIR2130	5	3177	Yes
NDWI	6	2937	Yes
NDMI2130	7	2812	Yes
NDMI1640	8	2770	Yes
NDVI	9	2768	Yes
SAVI	10	2740	No
NDFI	11	2712	No

). Predictors with the lowest relative importance, specifically SAVI and NDFI, were excluded from the final model in accordance with the MuRaF-LULC predictor-selection procedure.

Using the optimized predictor subset, MuRaF-LULC produced annual LULC classifications for the period 2018–2024. Because spatial differences among annual maps are subtle at the reserve scale and difficult to visualize, only the most recent classification is shown to avoid redundancy. Figure 2 shows the 2024 classification alongside a contemporaneous NIR false-color composite to provide visual context for the spatial distribution of major LULC classes. Forest dominates the central and northeastern portion of the reserve, which largely corresponds to the multiple-use zone, while agropecuario land use is concentrated along the southern and western margins, primarily within the buffer zone and Laguna del Tigre National Park, where fragmentation patterns and clearings are evident. Savanna occurs in localized northwestern patches, mainly in transitional areas. Wetland and water classes are spatially coherent and largely confined to low-elevation areas and river corridors.

From the annual LULC classifications, MuRaF-LULC generated change maps that reveal the spatial patterns of agropecuario dynamics across the reserve (Figure 3). In general, agropecuario gain is concentrated along active deforestation frontiers and expansion zones, while agropecuario loss occurs more sparsely and in smaller, discontinuous patches. Stable agropecuario areas form contiguous regions in long-established agricultural areas within the buffer zone and Laguna del Tigre National Park, whereas stable non-agropecuario areas dominate the interior forest region. Changes exhibit temporal variability in their spatial expression. During the earliest interval, changes are more localized and aligned with established deforestation frontiers. In later intervals, spatial patterns become increasingly fragmented. In particular, the 2020–2022 interval is characterized by more widespread agropecuario loss. The independently estimated 2018–2024 change map integrates these dynamics and emphasizes areas where agropecuario gain and loss persisted across multiple years, which helps to distinguish sustained transitions from short-lived or transient changes.

Quantitative estimates of change area and associated uncertainty were not derived directly from mapped pixel counts, as map-based area summaries are subject to bias due to classification errors. Instead, unbiased area estimations for change classes, particularly for agropecuario gain, were obtained through the stratified validation framework described in Section 3.3.3 and reported in Section 4.3.

4.2. Single-Date LULC Classification Accuracy

The accuracy of the annual MuRaF-LULC classifications for 2018, 2020, 2022, and 2024 was evaluated using independent stratified reference observations (

Table 4. Confusion matrices expressed as sample counts and associated accuracy metrics for the MuRaF-LULC single-date classifications in Guatemala’s MBR.

			Reference
			Agropecuario	Forest	Savanna	Wetland	Water	Total	Producer’s Accuracy
MuRaF-LULC	2018	Agropecuario	56	5	0	1	0	62	0.90
		Forest	6	138	0	0	0	144	0.96
		Savanna	4	0	21	0	0	25	0.84
		Wetland	3	2	0	20	0	25	0.80
		Water	0	0	0	0	20	20	1.00
		Total	69	145	21	21	20	276
		User’s Accuracy	0.81	0.95	1.00	0.95	1.00
	2020	Agropecuario	75	13	1	3	0	92	0.82
		Forest	3	113	1	0	0	117	0.97
		Savanna	0	0	25	0	0	25	1.00
		Wetland	2	0	0	22	0	24	0.92
		Water	0	0	0	1	20	21	0.95
		Total	80	126	27	26	20	279
		User’s Accuracy	0.94	0.90	0.93	0.85	1.00
	2022	Agropecuario	69	3	0	1	0	73	0.95
		Forest	9	126	1	1	0	137	0.92
		Savanna	7	0	26	0	0	33	0.79
		Wetland	3	0	0	22	0	25	0.88
		Water	0	0	0	0	20	20	1.00
		Total	88	129	27	24	20	288
		User’s Accuracy	0.78	0.98	0.96	0.92	1.00
	2024	Agropecuario	90	14	1	2	0	107	0.84
		Forest	4	97	0	0	0	101	0.96
		Savanna	3	0	22	0	0	25	0.88
		Wetland	2	1	0	20	0	23	0.87
		Water	0	0	0	0	20	20	1.00
		Total	99	112	23	22	20	276
		User’s Accuracy	0.91	0.87	0.96	0.91	1.00

). Overall accuracy was consistently high across all years: 0.92 in 2018, 0.91 in 2020, 0.91 in 2022, and 0.90 in 2024. Forest and water classes exhibited the highest and most stable accuracies. User’s accuracy for forest ranged between 0.87 and 0.98, with corresponding producer’s accuracy values between 0.92 and 0.97. Water was classified with perfect performance in all years, except in 2020, when producer’s accuracy decreased slightly to 0.95. Wetland and savanna classes showed a moderately strong performance. Wetland user’s accuracy oscillated between 0.85 and 0.95, and producer’s accuracy between 0.80 and 0.92. Savanna user’s accuracy was consistently high, between 0.93 and 1.00, while its producer’s accuracy showed greater interannual variability, ranging from 0.79 to 1.00.

Agropecuario, the target class of this study and the basis for change detection and area estimation, exhibited user’s accuracy ranging between 0.78 and 0.94, and producer’s accuracy from 0.82 to 0.95. Lower user’s accuracy in some years indicates the presence of commission errors, which means that a proportion of pixels classified as agropecuario corresponds to other LULC types in the reference data. These errors are primarily concentrated along fragmented forest and savanna interfaces, where mixed LULC signals and fine-scale spatial heterogeneity increase classification ambiguity. Overall, the single-date accuracy assessment indicates that MuRaF-LULC produced annual LULC maps that are reliable for subsequent change-detection analyses in the MBR.

4.3. Multi-Year LULC Change Map Accuracy and Area Estimation

The accuracy and area of agropecuario change were evaluated for two-year intervals and for the overall period using a stratified random sampling framework designed to provide unbiased, uncertainty-aware estimates for each change class, as described in Section 3.3.3.

Table 5. Estimated class area proportions and associated accuracy metrics with standard errors (SE) for the MuRaF-LULC change maps in Guatemala’s MBR.

	Class	Estimated Area Proportion	SE Estimated Area Proportion	User’s Accuracy	SE User’s Acc.	Producer’s Accuracy	SE Producer’s Accuracy
2018–2020	Agropecuario gain	0.03	0.01	0.48	0.06	0.73	0.2
	Agropecuario loss	0.01	0	0.47	0.09	0.95	0.05
	Stable agropecuario	0.26	0.01	0.95	0.04	0.89	0.03
	Stable non-agropecuario	0.7	0.02	0.97	0.02	0.96	0.01
2020–2022	Agropecuario gain	0.04	0.01	0.75	0.05	0.31	0.11
	Agropecuario loss	0.05	0.01	0.57	0.09	0.82	0.15
	Stable agropecuario	0.22	0.02	0.88	0.06	0.88	0.04
	Stable non-agropecuario	0.69	0.02	0.95	0.02	0.96	0.01
2022–2024	Agropecuario gain	0.04	0.01	0.61	0.06	0.85	0.13
	Agropecuario loss	0.01	0.01	0.17	0.07	0.31	0.23
	Stable agropecuario	0.23	0.01	0.91	0.05	0.87	0.03
	Stable non-agropecuario	0.72	0.01	0.99	0.01	0.97	0.01
2018–2024	Agropecuario gain	0.04	0	0.69	0.05	0.97	0.03
	Agropecuario loss	0.04	0.01	0.6	0.09	0.64	0.16
	Stable agropecuario	0.22	0.01	0.94	0.04	0.94	0.02
	Stable non-agropecuario	0.7	0	1	0	0.97	0.01

shows that stable non-agropecuario consistently accounted for the largest proportion of the mapped area across all intervals, with estimated areas ranging from 69% to 72% of the reserve and low standard error (SE ≤ 0.02) associated. This class showed very high user’s and producer’s accuracies (0.95–1.00) and similarly low uncertainty (SE ≤ 0.02). Stable agropecuario represented the second largest area fraction, accounting for about one fourth of the reserve across all intervals (22% to 26%, SE ≤ 0.02). User’s and producer’s accuracies for this class remained between 0.87 and 0.95, with slightly higher associated uncertainty (SE ≤ 0.06). Overall, the stable classes were robustly estimated.

As expected, change classes represented relatively small but highly dynamic fractions of the landscape (Table 5). Across the evaluated two-year intervals, agropecuario gain, extended over 3.2% to 4.4% of the study area, with the lowest extent observed during the earliest period (2018–2020) (Figure 4). Agropecuario gain exceeded agropecuario loss in most intervals, which reflects a net expansion of agropecuario land use, except during 2020–2022, when both processes showed comparable magnitudes. Notably, this interval was also characterized by higher uncertainty for both agropecuario gain and loss (SE = 0.015 and 0.010, respectively). Over the overall 2018–2024 period, agropecuario gain (4.1%) slightly exceeded agropecuario loss (3.9%); however, uncertainty associated with agropecuario loss was substantially higher (SE = 0.010) than that of agropecuario gain (SE = 0.003), which suggest that estimates of agropecuario expansion, the target class in this study, are more robust when assessed over longer periods.

The accuracies of the target class further highlight the influence of longer temporal periods on change detection performance (Figure 5). User’s accuracy for agropecuario gain remained moderate to high across all evaluated two-year intervals (0.48–0.75) with consistently similar associated uncertainty (SE = 0.05–0.06), which indicates a relatively stable reliability of mapped gain events. In contrast, producer’s accuracy exhibited substantially greater variability across the same intervals (0.31–0.85), coupled with a wider range of uncertainty (SE = 0.13–0.20). This reflects the inherent difficulty of capturing all true short-term gain events in fragmented agricultural frontiers. By comparison, the overall 2018–2024 change map exhibited more balanced and less uncertain user’s (0.69; SE = 0.05) and producer’s (0.97; SE = 0.03) accuracies. This pattern suggests that longer temporal windows reduce the influence of short-lived, ambiguous, or reversible transitions.

Together, these results establish a statistically robust foundation for interpreting the dominant LULC transition in Guatemala’s MBR and support an uncertainty-aware, multi-temporal assessment of this dynamic.

5. Discussion

Within its intended design scope, MuRaF-LULC demonstrates reliable performance for annual LULC mapping and change analysis in a complex tropical environment. Consistently high classification accuracy across all evaluated single-date maps indicates that the multivariate predictor design effectively captures the dominant spectral and environmental gradients within Guatemala’s MBR. At the same time, several methodological and interpretative nuances merit discussion to contextualize these results and clarify the algorithm’s strengths and limitations.

5.1. MuRaF-LULC Performance

Variable-importance diagnostics indicate that geographic location and elevation were among the most influential predictors. This reflects broad physiographic, climatic, and land-use practices across the administrative zones of the reserve. Although spatial predictors have been shown to increase the risk of overfitting in some machine learning applications (e.g., Ref. [46]), the consistency of MuRaF-LULC performance across multiple years, as demonstrated through independent validation samples, supports its use for regional-scale LULC mapping within the trained area. Because latitude and longitude encode location-specific context rather than transferable physical processes, their contribution may be less robust under direct spatial extrapolation. Consequently, applications of MuRaF-LULC beyond the training domain should rely on region-specific retraining rather than assuming direct geographic portability.

Single-date classifications achieved high overall accuracy (0.90–0.92) and strong class-specific performance across years, including for the agropecuario class (≥0.78), which formed the basis for change detection and area estimation in this study. In contrast, validation of interannual change maps yielded lower class-specific accuracies for agropecuario gain and loss, particularly over two-year intervals. This behavior is consistent with previous studies showing that change detection is inherently more challenging than single-date classification due to the propagation of classification errors across time and the typically small spatial extent of change classes [43,45].

Lower producer’s accuracy for agropecuario gain during specific intervals (most notably 2020–2022; PA = 0.31) indicates increased omission of true change events, whereas lower user’s accuracy (most notably 2018–2020; UA = 0.48) reflects a higher rate of commission errors, in which mapped gains do not correspond to true transitions. This behavior likely results from a combination of factors, including partially reversible or short-lived transitions, mixed pixels along fragmented deforestation frontiers at Landsat spatial resolution, and a temporal mismatch between the spectral observation selected by the annual medoid composite and the exact timing of on-the-ground conversion.

Cumulative change assessment over the full 2018–2024 period yielded a substantially improved performance. Agropecuario gain exhibited very high producer’s accuracy (0.97) and low associated uncertainty (SE = 0.03), and moderate user’s accuracy (0.69) with similarly low uncertainty (SE = 0.05). These results suggest that longer temporal windows effectively integrate persistent LULC signals and reduce the influence of short-lived, ambiguous, or reversible changes. This performance aligns with the intended use of MuRaF-LULC for long-term monitoring and policy-relevant LULC change analysis, where sustained transitions are of greater importance than ephemeral events.

5.2. Comparison with Existing Algorithms

Many well-established Landsat-based approaches to detect changes have prioritized the identification of disturbance timing and magnitude from spectral time series. Numerous extensions and refinements of these approaches have been proposed over the past 15 years; however, they largely retain the same core design principles related to temporal segmentation and disturbance-focused change detection. In contrast, MuRaF-LULC is designed to generate annual, multi-class LULC maps, with interannual change derived as a secondary product from year-to-year label comparisons (see

Table 6. Comparison between the MuRaF-LULC framework and selected Landsat-based change detection mapping algorithms.

Algorithm	Main Objective	Core Output (Representative, Not Exhaustive)	Comparison Relative to MuRaF-LULC
VCT [47]	Detection of forest disturbances using spectral trajectories	Disturbance timing and magnitude	Forest-focused; not designed to produce multi-class annual LULC maps
LandTrendr [13]	Detection of vegetation disturbance and recovery using segmented spectral trajectories	Disturbance year, magnitude, duration, and rate of change	Vegetation-focused; not designed to produce multi-class annual LULC maps
BFAST [14]	Detection of change events in trend and seasonal components of spectral time series	Time and magnitude of change events	Generic change detection; no explicit LULC labeling
CMFDA [15]	Detection of forest disturbances based on deviations from modeled trends	Disturbance timing and signal magnitude	Forest-focused; not designed to produce multi-class annual LULC maps
CCDC [19]	Detection and classification of change events based on deviations from modeled trends	Change timing and segment-based LULC labels	LULC class labels remain fixed within segments; not independent annual LULC classifications
CCDC-SMA [16]	Detection of forest degradation based on deviations from modeled trends using spectral mixture trajectories	Timing of degradation events and subpixel fraction trajectories	Forest degradation-focused; not designed to produce multi-class annual LULC maps
CODED [17]	Detection of forest disturbances based on changes in NDFI trajectories	Change timing, magnitude, and forest condition classes	Forest degradation and deforestation mapping; produces stratified forest condition fixed to change timing; not designed to produce multi-class annual LULC maps
MTDD [18]	Detection of annual forest disturbance events using multivariate trajectory metrics	Annual forest disturbance maps	Focused on forest disturbances rather than full multi-class LULC mapping
MuRaF-LULC	Detection of annual LULC classes and interannual transitions	Annual LULC maps and derived changes	Reference framework

for a conceptual comparison).

Early segmentation-based frameworks that focus on identifying structural changes in spectral or index trajectories include Vegetation Change Tracker (VCT), LandTrendr, BFAST, and CMFDA. While highly effective for characterizing disturbance timing and magnitude, these approaches are not designed to produce temporally consistent, multi-class annual LULC maps that are directly comparable from year to year. VCT was developed to identify stand-replacing forest disturbances by tracking sustained declines in vegetation indices through time [47]. LandTrendr fits piecewise linear segments to annual stacks and interprets breakpoints and subsequent trends as disturbance and recovery dynamics [13]. BFAST decomposes time series into trend and seasonal components and detects breaks in those components, which enables generic change detection [14]. CMFDA supports continuous disturbance monitoring by fitting harmonic-type models to dense observations and flagging disturbance when incoming observations deviate persistently from model predictions [15].

CCDC represents the closest conceptual bridge to MuRaF-LULC among the segmentation family because it combines continuous change detection with classification of temporal segments [19]. The key distinction lies in how class labels are assigned through time: in CCDC, LULC labels are associated with model segments and remain fixed until a new change is detected, whereas MuRaF-LULC produces an independent class label for each year. CCDC-SMA builds on this framework by incorporating spectral mixture analysis to track subpixel fraction trajectories and improve sensitivity to forest degradation processes [16]. CODED similarly employs spectral unmixing analysis through NDFI metrics to detect degradation dynamics [17]. These designs are powerful for monitoring forest degradation and deforestation, but their target product is typically deforestation/degradation, rather than a general multi-class annual LULC product intended to support broader LULC transition analysis.

MTDD differs from segmentation-based methods by using multivariate metrics derived from spectral trajectories within a machine-learning framework to detect annual disturbance events [18]. MuRaF-LULC similarly leverages multivariate predictors and machine-learning methods, but it is trained to discriminate multi-class LULC annual states, from which change is computed.

These design distinctions have important implications for how MuRaF-LULC should be interpreted. MuRaF-LULC performance is best evaluated in terms of the thematic reliability of annual LULC products and the robustness of change signals derived from their comparisons across time. Within this framework, strong single-date classification accuracy is a prerequisite for meaningful change analysis, where reduced class-specific accuracy for short-interval transitions in fragmented frontiers is an expected outcome. At the same time, the improved performance observed over longer temporal windows highlights the suitability of MuRaF-LULC for policy-relevant assessments of persistent LULC transitions. Beyond performance considerations, MuRaF-LULC is intentionally designed to be operationally transparent and straightforward to implement by relying on standard annual composites and supervised classification rather than complex temporal segmentation or breakpoint modeling. This design facilitates interpretation and application by remote sensing practitioners in governmental and private organizations, where reproducibility, clarity, and ease of deployment are often as critical as methodological sophistication.

5.3. Limitations and Future Directions

Despite its demonstrated reliability for annual LULC mapping and multi-year change analysis, MuRaF-LULC is subject to several limitations that are important to acknowledge and that also motivate future methodological development. First, because MuRaF-LULC derives change from differences between consecutive annual classifications, short-lived or partially reversible transitions may be underrepresented, particularly in fragmented agricultural frontiers and at Landsat spatial resolution. This limitation is inherent to this remote sensing workflow and reflects the compounded uncertainty of consecutive classifications, rather than deficiencies in single-date thematic accuracy. As shown in this study, this effect is most pronounced over short temporal intervals and diminishes when change is assessed over longer periods.

Second, MuRaF-LULC relies on annual surface-reflectance composites summarized using a medoid approach. While this strategy ensures spatial completeness and reduces the influence of clouds and outliers, it may not capture the precise timing of LULC transitions within the year. Consequently, abrupt changes that occur outside the temporal window represented by the selected observation may be delayed or smoothed in the annual record. Third, the framework employs supervised classification trained on reference data from a single year, assuming relative stability in class spectral separability through time. Although this assumption is shared with other supervised classification approaches and supported by the strong single-date performance observed across multiple years in this study, variability in short-term change detection likely reflects a combination of classification uncertainty and gradual shifts in LULC characteristics. In particular, interannual climate variability associated with El Niño/La Niña events may influence vegetation condition and spectral behavior in ways not fully captured by a single training year. While the robust performance observed over longer temporal windows suggests that these effects stabilize when change is aggregated through time, future extensions could explore adaptive or multi-year training strategies to further enhance temporal robustness.

Finally, while the MuRaF-LULC framework is generalizable, its application to new regions requires appropriate reference data and model retraining to capture region-specific landscape structure and LULC dynamics. This requirement is common to supervised LULC classification approaches and reflects standard practice rather than a limitation of the framework. Nevertheless, transferability across distinct environmental settings may introduce additional biome-specific challenges. In arid or semi-arid regions, for example, reduced vegetation seasonality and spectral similarity between bare soil and sparsely vegetated classes may require adjustments in candidate input variables. In contrast, in temperate systems, strong phenological cycles and seasonal land management practices may necessitate finer temporal resolution and the selection of alternative temporal compositing strategies. These considerations highlight that, while MuRaF-LULC is spatially transferable at the methodological level, optimal performance depends on context-aware calibration rather than direct model portability.

Looking forward, MuRaF-LULC is intentionally structured to support continuous improvement, extension, and adaptation. Future developments may involve the incorporation of multi-year training strategies, further candidate inputs, including ancillary socio-environmental and climate variables, and the integration of additional sensors (e.g., Sentinel-2), which could improve temporal sampling, reduce mixed-pixel effects in fragmented landscapes, and there-fore enhance sensitivity to short-lived or small-scale LULC transitions. In landscapes characterized by strong seasonal vegetation dynamics, such as the MBR, future extensions of MuRaF-LULC could also explore the incorporation of intra-annual variability metrics (e.g., annual NDVI standard deviation). Such metrics may help better discriminate classes with similar mean spectral responses but distinct seasonal behavior, such as savanna and agropecuario areas where spectral confusion was observed. Because the framework is modular and implemented in a cloud-computing environment, it can readily accommodate methodological advances and evolving user needs.

6. Conclusions

Given the global impacts of LULC change on ecosystems and climate, detecting and monitoring LULC trajectories remains a central objective of environmental research. This study introduced MuRaF-LULC, a generalizable multivariate Random Forest framework designed to address this challenge by producing annual, multi-class LULC maps from Landsat time series, with interannual change derived through year-to-year comparisons. The framework was evaluated in Guatemala’s Maya Biosphere Reserve, a complex tropical forest frontier, using single-date thematic accuracy assessment and uncertainty-aware accuracy and area estimation for multi-year LULC change.

Results demonstrate that MuRaF-LULC generates robust annual LULC classifications across multiple years and reliable cumulative estimates of agropecuario expansion when change is assessed over longer temporal windows. Design-based comparisons with existing change-detection algorithms show that MuRaF-LULC is complementary to segmentation- and disturbance-oriented approaches: rather than prioritizing the precise timing of disturbance events, it emphasizes interpretable annual LULC states and year-to-year trajectories, which are well suited for regional-scale LULC analysis and policy-relevant applications.

MuRaF-LULC is designed to be operationally transparent and straightforward to implement. It facilitates interpretation and deployment by remote sensing practitioners in governmental and private organizations, where reproducibility, clarity, and ease of implementation are often as critical as methodological sophistication. At the same time, its modular structure and cloud-based implementation support continued extension and adaptation, including the integration of additional sensors, alternative compositing strategies, and expanded predictors. Together, these characteristics present MuRaF-LULC as a flexible and operationally accessible foundation for long-term LULC monitoring across diverse socio-environmental contexts.