Object-Based Random Forest Approach for High-Resolution Mapping of Urban Green Space Dynamics in a University Campus

Abstract

Urban green space is essential for ecological functions, environmental quality, and human well-being, yet campus expansion can reduce vegetated areas. This study assessed UGS dynamics at Universitas Padjadjaran’s Jatinangor campus from 2015 to 2025 and evaluated an object-based machine learning approach for fine-scale land cover mapping. High-resolution WorldView-2, WorldView-3, and Legion-03 imagery were pan-sharpened, geometrically corrected, normalized, and used to compute NDVI and NDWI indices. Object-based image analysis segmented the imagery into homogeneous objects, followed by random forest classification into six land cover classes; UGS was derived from dense and sparse vegetation. Accuracy assessment included confusion matrices, overall accuracy 0.810–0.860, kappa coefficients 0.747–0.826, weighted F1 scores 0.807–0.860, and validation with 43 field points. The total UGS increased from 68.89% to 74.69%, bare land decreased from 13.49% to 5.81%, and building areas moderately increased from 10.36% to 11.52%. The maps captured vegetated and developed zones accurately, demonstrating the reliability of the classification approach. These findings indicate that campus expansion has been managed without compromising ecological integrity, providing spatially explicit, reliable data to inform sustainable campus planning and support green campus initiatives.

1. Introduction

Urban green space (UGS) plays a vital role in enhancing environmental quality, providing ecosystem services, and supporting human well-being in rapidly urbanizing areas. Previous research highlights that UGS contributes not only to ecological functions such as carbon sequestration, microclimate regulation, and biodiversity conservation but also to social and health benefits through recreational opportunities and improved mental health [1]. These multiple functions make UGS monitoring a central issue in urban sustainability studies.

In Indonesia, the importance of UGS is reinforced by legal instruments. Law No. 26 of 2007 on Spatial Planning stipulates that urban areas must allocate at least 30% of their total area as green space, with a minimum of 20% dedicated to public use [2]. At the regional level, the government of Sumedang Regency, through Regional Regulation No. 2 of 2012 on the Regional Spatial Plan for 2011–2031, emphasizes the need to protect ecological functions and secure sufficient UGS to maintain environmental balance [3]. These frameworks provide the normative foundation for integrating UGS into spatial planning at both national and local scales.

Universitas Padjadjaran (Unpad), located in Jatinangor, West Java, has undergone significant expansion in infrastructure and student population over the past decade. While this growth supports academic activities, it also puts pressure on the existing UGS and may reduce vegetation cover and ecological functions such as water retention, air quality improvement, and thermal regulation. Research on UGS dynamics at the campus level is therefore essential, as universities act as microcosms of urban systems where spatial planning decisions strongly influence sustainability outcomes [4].

Studies on urban environments have widely employed remote sensing to examine land use and land cover (LULC) changes, often relying on medium-resolution imagery such as Landsat or Sentinel. However, these sensors are limited in capturing the fine spatial heterogeneity of smaller environments. High-resolution imagery such as WorldView enables detailed mapping at the local scale [5]. Recent advances in urban remote sensing have also demonstrated the integration of deep learning and multi-source imagery, such as the combination of remote sensing and street view data for urban mapping in Bandung City [6]. Despite extensive research on urban green space at city or regional scales, several studies have demonstrated strong classification performance. For example, Ismayilova & Timpf (2022) [7] used Sentinel-2 and a pixel-based Random Forest for Augsburg, Germany, achieving 97% accuracy. Another study by Guo et al. (2022) [8] applied OBIA with Random Forest, SVM, and KNN for the Mawei District, with Random Forest reaching 91.15% accuracy. However, studies at the campus scale remain relatively scarce. Previous work at Unpad has primarily relied on vegetation indices [9], which tend to oversimplify land cover classification and may misinterpret sparse or mixed vegetation. This highlights a notable research gap in high-resolution, long-term UGS monitoring for micro-scale systems such as university campuses.

To overcome these challenges, object-based image analysis (OBIA) has been increasingly adopted in high-resolution remote sensing. OBIA integrates spectral, spatial, and textural information through segmentation and classification, producing more reliable land cover results than traditional pixel-based approaches [10]. Within this framework, machine learning methods such as random forest are particularly effective because they achieve high accuracy, are resistant to overfitting, and can manage the complexity of high-dimensional data [11,12].

Therefore, this study implements an object-based random forest approach to analyze the dynamics of urban green space at Unpad from 2015 to 2025. The research aims to quantify spatiotemporal changes in UGS and evaluate the performance of an object-based machine learning framework at the campus scale. The findings are expected to contribute theoretically by advancing the application of OBIA and Random Forest for small-scale urban research and practically by providing spatially explicit evidence to support sustainable campus planning. In addition, this study provides a methodological contribution by systematically validating, for the first time at the campus scale, the OBIA–Random Forest approach for decadal monitoring using high-resolution imagery. The results also reveal a unique pattern at Unpad, where infrastructure expansion and ecological restoration progress in parallel. This pattern is rarely documented in previous campus-scale remote sensing studies.

2. Materials and Methods

2.1. Study Area

Unpad Jatinangor Campus is located in Jatinangor District, Sumedang Regency, West Java, Indonesia, at approximately ${107}^{°}{46}^{\prime }{28}^{″}\mathrm{E}$ and $6^{°}{55}^{\prime }{33}^{″}\mathrm{S}$. It is ∼725–810 m above sea level, covering an area of around 180 hectares and comprising 18 faculties [13]. The land cover is diverse, with the western part containing an arboretum, the northern part includes the Ciparanje area, and the southern and central zones are dominated by academic buildings and supporting infrastructure. In addition, two major water bodies are present on campus: the Ecoriparian, a managed riparian ecological zone designed for water retention, biodiversity conservation, and environmental education; and Embung Leuwi Padjadjaran, an artificial reservoir functioning as a water storage system and a hydrological buffer for the surrounding area. An overview of the study area is shown in Figure 1.

Beyond its campus boundary, Jatinangor has developed into one of the prominent higher education clusters in Indonesia, hosting four major universities: Unpad, Institut Teknologi Bandung (ITB), Institut Koperasi Indonesia (IKOPIN), and Institut Pemerintahan Dalam Negeri (IPDN). This concentration of institutions has established Jatinangor as an acknowledged educational hub in West Java [13].

2.2. Data

This study utilized high-resolution satellite imagery from WorldView-2, WorldView-3, and Legion-03, acquired between 2015 and 2025 to capture temporal dynamics of land cover at the campus scale. These datasets, provided by Maxar Technologies, were selected due to their sub-meter spatial resolution, which is essential for detecting fine-scale variations in urban green space within a relatively small study area [14,15,16,17]. The images were supplied as Ortho Ready (Level 2A) products that have undergone radiometric correction and are referenced to the WGS84 coordinate system. The specific acquisition dates and spatial resolutions are summarized in

Table 1. Overview of satellite data used in this study.

Satellite	Acquisition Date	Panchromatic Resolution (m)	Multispectral Resolution (m)
WorldView-2	15 June 2015	0.46	1.84
WorldView-2	20 July 2017	0.46	1.84
WorldView-3	10 July 2020	0.31	1.24
Legion-03	18 May 2025	0.30	1.20

, ensuring consistency in temporal coverage and comparability across sensors.

For vegetation analysis and LULC classification, the visible (Red, Green, Blue) and Near-Infrared (NIR) bands were primarily employed. These bands are widely recognized for their sensitivity to vegetation vigor and are therefore well suited to mapping urban green space. To further support the analysis, the complete spectral configurations of the WorldView and Legion sensors are presented in

Table 2. Spectral characteristics of the WorldView and Legion satellites.

Band	Spectral Range (μm)
Coastal	0.400–0.450
Blue	0.450–0.510
Green	0.510–0.580
Yellow	0.585–0.625
Red	0.630–0.690
Red Edge	0.705–0.745
NIR-1	0.770–0.895
NIR-2	0.860–1.040

, providing an overview of the spectral ranges available for feature extraction.

The overall methodological workflow is illustrated in Figure 2, integrating object-based image analysis (OBIA) and random forest (RF) classification for multi-temporal high-resolution satellite imagery. The preprocessing stage involves defining land cover classes, applying pan-sharpening to enhance spatial resolution and geometric correction to improve positional accuracy. Processing involves image segmentation and RF classification to generate object-based LULC maps. Post-processing then begins with an accuracy assessment, including overall accuracy, Kappa, F1 score, and weighted F1 score. Ground truth data are used to support this accuracy evaluation. Subsequently, manual correction is performed to refine thematic consistency, followed by the computation of areal extent and statistical performance of each land cover class.

2.3. Preprocessing

Raw satellite imagery requires preprocessing to enhance accuracy and reliability in subsequent analyses. A key step in this process is pan-sharpening, which combines high-resolution panchromatic bands with lower-resolution multispectral bands to improve image sharpness without losing spectral information [18]. Before fusion, the multispectral bands were resampled to match the panchromatic spatial resolution using cubic spline interpolation, ensuring pixel-level alignment and smooth spectral transitions. The Brovey transform was then applied to integrate the spatial detail from the panchromatic image with the spectral characteristics of the multispectral bands, resulting in a high-resolution dataset suitable for campus-scale analysis. This method was selected due to its ability to preserve relative spectral proportions while enhancing spatial sharpness, making it effective for urban-scale vegetation mapping. The overall process is illustrated in Figure 3.

Geometric correction is also crucial to ensure spatial consistency across images acquired at different times, enabling precise comparison of land cover changes [19]. This step corrects distortions caused by satellite acquisition angle and sensor geometry. Geometric correction was conducted in QGIS (Figure 4), ensuring that all imagery aligns accurately within the study area.

Spectral indices were computed to enhance feature extraction. Specifically, the Normalized Difference Vegetation Index (NDVI) and the Normalized Difference Water Index (NDWI) were calculated to distinguish vegetation and water bodies from other land cover types. NDVI is defined as:

$$NDVI={\displaystyle \frac{NIR-RED}{NIR+RED}},$$

(1)

where NIR and RED represent the reflectance values of the near-infrared and red bands, respectively. NDVI values range from −1 to +1, with higher values indicating dense and healthy vegetation, near-zero values corresponding to non-vegetated surfaces, and negative values representing water, snow, or clouds [20]. In contrast, NDWI emphasizes water features and is calculated as:

$$NDWI={\displaystyle \frac{GREEN-NIR}{GREEN+NIR}},$$

(2)

where GREEN and NIR are the reflectance values in the green and near-infrared bands, respectively. Positive NDWI values indicate water, while negative values represent vegetation or dry land [21]. Both indices were computed in Google Earth Engine (GEE), providing additional features for improved land cover classification. Finally, data normalization was applied to ensure uniform scaling across all bands and spectral indices. Each variable was rescaled to a 0–1 range using min-max normalization, preventing variables with larger numeric ranges from dominating machine learning algorithms and ensuring equitable weighting in classification [22]. This step was performed in GEE to leverage cloud-based computation for consistent and efficient processing across all imagery.

2.4. OBIA Segmentation and Classification

Object-based image analysis (OBIA) is an advanced image processing approach that uses segmented objects as the unit of analysis rather than individual pixels. In OBIA, imagery is partitioned into homogeneous polygons that spatially represent real-world features such as vegetation, water bodies, or buildings [10]. The main advantage of OBIA is its ability to integrate spectral information with spatial, shape, and texture characteristics, making it more robust than conventional pixel-based classification [23].

The general workflow of OBIA begins with preprocessing, including radiometric correction, mosaicking, and defining the classification system. Next is image segmentation, where pixels with similar spectral and spatial properties are grouped into homogeneous segments called image objects. After segmentation, training samples representing the target land cover classes are extracted and used for classification with the RF algorithm. Level 1 classification identifies objects into basic categories, followed by level 2 refinement using contextual rules or ancillary data to achieve higher accuracy [24]. Statistical analysis then estimates the area per class and assesses accuracy using reference data or higher resolution imagery.

Segmentation quality directly affects subsequent analysis. Mean shift segmentation, a non-parametric algorithm that groups pixels based on spectral and spatial similarity without predefining the number of segments, was applied in this study [25]. The segmentation parameters used in this study are summarized in

Table 3. Segmentation parameters used in this study.

Parameter	Description	Value
Spatial radius	Defines the neighborhood search range in the spatial domain	5
Range radius	Sets the tolerance for spectral similarity among pixels	0.05
Minimum region size	Specifies the minimum object size to reduce noise	100
Maximum iterations	Ensures convergence during segmentation	100

.

The selection of these parameters followed a combination of literature guidance and empirical calibration. The spatial radius value of 5 pixels is consistent with previous mean shift applications on high-resolution imagery, which commonly use radii between 3–7 pixels to preserve object boundaries while avoiding excessive merging [26,27]. The maximum iteration setting (100) was adopted to ensure cluster convergence as recommended in similar OBIA–mean shift implementations [27], while maintaining acceptable computational time. The range radius (0.05) was calibrated based on the normalized spectral indices used in this study (NDVI, NDWI, RGB, and NIR), ensuring adequate separability between vegetation, built-up areas, and water. The minimum region size of 100 pixels was selected to correspond with the real-world minimum size of target objects (approximately 30–50 m²;), preventing fragmentation of vegetation patches and reducing noise. During the segmentation process, shadow regions characterized by low reflectance were sometimes delineated as separate objects when sufficiently large, whereas smaller shadow patches tended to merge with adjacent classes. These inconsistencies were later corrected during the post-classification manual refinement stage.

Geometric and statistical features were extracted from segmented objects, including area (A), perimeter (P), and compactness (C), which describes the closeness of an object’s shape to a circle:

$$C={\displaystyle \frac{4\pi A}{P^2}}$$

(3)

Values approaching 1 indicate more circular objects, while smaller values indicate elongated or irregular shapes [28]. Statistical features include the mean and standard deviation of pixel values within each object:

$$SD=\sqrt{{\displaystyle \frac{1}{N}}\sum _{i=1}^N{(x_i-\overline{x})}^2}$$

(4)

where N is the number of pixels in the object, $x_i$ is the value of each pixel, and $\overline{x}$ is the mean pixel value. These features are crucial for distinguishing spectrally homogeneous objects, such as dense vegetation, from more heterogeneous objects like urban areas.

The ensemble-based machine learning algorithm RF constructs multiple decision trees from random subsets of data and aggregates their predictions by majority voting or averaging [11]. RF is resistant to overfitting, effectively managing high-dimensional data [12]. The algorithm parameters were determined based on previous studies and refined by trial and error to optimize classification performance while maintaining computational 184 efficiency (

Table 4. Random Forest parameters used in this study.

Parameter	Description	Value
N_estimators	Number of trees in the forest	500
Max_depth	Maximum depth of each tree	15
Max_features	Number of predictor variables considered at each split	5
Min_samples_split	Minimum number of samples required to split an internal node	5
Random seed	Ensures reproducibility of model results	42

).

The land cover was classified into six target categories to capture the spatial characteristics of the campus environment, including both vegetated and non-vegetated areas: water, buildings, dense vegetation, sparse vegetation, bare soil, and roads. These classes were selected to represent the main land cover types influencing UGS distribution and dynamics.

Training samples for each class were manually selected through visual interpretation of high-resolution satellite imagery, ensuring that each point clearly represented its respective LULC class. In this study, the labeling process was initiated using the most recent imagery (2025), which provided the clearest visual delineation of objects such as newly constructed buildings, vegetation boundaries, and road networks. Stable training samples were then propagated backward to earlier years, while additional samples were added through visual interpretation when temporal changes were observed. This iterative and cumulative approach explains the fluctuation in training sample counts across years, reflecting both real landscape transformations and progressive enrichment of the training dataset. Comparability across years was maintained by preserving consistent sample distributions for stable classes while incorporating new samples only where land cover change occurred, ensuring that classification performance remained balanced and temporally coherent. The distribution of training samples for each year is summarized in

Table 5. Distribution of training data samples per LULC class and year.

Class	2015	2017	2021	2025
Water	16	31	16	13
Building	180	186	181	154
Dense vegetation	191	205	169	79
Sparse vegetation	118	157	123	65
Bare soil	78	145	64	28
Road	59	105	82	61
Total	642	829	635	400

.

Each class was then defined according to its spectral, spatial, and contextual characteristics, as detailed in

Table 6. Description of land cover classes used in the OBIA–RF classification.

Class	Description	Illustration
Water	Rivers, ponds, other water bodies
Buildings	Academic, residential, and infrastructure
Dense Vegetation	Forest canopy or thick tree cover
Sparse Vegetation	Scattered trees, grasslands, low-density cover
Bare Soil	Exposed soil, bare ground, construction areas
Roads	Paved or unpaved roads and pathways

. UGS was derived by combining dense and sparse vegetation classes to provide a focused metric of green space provision. This hierarchical approach from general LULC classification to a specific focus on UGS, allows comprehensive evaluation of vegetation cover changes and campus development trends between 2015 and 2025.

2.5. Accuracy Assessment Methodology

To ensure the reliability of the classification results, accuracy assessment was conducted using a set of complementary evaluation metrics. A confusion matrix was first generated to summarize the relationship between predicted and reference data [29]. In this matrix, the predicted data represent the land cover classes assigned by the classification algorithm, while the actual data correspond to the ground truth reference used for validation. The diagonal elements of the matrix indicate the number of correctly classified samples for each class, whereas the off-diagonal elements represent misclassifications between classes. The overall accuracy (OA) was calculated as the proportion of correctly classified samples relative to the total number of samples:

$$OA={\displaystyle \frac{{\sum }_{i=1}^{k}T{P}_i}{N}}$$

(5)

where $T{P}_i$ is the number of correctly classified samples for class i, k is the total number of classes, and N is the total number of samples. OA provides a general measure of classification performance but can be misleading in imbalanced datasets, where the accuracy may be dominated by the majority class [30].

The F1 score was employed to address class imbalance. Precision and recall were computed for each class, and the F1 score was derived as their harmonic mean [31]:

$$F1=2\times {\displaystyle \frac{Precision\times Recall}{Precision+Recall}}$$

(6)

In this study, the F1 Score was calculated in a weighted form, where the contribution of each class was weighted by the number of samples to ensure that both the majority and minority classes are proportionally represented in the final evaluation [32]. Additionally, Cohen’s Kappa coefficient was used to evaluate agreement beyond chance [33]:

$$\kappa ={\displaystyle \frac{p_o-p_e}{1-p_e}}$$

(7)

where $p_o$ is the observed agreement (equivalent to OA) and $p_e$ is the expected agreement by chance. Kappa is particularly valuable in multiclass classification settings and is frequently used in OBIA-based studies to validate model robustness [34].

A K-fold cross-validation strategy was applied to enhance reliability further [35]. The dataset was divided into four folds ($k=4$), selected to balance between computational efficiency and a sufficient validation sample size. The four-fold configuration was determined based on the smallest training class (water), ensuring that the number of validation samples in each iteration remained adequate for statistical evaluation. Consequently, each iteration used 75% of the samples for training and 25% for testing.

During each iteration, three folds were used to train the RF classifier and the remaining fold served as the validation subset. This process was repeated until every fold was used once as validation, and the final performance metrics (OA, weighted F1, and Kappa) were averaged across all folds, as shown in Figure 5. This cross-validation procedure reduces dependence on a single train–test split, providing a more stable and representative estimate of model performance [36].

After statistical validation using OA, Kappa, and weighted F1, ground truth data from 2025 were employed for an additional accuracy assessment. This step ensures that the evaluation not only depends on internal validation metrics but is also anchored to real-world conditions. Manual correction was subsequently applied in the post-processing stage to refine thematic consistency, followed by computation of the areal extent and statistical performance of each land cover class.

3. Results

3.1. Land Cover Classification and Area Statistics

The classification results for 2015, 2017, 2021, and 2025 are shown in Figure 6. Building areas were predominantly concentrated in the southern and central parts of the campus, while dense and sparse vegetation dominate the northern and peripheral zones. These spatial patterns remain relatively consistent across the years, with noticeable increases in vegetated coverage and minor infrastructure expansion.

The temporal distribution of land cover classes in Figure 7 clearly illustrates the dynamic changes in vegetation, building areas, and bare surfaces over the study period. Overall, UGS—comprising dense and sparse vegetation—increased steadily from 68.89% (1,243,174 m²) in 2015 to 74.69% (1,348,497 m²) in 2025, highlighting positive ecological progress within the campus. Sparse vegetation accounted for the largest share, expanding from 40.59% (732,570 m²) to 44.49% (803,243 m²), while dense vegetation recovered substantially after 2017, rising from 25.48% (459,780 m²) to 30.20% (545,254 m²) by 2025. In contrast, bare land declined sharply from 13.49% (243,413 m²) in 2015 to 5.81% (104,859 m²) in 2025, reflecting gradual land stabilization and conversion into vegetated or developed zones. Building areas exhibited moderate but consistent growth, increasing from 10.29% (185,758 m²) to 11.52% (207,929 m²), while road networks expanded slightly from 6.56% (118,301 m²) to 6.65% (120,098 m²), peaking in 2021 at 7.65% (138,065 m²). The building coverage ratio (BCR) decreased from 31.15% (562,220.6 m²) in 2015 to 25.31% (456,955.3 m²) in 2025, aligning with the observed greening trend. Meanwhile, water bodies remained stable (below 1.5%), increasing slightly from 0.82% (14,748.6 m²) in 2015 to 1.33% (24,069.3 m²) in 2025 following the construction of the artificial lake after 2017.

The distribution patterns shown in Figure 7 confirm that vegetation consistently dominated campus land use throughout the study period. Despite localized infrastructure expansion, the overall trend indicates enhanced ecological balance, with substantial vegetation recovery compensating for minor increases in building and road areas. These findings underline the effectiveness of spatial planning and green management policies that prioritize UGS maintenance while accommodating necessary campus development.

3.2. UGS Dynamics in Key Areas

Figure 8 shows the spatial dynamics in areas that experienced significant building expansion between 2015 and 2025, including the bank, hospital, and chicken house. Each subfigure illustrates temporal changes in land cover distribution for the respective sites.

Figure 9 presents the corresponding area statistics for each land cover class within these sites. The building area in the bank region increased from 0 m² in 2015 to 3510.21 m² in 2021, followed by a slight decrease to 3440.87 m² in 2025. In the hospital area, the building class expanded considerably from 2586.94 m² to 7501.13 m², accompanied by a decline in bare land from 3358.15 m² to 822.77 m². The chicken house area was developed, reaching 2164.09 m².

In contrast, Figure 10 displays land cover changes in regions characterized by notable UGS expansion, including the Ecoriparian, Faculty of Fisheries and Marine Sciences, and Faculty of Economics and Business areas. Each subfigure visualizes vegetation growth and associated land transformations over time.

Figure 11 presents the corresponding area statistics for the major land cover classes within the Ecoriparian, Faculty of Fisheries and Marine Sciences, and Faculty of Economics and Business regions. In Ecoriparian, dense vegetation expanded from 11,925.5 m² in 2015 to 15,125.5 m² in 2025, indicating significant greening efforts. The Faculty of Fisheries and Marine Sciences area also showed vegetation growth, with sparse vegetation increasing from 8063.24 m² to 16,756.3 m². Similarly, the Faculty of Economics and Business experienced a rise in sparse vegetation from 2750.37 m² to 9548.97 m², accompanied by a sharp reduction in bare land from 5179.92 m² to 36 m².

3.3. Accuracy Assessment Result

The classification accuracy was evaluated using confusion matrices for the four reference years (2015, 2017, 2021, and 2025). The results are shown in Figure 12, where the rows represent the reference classes and the columns represent the predicted classes.

Overall, the confusion matrices indicate that the classification models produced reliable results across all years, with clear class separability in most land cover categories.

Table 7. Classification accuracy metrics (OA, Weighted F1, and Kappa) for each year.

Year	OA	Weighted F1	Kappa
2015	0.844	0.845	0.799
2017	0.860	0.860	0.826
2021	0.836	0.835	0.790
2025	0.810	0.807	0.747

presents OA, weighted F1 score, and Kappa coefficient for each year.

To provide a comparative benchmark, pixel-based Random Forest classification was also evaluated for the same four years. The pixel-based approach resulted in consistently lower accuracy than the OBIA-based method, confirming that object segmentation improves class separability and reduces spectral noise. The accuracy results for the pixel-based model are summarized in

Table 8. Pixel-based Random Forest accuracy metrics (OA, Weighted F1, and Kappa).

Year	OA	Weighted F1	Kappa
2015	0.736	0.730	0.656
2017	0.773	0.770	0.717
2021	0.715	0.712	0.634
2025	0.685	0.680	0.579

.

In addition, a ground truth validation was conducted to further confirm spatial consistency. A total of 43 ground truth points were collected in 2025 using a 200-m grid spacing across the study area. These points were compared against the OBIA RF classification results before manual correction to evaluate the machine learning accuracy independently of post-processing edits. Figure 13 illustrates the spatial distribution of the ground truth sampling locations.

Table 9. Ground truth reference points and corresponding mapped land cover classes for accuracy validation (year 2025). Coordinates are given in decimal degrees (EPSG:4326).

Id	Longitude	Latitude	Actual Class	Predicted Class
1	107.7725243	−6.932253645	Sparse Vegetation	Sparse Vegetation
2	107.7725137	−6.930446579	Dense Vegetation	Dense Vegetation
3	107.7743221	−6.930436016	Dense Vegetation	Dense Vegetation
4	107.7761304	−6.930425446	Building	Building
5	107.7779388	−6.930414869	Building	Building
6	107.7779282	−6.928607811	Sparse Vegetation	Sparse Vegetation
7	107.7761199	−6.928618385	Sparse Vegetation	Sparse Vegetation
8	107.7743115	−6.928628952	Building	Building
9	107.7725032	−6.928639512	Dense Vegetation	Dense Vegetation
10	107.7706843	−6.926842996	Dense Vegetation	Dense Vegetation
11	107.7724926	−6.926832446	Dense Vegetation	Dense Vegetation
12	107.7743009	−6.926821888	Sparse Vegetation	Sparse Vegetation
13	107.7761093	−6.926811324	Road	Road
14	107.7779176	−6.926800752	Building	Building
15	107.7760987	−6.925004262	Sparse Vegetation	Sparse Vegetation
16	107.7742904	−6.925014824	Bare Soil	Bare Soil
17	107.7724820	−6.925025379	Road	Road
18	107.7706737	−6.925035926	Sparse Vegetation	Sparse Vegetation
19	107.7706631	−6.923228857	Road	Road
20	107.7724715	−6.923218311	Building	Building
21	107.7742798	−6.923207759	Building	Building
22	107.7742692	−6.921400695	Building	Building
23	107.7724609	−6.921411244	Road	Road
24	107.7706526	−6.921421786	Bare Soil	Bare Soil
25	107.7706421	−6.919614716	Road	Road
26	107.7724504	−6.919604177	Building	Building
27	107.7742481	−6.917786566	Dense Vegetation	Dense Vegetation
28	107.7724398	−6.917797109	Sparse Vegetation	Sparse Vegetation
29	107.7706315	−6.917807646	Bare Soil	Bare Soil
30	107.7688127	−6.916011103	Sparse Vegetation	Sparse Vegetation
31	107.7706210	−6.916000575	Dense Vegetation	Dense Vegetation
32	107.7724293	−6.915990041	Dense Vegetation	Dense Vegetation
33	107.7742376	−6.915979501	Dense Vegetation	Dense Vegetation
34	107.7742270	−6.914172435	Sparse Vegetation	Sparse Vegetation
35	107.7724187	−6.914182974	Sparse Vegetation	Sparse Vegetation
36	107.7706104	−6.914193505	Building	Building
37	107.7688021	−6.914204029	Dense Vegetation	Dense Vegetation
38	107.7687916	−6.912396956	Bare Soil	Bare Soil
39	107.7705999	−6.912386434	Dense Vegetation	Dense Vegetation
40	107.7724082	−6.912375906	Dense Vegetation	Dense Vegetation
41	107.7742165	−6.912365370	Sparse Vegetation	Sparse Vegetation
42	107.7723977	−6.910568837	Sparse Vegetation	Sparse Vegetation
43	107.7705894	−6.910579363	Dense Vegetation	Dense Vegetation

presents the ground truth validation results, showing the correspondence between field-observed classes and the classified land cover map for 2025. All field photographs and location documentation are provided in the Appendix A (Figure A1) for reference.

4. Discussion

4.1. UGS Transformation and Spatial Dynamics

UGS transformation within Unpad’s Jatinangor campus from 2015 to 2025 shows a gradual yet consistent ecological improvement accompanied by controlled infrastructure expansion. Between 2015 and 2017, the total UGS slightly decreased from 68.89% (1,243,174 m²) to 68.57% (1,237,387 m²), mainly due to an increase in bare land and minor construction activities, while dense vegetation decreased from 28.29% (510,604 m²) to 25.48% (459,780 m²) (Figure 7). This early phase likely represents the initial stage of development that temporarily reduced vegetated areas before reestablishment. From 2017 to 2021, UGS began to recover, rising from 68.57% (1,237,387 m²) to 71.92% (1,297,952 m²), with dense vegetation notably increasing by over 2% from 459,780 m² to 502,578 m², indicating early signs of regreening across several parts of the campus (Figure 7). During this period, bare land decreased significantly from 13.44% (242,520 m²) to 8.66% (156,233 m²), suggesting successful land cover restoration efforts. Building areas remained relatively stable, increasing only slightly from 10.36% (186,949 m²) to 10.41% (187,927 m²), showing that campus infrastructure growth was managed without major spatial expansion. The most substantial improvement occurred between 2021 and 2025, when UGS coverage reached 74.69% (1,348,497 m²), with both dense and sparse vegetation increasing while bare surfaces continued to shrink to only 5.81% (104,859 m²) (Figure 6 and Figure 7). Building areas grew moderately from 10.41% (187,927 m²) to 11.52% (207,929 m²), but the increase was proportionally small compared to vegetation gains.

Overall, UGS expanded by nearly 6% from 1,243,174 m² in 2015 to 1,348,497 m² in 2025, whereas bare land declined by more than half, from 13.49% (243,413 m²) to 5.81% (104,859 m²) (Figure 6 and Figure 7). This inverse trend between vegetated and bare surfaces highlights a decade-long shift toward more sustainable land management. Meanwhile, the overall BCR decreased from 31.15% (562,220.6 m²) to 25.31% (456,955.3 m²), indicating that despite new infrastructure, green spaces continued to dominate the campus area (Figure 7). These results suggest that Unpad has successfully balanced academic development with environmental preservation, strengthening both ecological resilience and spatial sustainability within the Jatinangor campus.

From a policy perspective, the dominance of vegetated areas across the Unpad campus, which consistently exceeded 70% throughout this period, demonstrates strong alignment with Indonesia’s national spatial planning regulation, which mandates at least 30% of urban areas to function as UGS according to Law No. 26 of 2007 on Spatial Planning. This compliance indicates a firm institutional commitment to sustainable spatial governance and provides an exemplary case of localized implementation of the UGS policy within an academic environment. Maintaining such a high proportion of green coverage enhances ecological resilience against urban heat and surface runoff, while reinforcing the university’s role as a living laboratory for sustainable development. To ensure that this performance is preserved amid future campus expansion, spatial planning strategies should continue to emphasize UGS protection through vertical greening initiatives, the preservation of riparian buffer zones, and the optimization of existing vegetated areas in accordance with both national and regional spatial frameworks.

A comparison with previous studies further reinforces the significance of Unpad’s UGS performance. Earlier internal assessments using NDVI thresholding [9] produced UGS proportions of 69.84% in 2015 and 69.31% in 2017, which closely align with the OBIA–RF results of this study (68.89% and 68.57%). The high level of agreement indicates that despite differences in methodological approaches, the overall trajectory of UGS change remains consistent, strengthening the reliability of the observed long-term greening trend. Broader comparisons with other university campuses also highlight the uniqueness of this pattern. At Banaras Hindu University in India, dense vegetation experienced a dramatic decline between 2008 and 2018, falling from 11.14% to only 2.68%. More importantly, total vegetated land (dense vegetation + thorny bushes) dropped from 70.56% to 38.43%, a reduction of approximately 32% over a single decade due to rapid land conversion [37]. A similar trend was reported at the University of Baghdad, where combined low, medium, and dense vegetation decreased from 64% in 1988 to 61% in 2022 [38]. These findings contrast with the Jatinangor campus, where UGS increased from 68.89% in 2015 to 74.69% in 2025 despite ongoing infrastructure development. This comparison underscores that, unlike many universities experiencing long-term vegetation decline under development pressure, Unpad demonstrates a rare and successful model of sustained ecological restoration and green-space governance.

Although several new infrastructures were developed during the study period, the increase in building areas remained relatively small compared to the significant expansion of vegetated zones (Figure 8 and Figure 9). This balance reflects how development at Unpad was directed toward optimizing existing spaces rather than uncontrolled land conversion. Spatially, six primary areas of interest (AOIs) consisting of the bank, hospital, chicken house, Ecoriparian, Faculty of Fisheries and Marine Sciences, and Faculty of Economics and Business areas illustrate contrasting patterns between building and vegetated changes. The bank, hospital, and chicken house regions represent focal points of infrastructure growth. The establishment of the bank in 2017 contributed to an increase of approximately 3000–3500 m² in building area, with a corresponding decline in sparse vegetation (Figure 8a–d and Figure 9a). The hospital zone recorded the largest expansion, exceeding 7500 m² of new construction by 2025 (Figure 8e–h and Figure 9b), reflecting the major enhancement of healthcare facilities within the campus. Similarly, the chicken house area, which was initially dominated by vegetation, transitioned to include more than 2000 m² of new buildings by 2025 (Figure 8i–l and Figure 9c), indicating the establishment of new academic or operational infrastructures. In contrast, the other three AOIs, namely Ecoriparian, Faculty of Fisheries and Marine Sciences, and Faculty of Economics and Business, exhibited notable vegetation recovery. The Ecoriparian zone recorded dense vegetation growth exceeding 15,000 m² by 2025 (Figure 10a–d and Figure 11a), replacing previously open or bare surfaces and demonstrating the success of riparian rehabilitation efforts. The Faculty of Fisheries and Marine Sciences showed a steady increase in sparse vegetation (Figure 10e–h and Figure 11b), largely resulting from the restoration of previously compacted or paved areas, while the Faculty of Economics and Business transformed its bare grounds into green cover, reducing bare land to nearly zero by 2025 (Figure 10i–l and Figure 11c).

In addition to natural regeneration or landscape changes that may influence vegetation dynamics, the consistent increase in green cover observed in this study is strongly associated with the long-term, planned greening initiatives implemented by Unpad since 2016. These initiatives were carried out in multiple phases over nearly a decade, including the planting of 32,956 seedlings during 2016–2017, continued planting with 900 seedlings in 2018–2019, further large-scale greening activities involving 7937 seedlings between 2020–2022, and collaborative tree-planting programs around key ecological areas with 301 seedlings in 2023–2024 and 200 additional seedlings in 2025. This sustained and progressive annual intervention demonstrates that the improvement in vegetation is not solely the result of passive ecological processes, but is predominantly driven by structured restoration programs, strategic landscape management, and stakeholder-supported environmental initiatives. Consequently, the positive vegetation trend identified in the classification results reflects UNPAD’s active and continuous commitment to ecological enhancement and long-term campus sustainability.

Overall, these localized changes reflect a dual approach in campus spatial planning: targeted infrastructure expansion in specific functional areas accompanied by consistent ecological restoration in environmentally sensitive and academic zones. Despite localized construction, the overall campus transformation remained dominated by UGS expansion, maintaining a green ratio well above the national minimum standard of 30% (Figure 7, Figure 8 and Figure 10). This outcome underscores Unpad’s success in integrating development needs with sustainable land management, ensuring both spatial functionality and ecological resilience.

4.2. Accuracy and Sustainability Implications

The accuracy assessment conducted for the LULC classification in Unpad Jatinangor campus demonstrates a high level of reliability across all reference years (2015, 2017, 2021, and 2025), as presented in Figure 12 and Table 7. The OA ranged from 0.810 in 2025 to 0.860 in 2017, which is considered good according to standard remote sensing classification guidelines [39]. The Kappa coefficient varied between 0.747 and 0.826, indicating substantial agreement beyond chance [40]. Similarly, the weighted F1 score, which combines precision and recall for each class, closely mirrors the OA values, reflecting reliable per-class performance and a well-balanced classifier [31]. Analysis of the confusion matrices reveals that certain classes achieved particularly high accuracy. Dense vegetation and building areas consistently showed the highest classification reliability, with minimal misclassification to other categories (Figure 12). Sparse vegetation and bare land, while slightly more prone to confusion, still maintained high per-class accuracy, demonstrating the reliability of the RF classifier employed in the study. The results indicate that the model effectively captured both highly vegetated and developed regions, which are critical for monitoring UGS expansion and building encroachment.

To further contextualize the performance, a comparative baseline was established using a pixel-based Random Forest classifier applied to the same years (Table 8). The pixel-based approach consistently underperformed relative to OBIA–RF, with OA ranging from 0.685 in 2025 to 0.773 in 2017, Weighted F1 scores from 0.680 to 0.770, and Kappa coefficients from 0.579 to 0.717. This corresponds to an improvement of approximately 12–13% in OA, 12–13% in Weighted F1, and 13–17% in Kappa when employing OBIA–RF compared to the pixel-based approach across the four years. The advantage of object-based segmentation is particularly pronounced for classes exhibiting high textural heterogeneity, such as sparse vegetation and bare land, which were frequently misclassified under the pixel-based approach due to overlapping spectral signatures. Built-up areas also exhibited clearer and more spatially coherent boundaries under the OBIA–RF framework. These findings confirm that OBIA–RF is better suited for detailed, campus-scale LULC monitoring using high-resolution imagery, offering superior classification accuracy and enhanced thematic consistency for long-term UGS assessment.

Following the automated classification, a manual correction was applied to ensure spatial consistency, particularly in ambiguous boundary areas between classes. This step was crucial to align the LULC map with visual interpretation from high-resolution imagery. Ground truth validation was then conducted using 43 sampling points collected in 2025 (Figure 13 and Table 9), which were kept fully independent and were not affected by the manual correction process.Although the classified map showed a high degree of agreement with the ground-observed points, it cannot be regarded as a perfect match due to the limited number of validation samples relative to the total campus area. Validation was performed by visually comparing each sampling location against very high-resolution imagery and on-site observations, ensuring that the assessment reflected the classifier’s performance rather than the effects of manual editing. Despite the small sample size, the results indicate a strong correspondence between classification outputs and field data, suggesting that the final LULC map provides a reliable representation of on-ground conditions. The validated maps therefore serve as a dependable basis for monitoring UGS dynamics and assessing spatial development on campus. The reliable detection of vegetation and built-up expansion supports informed decision-making for conservation and infrastructure planning, contributing to the university’s long-term sustainability strategy.

Overall, the integration of quantitative accuracy metrics, confusion matrix analysis, and an independent ground truth validation performed prior to manual correction ensures that the LULC dataset is scientifically reliable. The subsequent manual correction further enhances thematic consistency without influencing the validation results. Such high confidence in classification outcomes supports evidence-based decision-making, aligning spatial monitoring with the university’s objectives for maintaining a green and sustainable campus environment.

4.3. Limitations and Future Improvements

Despite the strong performance of the OBIA–RF classification and the overall consistency of long-term UGS trends, several limitations should be acknowledged to appropriately contextualize the findings of this study. First, the analysis depends heavily on the quality and temporal consistency of high-resolution satellite imagery. Variability in illumination, seasonal conditions, and shadowing, particularly under dense canopy cover, may reduce the classifier’s ability to differentiate sparse vegetation, bare land, and built-up features with precision. Some narrow structures, small rooftops, and pedestrian pathways were also partially obscured by tree crowns, resulting in occasional misclassifications even after manual refinement. Additionally, the LULC scheme employed in this study includes only four major categories. Although these categories are adequate for campus-scale UGS monitoring, they do not capture finer ecological distinctions such as grasslands, shrublands, water bodies, or mixed vegetation. This simplification may underrepresent the campus’s ecological heterogeneity, indicating the need for future studies to adopt a more detailed class hierarchy or incorporate multi-temporal phenological metrics to improve vegetation subtype discrimination.

Beyond classification considerations, this study does not integrate landscape ecology metrics such as fragmentation, patch connectivity, edge density, or core-area analysis. These indicators are essential for evaluating the ecological functionality of UGS beyond its spatial extent, particularly in relation to habitat continuity, microclimatic regulation, and long-term ecosystem resilience. Incorporating landscape metrics in future research would therefore provide deeper insight into the spatial quality and ecological stability of the campus green infrastructure. Finally, ground truth data were collected only for 2025, and the number of validation samples was limited relative to the total study area. Although the resulting accuracy metrics were sufficiently high, a more robust sampling strategy that includes stratified sampling and multi-year field observations would enhance the reliability of temporal comparisons. Addressing these limitations in future work will strengthen the reliability of UGS monitoring and support a more comprehensive understanding of ecological dynamics across the Unpad campus.

5. Conclusions

This study assessed the dynamics of UGS at Unpad’s Jatinangor campus over a ten-year period (2015–2025) using an object-based RF classification approach applied to high-resolution imagery. The analysis revealed a consistent increase in total UGS coverage from 68.89% (1,243,174 m²) in 2015 to 74.69% (1,348,497 m²) in 2025, accompanied by substantial recovery of dense and sparse vegetation and a marked reduction in bare land. Building areas increased moderately from 10.36% (185,758 m²) in 2015 to 11.52% (207,929 m²) in 2025, but remained proportionally small relative to vegetation gains, demonstrating that campus expansion has been managed without compromising ecological integrity. Compared to other universities where UGS declined over similar periods, such as Banaras Hindu University and University of Baghdad, the increase in UGS at Unpad is notably more significant, highlighting a rare success in balancing campus development with ecological restoration. Accuracy assessment confirmed that the classification model reliably captured UGS and building areas, with dense vegetation and building classes showing the highest per-class accuracy. Moreover, the object-based RF approach improved overall classification accuracy by approximately 12–17% compared to a pixel-based RF method, particularly enhancing the detection of sparse vegetation, bare land, and building boundaries. An independent ground truth validation using 43 sampling points in 2025 showed strong agreement between mapped and observed classes. These results indicate that the LULC dataset is robust, spatially accurate, and suitable for supporting sustainable campus planning.

The study findings have several important implications. The observed UGS expansion exceeding the national minimum standard of 30% demonstrates the university’s success in integrating environmental preservation with infrastructure development, thereby enhancing ecological resilience and spatial sustainability. The reliable and validated LULC maps provide an evidence-based foundation for informed decision-making, enabling campus authorities to identify areas for conservation, monitor vegetation dynamics, and strategically plan new infrastructure with minimal impact on green spaces. Overall, this research contributes both theoretically by demonstrating the applicability of OBIA and random forest for campus-scale studies and practically by providing spatially explicit information to support the university’s green campus initiatives.