CLOUD-BASED MULTI-TEMPORAL SAR-OPTICAL FUSION USING RANDOM FOREST FOR GEOSPATIAL INTELLIGENCE MONITORING OF KIPP IKN NUSANTARA

Geographia Technica, Vol. 22, Issue 1, 2027, pp 106 to 119

CLOUD-BASED MULTI-TEMPORAL SAR-OPTICAL FUSION USING RANDOM
FOREST FOR GEOSPATIAL INTELLIGENCE MONITORING
OF KIPP IKN NUSANTARA

Fajar Sidik SUGANDA 1* & Asep Adang SUPRIYADI 1

DOI: 10.21163/GT_2027.221.08

ABSTRACT
The development of Indonesia's new capital (IKN) Nusantara drives dramatic land -cover
transformation across a strategic 17,599 -hectare area under the persistent cloud cover of Kalimantan,
where optical remote sensing alone is insufficient for continuous mo nitoring. This study proposes a
cloud-based multi -temporal SAR -optical fusion approach using Random Forest classification for
geospatial-intelligence (GEOINT) monitoring of the Core Government Area (KIPP) of IKN. Using
Google Earth Engine, we processed 70 Sentinel-1 GRD and 52 Sentinel-2 images for 2022 and 61 and
81 images for 2024, building a 12 -band fusion stack (SAR backscatter VV, VH, VV/VH; optical B2,
B3, B4, B8, B11, B12; and NDVI, NDBI, NDWI). Six experiments compared SAR-only, optical-only,
and fusion configurations. The Random Forest fusion achieved 88.89% overall accuracy and 0.8609
kappa for 2022 and 85.12% and 0.8127 for 2024 (assessed against ESA WorldCover 2021 class
definitions; see Section 2.7), outperforming SAR -only (63.20%) and optical -only (86.44%). Change
detection revealed 899.64 ha of direct vegetation-to-built-up conversion, a 60% expansion of built-up
area driven by IKN construction. This is the first cloud-based multi-temporal fusion framework applied
to KIPP IKN monitoring, with i mplications for spatial planning, security surveillance, and
environmental compliance.
Keywords: sar-optical fusion; random forest; Google Earth Engine; geospatial intelligence ; IKN
Nusantara.

1. INTRODUCTION
   Land use and land cover (LULC) classification using satellite remote sensing is an indispensable
   tool for monitoring environmental change, urban expansion, and infrastructure development (Pande
   et al., 2024), (Fitri et al., 2025). Free Sentinel -1 SAR and Sentinel-2 imagery, combined with cloud
   platforms such as Google Earth Engine (GEE), has democratized large-scale geospatial analysis (Latif
   et al., 2025). The Random Forest (RF) algorithm performs robustly in LULC mapping through its
   handling of high-dimensional features and resistance to overfitting (Dagne et al., 2023), (Al -Dousari
   et al., 2023), and fusing complementary SAR and optical information improves accuracy over single-
   modality approaches by combining SAR's all-weather capability with optical's spectral richness (Irfan
   et al., 2025), (Y. Li & Xiao, 2025). State-of-the-art Sentinel-2 classification reports 75-85% accuracy
   regionally, with multi-temporal analysis increasingly critical for capturing dynamic landscapes (Jocea
   et al., 2025).
   Geospatial intelligence (GEOINT) is the exploitation and analysis of imagery and geospatial
   information to describe, assess and visually depict physical features and geographically referenced
   activities on the Earth in support of decision -making (Yusfan et al., 2021; Djati & Gunawan, 2025).
   In this study GEOINT refers to the use of multi-temporal land-cover change over a strategic national
   asset the KIPP of IKN to derive indicators for spatial-planning verification and security-perimeter
   monitoring. Despite these advances, several gaps remain in applying SAR -optical fusion to rapidly
   transforming strategic infrastructure in tropical regions.

---

1 Sensing Technology Study Program, Faculty of Defense Science and Technology, Republic of Indonesia
Defense University, Sentul, West Java, Indonesia. * Corresponding author: fajarsidiksugandaa@gmail.com
(FJS); aadangsupriyadi@gmail.com (AAS).

107

First, fusion has mostly been demonstrated in temperate or sub -tropical contexts; its
generalizability to tropical Indonesian landscapes with persistent cloud cover (often 60 -80% year-
round) is underexplored.
Second, change-detection studies typically address long -term decadal analysis (Jagannathan et
al., 2025), yet megaproject construction produces major change within 2 -3 years, requiring shorter -
interval frameworks that are seldom evaluated. Third, comprehensive fusion-based monitoring of the
Core Government Area (KIPP) of Indonesia's new capital (IKN) -the largest greenfield capital
relocation in modern history -has not been published. Fourth, while urban built -up change detection
exists (Feng et al., 2021), f rameworks designed specifically for geospatial -intelligence (GEOINT)
applications on strategic national infrastructure are absent for Indonesian contexts (Yusfan et al.,
2021), (Djati & Gunawan, 2025), (Yusuf Samad et al., 2023).
Filling these gaps is urgent: during 2022 -2024 KIPP IKN underwent intensive construction -
Istana Garuda, Sumbu Kebangsaan, and Kemenko -transforming roughly 17,000 ha of Kalimantan
rainforest and demanding evidence -based monitoring (Cholidah et al., 2024). T his study proposes a
cloud-based multi -temporal SAR -optical fusion framework using Random Forest for GEOINT
monitoring of KIPP IKN, processed entirely in Google Earth Engine. Its objectives are: (1) to develop
a 12-band SAR-optical fusion stack and evaluate Random Forest performance against single-modality
baselines across two timeframes; (2) to quantify 2022 -2024 land -cover transitions revealing
GEOINT-relevant urban-transformation patterns; (3) to provide an actionable monitoring framework
supporting spatial planning and security surveillance for similar strategic areas (Adi et al., 2025).
2. MATERIALS AND METHODS
2.1. Study Area
The study area is the Kawasan Inti Pusat Pemerintahan (KIPP) of Ibu Kota Nusantara (IKN), in
Sepaku Subdistrict, Penajam Paser Utara Regency, East Kalimantan ( Fig. 1 ). A bounding box of
~17,599 ha (116.650 -116.780 E, 1.030 -0.920 S) encompasses the 6,671 -ha official KIPP boundary
and its immediate context, including Pemaluan, Bumi Harapan, and Bukit Raya villages. This
tropical-rainforest landscape has transformed rapi dly since IKN construction began in 2022, with
landmarks such as Istana Garuda and the Sumbu Kebangsaan boulevard built within the study period.

Fig. 1. Study area showing the Kawasan Inti Pusat Pemerintahan (KIPP) of Indonesia's new capital city (IKN)
Nusantara.

Fajar Sidik SUGANDA and Asep Adang SUPRIYADI / CLOUD-BASED MULTI-TEMPORAL … 108

2.2. Data Acquisition
Two complementary imagery sources were accessed through Google Earth Engine ( Fig. 2 )
(Nigar et al., 2024). Sentinel -1 SAR Ground Range Detected (GRD) imagery was acquired in
Interferometric Wide swath mode, descending orbit, dual polarization (VV, VH), yielding 70 images
for 2022 and 61 for 2024, consistent with multi -temporal SAR monito ring for Indonesian contexts
(Aldiansyah et al., 2024). Sentinel-2 Surface Reflectance Harmonized imagery filtered to <60% cloud
cover yielded 52 images for 2022 and 81 for 2024. ESA WorldCover v200 (10 m, 2021) provided
ground reference as pseudo -labels, remapped to five classes: built -up (0), bare land (1), vegetation
(2), cropland (3), and water (4) (Carneiro et al., 2025). Both collections were filtered to the study-area
boundary with GEE's filterBounds and filterDate.

Fig. 2. Sample multi-modal input data over the KIPP IKN study area.
The cloud-based strategy retrieved a comprehensive multi-temporal dataset (Table 1): 70 and 61
Sentinel-1 GRD images and 52 and 81 Sentinel -2 SR images for 2022 and 2024 respectively (60%
cloud threshold), over a 17,599.25-ha area confirmed by GEE's pixelArea reducer, encompassing the
6,671-ha KIPP boundary plus surrounding context.
Table 1.
Multi-temporal satellite imagery acquisition summary.
Parameter 2022 (baseline) 2024 (current) Specification
Sentinel-1 GRD images 70 61 IW mode, descending
Sentinel-2 SR images 52 81 Cloud cover < 60%
S1 polarization VV + VH VV + VH Plus VV/VH ratio
S2 spectral bands 6 bands 6 bands B2, B3, B4, B8, B11, B12
S2 cloud masking QA60 QA60 Bits 10 & 11
ROI total area 17,599.25 ha 17,599.25 ha Bounding box

109

2.3. Preprocessing and Fusion Stack Construction
Sentinel-1 imagery was provided as backscatter in decibels following ESA's standard processing
(thermal-noise removal, radiometric calibration, terrain correction). A 3x3 focal -mean filter was
applied to the median composite for speckle reduction, and a VV /VH ratio band (VV.divide(VH))
was added to enhance textural discrimination. Sentinel-2 imagery was cloud- and cirrus-masked using
the QA60 band (bits 10-11), then median-composited into a cloud-free annual composite. The optical
composite comprised six bands (B2, B3, B4, B8, B11, B12) plus three normalized-difference indices
(Aos & Putri, 2023): NDVI = (B8 -B4)/(B8+B4), NDBI = (B11 -B8)/(B11+B8), and NDWI = (B3 -
B8)/(B3+B8). The final fusion stack comprised 12 bands integrating SAR (3) and optical (9)
information per period. In this study, “multi -temporal” denotes the two annual epochs (2022 and
2024) used for post -classification change detection, not intra -annual seasonal compositing: for each
epoch every cloud -masked scene of the calendar year was reduced to a single per -pixel median
composite. Seasonal (e.g. quarterly) sub-composites were not constructed because the persistent cloud
cover over KIPP IKN leaves too few clear observations per quarter for a stable median, whereas an
annual median maximises the number of valid observations per pixel.
2.4. Training Sample Generation
To ensure spatially representative samples with minimal manual labelling, stratified random
sampling was applied to the ESA WorldCover label image, following auxiliary -data integration in
heterogeneous-landscape classification (Shandu et al., 2026). GEE's stratifiedSample generated 5,000
points (1,000 per class; seed 42), partitioned into training (70%), validation (15%), and testing (15%).
After feature extraction at 10 m, the final dataset comprised 3,528 training, 779 validation, and 693
testing points. ESA WorldCover 2021 is used solely to locate the stratified sample points; it is not
used as a wall-to-wall validation layer or overlaid on the classified maps.
The stratified sampling produced 5,000 points across five classes ( Table 2). The distribution is
extremely imbalanced, typical of pristine rainforest: vegetation dominated at 92.1% (16,287.95 ha)
while built -up was merely 0.2% (35.81 ha), confirming KIPP as effectively greenfield before
construction.
Table 2.
Class distribution in study area (2021 ESA WorldCover baseline).
Class Code Area (ha) Percentage Sample Points
Vegetation 2 16,287.95 92.1% 1,000
Water body 4 774.86 4.4% 1,000
Bare land 1 336.12 1.9% 1,000
Cropland 3 164.51 0.9% 1,000
Built-up 0 35.81 0.2% 1,000
Total - 17,599.25 100% 5,000
2.5. Random Forest Classification
The Random Forest (RF) algorithm was used for its robustness against overfitting, capacity for
high-dimensional features, and proven superiority over tree -based alternatives such as CART for
Sentinel-2 imagery (Gulci et al., 2025), (Syaban & Appiah -Opoku, 2024), (Kasahun & Legesse,
2024). Six experiments evaluated each modality's contribution:
RF SAR-only 2022 (3 bands: VV, VH, ratio)
RF Optical-only 2022 (9 bands: 6 spectral + 3 indices)
RF Fusion 2022 (12 bands combined) Proposed
RF SAR-only 2024
RF Optical-only 2024
RF Fusion 2024 - Proposed

Fajar Sidik SUGANDA and Asep Adang SUPRIYADI / CLOUD-BASED MULTI-TEMPORAL … 110

The classifier was implemented using GEE's ee.Classifier.smileRandomForest with 100 trees and
default hyperparameters for other settings, a configuration validated in recent multi-temporal Sentinel
land cover studies (Lemenkova, 2025). Training was perform ed using the 70% split, with each pixel
feature vector treated as an independent sample.
2.6. Multi-temporal Change Detection
Post-classification change detection encoded pairwise transitions between the two periods. A 5x5
transition matrix was built by computing pixel -wise area for every from -to class combination using
GEE's ee.Image.pixelArea() -robust for tropical regions where cloud cover challenges single -date
analyses. GEOINT -relevant transitions were flagged: vegetation -to-built-up (direct deforestation),
bare-to-built-up (completed construction), and vegetation -to-bare (pre -construction clearing).
Classified and change maps were exported as GeoTIFFs in EPSG:32750 at 10 m.
2.7. Accuracy Assessment
Accuracy was assessed from the error matrix: Overall Accuracy (OA), Cohen's Kappa, and per -
class Producer's and User's Accuracy, computed independently for each of the six configurations on
the held-out 15% test subset (Rangel et al., 2025). Both epochs are trained and tested against the same
ESA WorldCover 2021 label set; the 2024 sample points were not re -labelled. Because the 2024
imagery post-dates the reference by three years, the 2024 accuracy is therefore assessed against 2021-
vintage class definitions and carries label noise on the ~9% of the study area that genuinely changed
land cover in the interim (Section 3.5). The 2024 figures should be read relative to the WorldCover -
2021 definition rather than as a strict like-for-like comparison with 2022, whereas the one-year offset
makes this effect negligible for 2022. A reference -independent check of the 2024 result against
Dynamic World, which does not use ESA WorldCover, is reported in Section 3.4.1.

3. RESULTS
   3.1. Random Forest Classification Performance
   The six experiments consistently supported the hypothesis that SAR -optical fusion outperforms
   single-modality approaches (Table 3). For 2022, RF Fusion achieved 88.89% OA and 0.8609 Kappa-
   a 2.45-point improvement over optical -only (86.44%, kappa=0.8303) and 25.69 points over SAR -
   only (63.20%, kappa=0.5399); a McNemar test confirmed the fusion-over-optical gain was significant
   (chi2=4.83, p=0.028). For 2024 the hierarchy held-Fusion 85.12% OA, 0.8127 Kappa, versus optical-
   only (83.68%, kappa=0.7949) and SAR -only (62.40%, kappa=0.5286) -but the fusion -over-optical
   gain was smaller and not significant (chi2=1.69, p=0.19), indicating that SAR's principal benefit under
   rapidly changing conditions is robustness rather than a large accuracy margin. The modest 3.77-point
   decline from 2022 to 2024 reflects the mismatch between the 2021 labels and the evolving 2024
   landscape.
   Table 3.
   Random Forest classification performance across modalities and periods.
   Method 2022 OA (%) 2022 Kappa 2024 OA (%) 2024 Kappa
   RF SAR-only 63.20 0.5399 62.40 0.5286
   RF Optical-only 86.44 0.8303 83.68 0.7949
   RF Fusion (proposed) 88.89 0.8609 85.12 0.8127
   Δ (Fusion vs Optical) +2.45 +0.0306 +1.44 +0.0179
   Δ (Fusion vs SAR) +25.69 +0.3210 +22.72 +0.2842

Because the held -out test set is class -balanced (~139 points per class), the overall accuracy is
essentially the mean of the five per -class accuracies; the modest aggregate gain therefore averages
one large improvement against three near -flat classes ( Table 4 ). The optical -to-fusion gain is

111

concentrated in built -up (Producer’s Accuracy +6.7 percentage points in 2022) and cropland (+7.1
points in 2024) the transition classes central to GEOINT monitoring because SAR backscatter
(double-bounce from vertical structures, textural separation of cropland from bare soil) adds
information exactly where the optical signal is ambiguous, whereas vegetation and water are already
near-ceiling with the NDVI and NDWI i ndices. By 2024 built-up is spectrally mature and its fusion
gain disappears, shifting SAR’s marginal value to cropland and vegetation.
Table 4.
Per-class change in Producer’s and User’s Accuracy from optical-only to fusion (percentage points;
negative = fusion lower).
Class ΔPA 2022 ΔUA 2022 ΔPA 2024 ΔUA 2024
Built-up +6.7 +1.6 −1.2 +1.6
Bare land +0.0 +3.5 −0.7 −0.2
Vegetation +0.8 +5.6 +0.8 +5.4
Cropland +3.7 +2.6 +7.1 +2.8
Water +0.7 −0.6 +0.6 −1.5
3.2. Per-Class Performance Analysis (Confusion Matrix)
The confusion matrix for the RF Fusion 2022 model (Table 5) reveals strong diagonal dominance
across all five classes. Vegetation achieved the highest classification accuracy (122/130 = 93.8%
Producer's Accuracy), followed closely by water bodies (140/150 = 93.3%) and built -up (135/149 =
90.6%). Cropland reached 84.3% (113/134), while bare land was the most challenging class at 81.5%
(106/130), owing to its spectral overlap with both built-up and vegetation during active construction.
The uniformly high per-class accuracies all exceeding 81% indicate that the SAR-optical fusion stack
discriminates the five land-cover types reliably even amid the landscape heterogeneity of KIPP IKN.
Table 5.
Confusion matrix for RF Fusion 2022 classification (test set, n=693).
Actual /
Predicted
Built-up Bare Vegetation Cropland Water
Built-up
(n=149)
135 11 3 0 0
Bare land
(n=130)
17 106 2 5 0
Vegetation
(n=130)
4 0 122 2 2
Cropland
(n=134)
4 9 2 113 6
Water
(n=150)
0 5 0 5 140
The full classification maps reveal the spatial distribution of land cover transformation across the
KIPP IKN study area for both periods ( Fig. 3). The 2024 map (panel b) demonstrates substantially
expanded built-up coverage (red), particularly concentrated in the central area corresponding to the
core development zone of IKN.
3.3. Multi-temporal Land Cover Distribution
Application of the trained RF Fusion classifiers to the full 17,599 -hectare study area revealed
dramatic land cover transformation between 2022 and 2024 ( Table 6 ). Built -up area exhibited a
striking increase, expanding from 785.35 hectares in 2022 to 1,258.83 hectares in 2024 an absolute
increase of 473.48 hectares representing a 60.3% expansion within two years. Even more pronounced
in relative terms, bare land expanded from 151.32 to 717.06 hectares (+373.9%), reflecting extensive
land clearing that constitutes a pre-construction pipeline.

Fajar Sidik SUGANDA and Asep Adang SUPRIYADI / CLOUD-BASED MULTI-TEMPORAL … 112

Fig. 3. Multi-temporal land cover classification maps showing (a) 2022 baseline and (b) 2024 current period.
Cropland nearly doubled (+99.0%, 630.48 → 1,254.96 ha), likely reflecting agricultural
intensification by displaced communities, while vegetation declined by 1,660.04 hectares (10.9%
reduction) as forest was progressively converted. Water bodies remained e ssentially stable (−0.4%),
consistent with the absence of major impoundment within the core zone during this period.
Table 6.
Multi-temporal land cover distribution (hectares).
Class 2022 (ha) 2024 (ha) Change (ha) Change (%)
Built-up 785.35 1,258.83 +473.48 +60.3%
Bare land 151.32 717.06 +565.73 +373.9%
Vegetation 15,175.45 13,515.40 -1,660.04 -10.9%
Cropland 630.48 1,254.96 +624.48 +99.0%
Water 856.65 852.99 -3.65 -0.4%
3.4. Transition Matrix Analysis
The complete 5×5 transition matrix ( Table 7 ) quantified pixel -wise pathways of landscape
transformation between 2022 and 2024. The most prominent off -diagonal transition was vegetation-
to-built-up at 899.64 hectares, equivalent to 71.5% of the total 2024 built -up area and confirming
direct forest-to-construction conversion as the dominant driver of built-up growth.
Table 7.
Transition matrix between 2022 and 2024 land cover classes (hectares).
From / To Built-up Bare Vegetation Cropland Water Total 2022
Built-up 276.57 63.48 369.55 75.74 0.01 785.35
Bare land 39.28 63.54 16.54 18.10 13.86 151.32
Vegetation 899.64 532.33 12,876.30 860.96 6.22 15,175.45
Cropland 43.35 45.89 233.84 286.64 20.76 630.48
Water 0.00 11.80 19.17 13.52 812.15 856.65
Total 2024 1,258.83 717.06 13,515.40 1,254.96 852.99 17,599.25

113

The vegetation-to-bare-land pathway (532.33 ha) represents a secondary clearing route, while
bare-to-built-up transitions reached 39.28 hectares and cropland contributed 43.35 hectares to new
built-up area. The diagonal entry for vegetation -to-vegetation ( 12,876.30 ha) confirms that the
majority of forested cover remained intact despite concentrated transformation in the central KIPP
zone.
Fig. 4 visualizes the transition matrix as a heatmap with logarithmic scaling to accommodate the
wide range of transition magnitudes (from ~0.01 to 12,876 hectares). The diagonal cells (stable land
covers) and the highlighted Vegetation→Built -up cell (899.64 ha, blue outline) represent the two
most prominent patterns of landscape behavior during the 2022-2024 period.

Fig. 4. Heatmap visualization of the 5×5 land cover transition matrix between
2022 and 2024 (hectares, log-scale).
3.4.1. Verification of Detected Changes
To provide reference -independent evidence for the detected transitions, a stratified random
sample of 150 points (50 each on vegetation -to-built-up, built-up-to-vegetation and vegetation -to-
bare-land) was cross -checked against Dynamic World a 10 m deep -learning land -cover product
independent of both ESA WorldCover and the Random Forest classifier using annual modal labels
for 2022 and 2024, and inspected on Sentinel -2 true-colour composites (Fig. 5). All 100 vegetation-
origin points were vegetation in 2022 in Dynamic World, and 72% had left tree canopy by 2024; for
the vegetation-to-built-up class specifically, 48% were labelled built -up or bare in Dynamic World
for 2024, the remainder retaining a c anopy label at mixed construction -edge pixels where the true -
colour imagery nonetheless shows clearing (Fig. 5).
The built-up-to-vegetation anomaly was independently confirmed as commission error: none of
its 50 points was built-up in Dynamic World in either year, so this class (369.55 ha) is a classification
artefact rather than genuine revegetation and is excluded from the change total; the corrected area of
significant change is 1,560.49 ha (8.9% of the ROI; Table 8). Google Open Buildings Temporal was
also examined but does not yet register the IKN construction (building presence < 0.15 at its mid -
2023 epoch) and was not used.

Fajar Sidik SUGANDA and Asep Adang SUPRIYADI / CLOUD-BASED MULTI-TEMPORAL … 114

Fig. 5. Reference-independent verification (Section 3.4.1): Sentinel-2 true-
colour composites for 2022 (left) and 2024 (right) at six representative
vegetation-to-built-up sample points, each showing forest in 2022 and
cleared or built ground in 2024.

115

3.5. Critical Land Cover Change Categories for GEOINT
Translation of the transition matrix into geospatial intelligence categories produced five critical
change types plus one class (built -up-to-vegetation) excluded as commission error ( Table 8). After
that exclusion, confirmed by an independent Dynamic World check (Section 3.4.1), the total area of
significant change is 1,560.49 hectares (8.9% of the ROI). The largest critical category was
vegetation-to-built-up (899.64 ha, 5.1% of ROI) represe nting direct deforestation for construction,
followed by vegetation -to-bare-land (532.33 ha, 3.0%) representing pre -construction land clearing.
The combined critical built -up acquisition pathways (Veg→Built + Bare→Built + Crop→Built =
982.27 ha) directly s upport GEOINT operations including spatial planning verification and security
perimeter monitoring.
Table 8.
Critical land cover change categories between 2022–2024.
Category Description Area (ha) % ROI GEOINT Significance
Veg → Built Direct deforestation 899.64 5.1% Critical (rapid
impact)
Bare → Built Construction
completion
39.28 0.2% High (planned)
Crop → Built Agricultural
conversion
43.35 0.2% Moderate
Veg → Bare Land clearing
pipeline
532.33 3.0% Future
construction signal
Crop → Bare Cropland clearing 45.89 0.3% Low

 Built → Veg 
  (excluded) 

 Commission error,   

confirmed by Dynamic
World (Sec. 3.4.1)

  369.55 

    2.1% 

   not counted 

Total significant
change
Sum of the five
categories above

           1,560.49 

           8.9% 

            – 

The spatial distribution of these critical change categories is visualized in Fig. 6 , revealing
distinct geographic patterns of transformation across KIPP IKN. The dark red areas (Veg→Built,
critical) cluster predominantly in the central portion of the study area, corresponding to the core IKN
development zone, while the orange Veg→Bare p attern (pre-construction pipeline) extends radially
outward, indicating the spatial trajectory of future expansion.

Fig. 6. GEOINT change category map of KIPP IKN 2022→2024, identifying six mapped transition types; with
the built-up-to-vegetation class excluded as commission error (Section 3.4.1), total significant change is
1,560.49 ha (8.9% of ROI).

Fajar Sidik SUGANDA and Asep Adang SUPRIYADI / CLOUD-BASED MULTI-TEMPORAL … 116

4. DISCUSSION
   4.1. SAR-Optical Fusion Performance in Tropical Cloud-Affected Regions
   The RF Fusion approach achieved 88.89% OA and 0.8609 Kappa for 2022, outperforming SAR-
   only (63.20%) and optical -only (86.44%). The 25.69 -point gain over SAR -only confirms that
   complementary sensor information enhances classification under difficult atmosp heric conditions,
   while the 2.45 -point gain over optical -only reflects the marginal but consistent value of SAR
   integration in cloud-prone tropics. This aligns with prior tropical fusion studies: (Monsalve-Tellez et
   al., 2022) reported 82.14% OA in tropical Colombia (7.85 points over optical-only, 11.66 over SAR-
   only), a hierarchy remarkably similar to ours, supporting the generalizability of fusion beyond
   temperate settings. Recent Sentinel -1/2 fusion in Kuala Selangor, Malaysia -an analogous tropical
   Southeast Asian setting-similarly reported substantial benefits despite persistent cloud cover reducing
   optical continuity to fewer than 30 annual scenes (Jia et al., 2025). On the class -balanced test set the
   overall accuracy is the mean of the five per -class accuracies, so the small aggregate gain reflects
   averaging one large improvement against near -flat classes: the fusion benefit falls on built -up (+6.7
   pp Producer’s Accuracy, 2022) and cropland (+7.1 pp, 2024) the transition classes central to GEOINT
   monitoring (Table 4) while vegetation and water are already near-ceiling with the optical indices.
   Importantly, the 2024 fusion result (85.12% OA, kappa=0.8127) maintained the performance
   hierarchy despite the transforming landscape, indicating that fusion robustness extends to non -
   stationary conditions-a critical property for operational monitoring dur ing active construction. The
   3.77-point decline from 2022 to 2024 primarily reflects temporal label -data mismatch (2021 labels
   versus 2022 -2024 test data) rather than methodological weakness. Transformer -based fusion
   architectures could further improve accuracy by adaptively weighting sensors by local data reliability
   (Liu et al., 2024).
   4.2. Built-up Expansion as GEOINT Indicator for IKN Development
   The 60.3% expansion of built -up area (785.35 -> 1,258.83 ha) far exceeds typical urbanization
   rates for established Indonesian cities, placing KIPP IKN among the most rapidly transforming
   strategic zones globally (Purnama & Karondia, 2026). Critically, 71. 5% of the 2024 built -up total
   originated directly from vegetation (899.64 ha Veg ->Built), indicating an aggressive parallel -track
   approach where forest clearing and foundation work occur within compressed timeframes rather than
   the conventional clear -then-build sequence -a distinctive GEOINT signature of megaproject
   execution, with implications for environmental-impact and biodiversity-loss accounting. This extends
   the IKN-area characterization of (Habibie et al., 2025), who reported a forest-to-grassland shift across
   East Kalimantan; our short -interval, high -resolution approach reveals construction dynamics at a
   granularity not previously documented for KIPP. The vegetation -to-bare-land transition (532.33 ha),
   now secondary to the dominant Veg ->Built pathway , remains a predictive signal for continued
   expansion through 2025-2026, supporting spatial planning and security-perimeter monitoring before
   construction completion. An independent cross-check against Dynamic World (150 stratified points)
   confirms that al l sampled vegetation -to-built-up points were vegetated in 2022 and that 72% of
   vegetation-origin points had left tree canopy by 2024 (Section 3.4.1); the built -up-to-vegetation
   anomaly is confirmed as commission error and is excluded from the corrected significant-change total
   (1,560.49 ha, 8.9% of the ROI).
   4.3. Limitations and Future Work
   Several limitations warrant acknowledgment. First, training and reference labels derive from
   ESA WorldCover 2021 pseudo-labels rather than field-validated ground truth. For 2022 the one-year
   offset is minor; for 2024 the three -year offset adds label noise that both degrades training for the
   rapidly changing built -up and bare -land classes and shifts the accuracy reference away from 2024
   ground truth (Section 2.7). Manually annotated or higher -resolution reference data would improve
   reliability. Second, the b uilt-up-to-vegetation anomaly (369.55 ha) points to classification errors
   during construction transitions, where partially cleared areas, foundations, and exposed soil within
   single 10-m pixels create spectral ambiguity (Purnama & Karondia, 2026); an indep endent Dynamic

117

World check (Section 3.4.1) found none of a 50-point sample of this class to be built-up in either year,
confirming it as commission error. Multi -temporal fusion across phenological cycles is promising:
(H. Li et al., 2026) showed that integrating spectral , structural, and phenological attributes from
Sentinel-1/2 time series improves accuracy in subtropical cloud -prone regions. Third, future work
should incorporate ground -truthing of change hotspots, evaluate Sentinel -3 thermal data for urban
heat-island assessment, and develop automated comparison against the official KIPP spatial plan.
5. CONCLUSIONS
This study developed and validated a cloud-based SAR-optical fusion framework using Random
Forest for GEOINT monitoring of the KIPP of IKN Nusantara. Implemented in Google Earth Engine,
it integrated 70 Sentinel -1 and 52 Sentinel -2 images for 2022 and 61 a nd 81 for 2024, across the
17,599-ha cloud-prone study area.
The RF Fusion approach achieved 88.89% OA and 0.8609 Kappa for 2022 (2.45 points over
optical-only, significant at McNemar p=0.028; 25.69 points over SAR -only) and maintained the
hierarchy for 2024 (85.12% OA, 0.8127 Kappa). Change detection revealed built -up expansion of
60.3% (785.35 -> 1,258.83 ha), of which 71.5% came directly from vegetation (899.64 ha Veg -

Built)-a distinctive GEOINT signature of accelerated megaproject construction. The accompanying
surge in bare land (+373.9%) and the 532.33-ha vegetation-to-bare-land transition together provide a
predictive indicator for continued expansion through 2025-2026.
The novelty lies in integrating cloud -based SAR -optical fusion with short -interval (2 -year)
change detection calibrated for tropical megaproject monitoring, providing operational GEOINT
capability for strategic infrastructure. Its reproducibility in Google Earth Engine supports scaling to
the broader 256,000 -ha IKN extent. Future work should add manually validated samples, deep -
learning architectures, and automated comparison against the official KIPP spatial plan.
ACKNOWLEDGEMENTS
The authors thank the providers of the openly available datasets used in this study (Sentinel -1,
Sentinel-2, and ESA WorldCover) and Google Earth Engine for cloud processing. The authors declare
no conflict of interest. This research received no external funding. Reproducible Google Earth Engine
code is available from the corresponding author on request.

REFERENCES

Adi, W.B., Anshar, A. & Muladi, A. (2025) Analisis Persebaran Zona Potensi Kawasan Rawan
Kebakaran Hutan dan Lahan di Kecamatan Sambelia Menggunakan Metode Penginderaan Jauh
dan Analisis Spasial. Jurnal Geosains Dan Remote Sensing, 6(1), 15 –34.
https://doi.org/10.23960/jgrs.ft.unila.382
Al-Dousari, A.E., Mishra, A. & Singh, S. (2023) Land use land cover change detection and urban
sprawl prediction for Kuwait metropolitan region, using multi -layer perceptron neural networks
(MLPNN). Egyptian Journal of Remote Sensing and Space Science, 26( 2), 381 –392.
https://doi.org/10.1016/j.ejrs.2023.05.003
Aos, A.N.A. & Putri, N. (2023) Dinamika Vegetasi dan Suhu Permukaan Lahan Berbasis Remote
Sensing di Waduk Jatigede Provinsi Jawa Barat: Studi Pendahuluan. Jurnal Geosains Dan
Remote Sensing, 4(2), 67–76. https://doi.org/10.23960/jgrs.ft.unila.112
Carneiro, G.A., Svoboda, J., Cunha, A., Sousa, J.J. & Štych, P. (2025) Efficient 3D convolutional
neural networks for Sentinel-2 land cover classification with limited ground truth data. European
Journal of Remote Sensing, 58(1). https://doi.org/10.1080/22797254.2025.2512531

Fajar Sidik SUGANDA and Asep Adang SUPRIYADI / CLOUD-BASED MULTI-TEMPORAL … 118

Cholidah Nina Nila Ziyana & Irawan Listyo Yudha. (2024) Pemanfaatan Penginderaan Jauh Untuk
Pemetaan Alih Fungsi Lahan Sawah di Kabupaten Nganjuk. Jurnal Geosains Dan Remote
Sensing. https://doi.org/10.23960/jgrs.ft.unila.203
Dagne, S.S., Hirpha, H.H., Tekoye, A.T., Dessie, Y.B. & Endeshaw, A.A. (2023) Fusion of sentinel-
1 SAR and sentinel-2 MSI data for accurate Urban land use -land cover classification in Gondar
City, Ethiopia. Environmental Systems Research, 12(1). https://do i.org/10.1186/s40068-023-
00324-5
Djati, W.H. & Gunawan, D.S. (2025) Utilization Of Geospatial Intelligence (Geoint) In The
Intelligence Cycle To Support The Role Of State Intelligence In Indonesia. Security Intelligence
Terrorism Journal (SITJ), 2(3), 290–296. https://doi.org/10.70710/sitj.v2i3.65
Feng, X., Li, P. & Cheng, T. (2021) Detection of Urban Built -Up Area Change from Sentinel -2
Images Using Multiband Temporal Texture and One -Class Random Forest. IEEE Journal of
Selected Topics in Applied Earth Observations and Remote Sensing, 14, 6974 –6986.
https://doi.org/10.1109/jstars.2021.3092064
Fitri, R., Fauzi, R., Seanders, O. & Danniswari, D. (2025) Land use changes and residential area
expansion in South Tangerang City, Indonesia. Southeast Asia: A Multidisciplinary Journal,
25(2), 86–94. https://doi.org/10.1108/seamj-01-2023-0003
Gülci, S., Wing, M. & Akay, A.E. (2025) Land Use and Land Cover (LULC) Mapping Accuracy
Using Single -Date Sentinel -2 MSI Imagery with Random Forest and Classification and
Regression Tree Classifiers. Geomatics, 5(3). https://doi.org/10.3390/geomatics503002 9
Habibie, M.I., Nurda, N., Fernando, D., Arifandri, R., Putra, P.K., Prayogi, H. & Sencaki, D.B. (2025)
Integrating Sentinel -2 and ESA world cover for effective land use and land cover assessment
using machine learning. Advances in Space Research, 76(9), 49 25–4958.
https://doi.org/10.1016/j.asr.2025.07.083
Irfan, A., Li, Y., E, X. & Sun, G. (2025) Land Use and Land Cover Classification with Deep Learning-
Based Fusion of SAR and Optical Data. Remote Sensing, 17(7).
https://doi.org/10.3390/rs17071298
Jagannathan, J., Vadivel, M.T. & Divya, C. (2025) Land use classification using multi-year Sentinel-
2 images with deep learning ensemble network. Scientific Reports, 15(1).
https://doi.org/10.1038/s41598-025-12512-7
Jia, P., Mohd Shafri, H.Z., Yu, S., Zheng, Z., You, S. & Mohamed Shariff, A.R. (2025) Radar-optical
fusion of Sentinel-1/2 for high-resolution NDVI reconstruction and landscape-driven carbon flux
assessment in Kuala Selangor, Malaysia (2020 –2024). Internat ional Journal of Applied Earth
Observation and Geoinformation, 145. https://doi.org/10.1016/j.jag.2025.104966
Jocea, A.F., Porumb, L., Necula, L. & Raducanu, D. (2025) Sentinel -2 Land Cover Classification:
State-of-the-Art Methods and the Reality of Operational Deployment
https://doi.org/10.3390/su172210324 A Systematic Review. In Sustainability (Switzerland) (Vol.
17, Number 22). Multidisciplinary Digital Publishing Institute (MDPI).
Kasahun, M. & Legesse, A. (2024) Machine learning for urban land use/ cover mapping: Comparison
of artificial neural network, random forest and support vector machine, a case study of Dilla town.
Heliyon, 10(20). https://doi.org/10.1016/j.heliyon.2024.e39146
Latif, R.M.A., Arshad, A., He, J., Ahmad, T. & Ahmed, A. (2025) Geospatial modeling and
forecasting of urban land use change using Google Earth Engine and machine learning. PLOS
ONE, 20(12 December). https://doi.org/10.1371/journal.pone.0338920
Lemenkova, P. (2025) Machine Learning Algorithms of Remote Sensing Data Processing for
Mapping Changes in Land Cover Types over Central Apennines, Italy. Journal of Imaging, 11(5).
https://doi.org/10.3390/jimaging11050153
Li, H., Luo, C., Kang, X., Luan, H. & Li, L. (2026) Multi-Temporal Fusion of Sentinel-1 and Sentinel-
2 Data for High-Accuracy Tree Species Identification in Subtropical Regions. Remote Sensing,
18(4). https://doi.org/10.3390/rs18040592

119

Li, Y. & Xiao, X. (2025) Deep Learning -Based Fusion of Optical, Radar, and LiDAR Data for
Advancing Land Monitoring. In Sensors (Vol. 25, Number 16). Multidisciplinary Digital
Publishing Institute (MDPI). https://doi.org/10.3390/s25164991
Liu, R., Ling, J. & Zhang, H. (2024) SoftFormer: SAR -optical fusion transformer for urban land use
and land cover classification. ISPRS Journal of Photogrammetry and Remote Sensing, 218, 277–
293. https://doi.org/10.1016/j.isprsjprs.2024.09.012
Monsalve‐tellez, J.M., Torres‐león, J.L. & Garcés‐gómez, Y.A. (2022) Evaluation of SAR and
Optical Image Fusion Methods in Oil Palm Crop Cover Classification Using the Random Forest
Algorithm. Agriculture (Switzerland), 12(7). https://doi.org/10.3390/agric ulture12070955
Nigar, A., Li, Y., Jat Baloch, M.Y., Alrefaei, A.F. & Almutairi, M.H. (2024) Comparison of machine
and deep learning algorithms using Google Earth Engine and Python for land classifications.
Frontiers in Environmental Science, 12. https://doi.org/10.3389/fenvs.2024.1378443
Pande, C.B., Srivastava, A., Moharir, K.N., Radwan, N., Mohd Sidek, L., Alshehri, F., Pal, S.C.,
Tolche, A.D. & Zhran, M. (2024) Characterizing land use/land cover change dynamics by an
enhanced random forest machine learning model: a Google Earth Engine i mplementation.
Environmental Sciences Europe, 36(1). https://doi.org/10.1186/s12302 -024-00901-0
Purnama, S.M. & Karondia, L.A. (2026) Utilizing Remote Sensing for Land Change and Night Lights
in Urbanization: Correlation between Built -Up Area Expansion in Berau City and Changes in
Climate Parameters. Geomatics and Environmental Engineering, 20(1), 93 –115.
https://doi.org/10.7494/geom.2026.20.1.93
Rangel, A., Terven, J., Córdova -Esparza, D.M., Romero -González, J.A., Ramírez -Pedraza, A.,
Chávez-Urbiola, E.A., Willars-Rodríguez, F.J. & Alfonso-Francia, G. (2025) Tracking U.S. Land
Cover Changes: A Dataset of Sentinel-2 Imagery and Dynamic World Labels (2016–2024). Data,
10(5). https://doi.org/10.3390/data10050067
Septianto Aldiansyah, Randi Adrian Saputra, Khalil Abdul Wahid, Ilyas Madani & Duwi Setiyo.
(2024) Rapid Flood Inundation Mapping Using Multi -Temporal Sentinel-1 SAR: An Example
from Kendari City. Jurnal Geosains Dan Remote Sensing.
https://doi.org/10.23960/jgrs.ft.unila.205
Shandu, I.D., Xulu, S. & Gebreslasie, M. (2026) Enhancing land cover classification in the
heterogeneous landscape by integrating auxiliary data with Sentinel-2 imagery using the random
forest algorithm. Frontiers in Remote Sensing, 6. https://doi.org/10.3 389/frsen.2025.1697897
Syaban, A.S.N. & Appiah -Opoku, S. (2024) Unveiling the Complexities of Land Use Transition in
Indonesia’s New Capital City IKN Nusantara: A Multidimensional Conflict Analysis. Land,
13(5). https://doi.org/10.3390/land13050606
Yusfan, M.A., Supriyadi, A.A., Martha, S., Agus, R., Gultom, G. & Kartika Sakti, S. (2021)
Geospatial Intelligence (GEOINT) As A National Defense Strategy for Information Overload in
Indonesia. International Journal of Engineering, 3(2). https://doi.org/10 .46923/ijets.v3i2.129
Yusuf Samad, M., Persadha, P., Ayu Permatasari, D., Adang Supriyadi, A., Prihanto, Y. & Dhae Ue,
E. (2023) Analisis Intelijen Terhadap Manfaat Penginderaan Jauh dalam Memahami Aksi
Kelompok Separatis di Indonesia Guna Mendukung Sistem Pertahanan Negara Int elligence
Analysis of the Benefits of Remote Sensing in Understanding the Actions of Separatist Groups
in Indonesia to Support the National Defense System Article info. 1, 55 –74.