Enhancing LULC Classification Using Satellite Image Embeddings and Machine Learning in A Data-scarce Region: A Case Study of Timor-Leste (Ongoing)
Land Use/Land Cover (LULC) classification plays a crucial role in understanding spatial patterns, environmental change, and climate change mitigation. Accurate LULC information supports sustainable urban planning, ecosystem management, and environmental conservation. In addition, reliable LULC data provides a foundation for various environmental applications, including carbon assessment and land monitoring.
Over the past decades, remote sensing combined with machine learning techniques has become a widely used approach for LULC classification. However, many existing studies rely on costly field inventory data, LiDAR, or proprietary datasets, which are often unavailable in developing or data-scarce regions (Pandey et al., 2021; Huang et al., 2025). As a result, there is a need for cost-effective and scalable approaches that utilize open-access datasets while maintaining reliable classification performance.
Timor-Leste is an example of data limited context (Cho and Lim, 2023). It is a small country located in Southeast Asia with a total area of approximately 14,954 km² and has experienced significant forest loss in recent decades due to repeated burning and land clearing for agricultural activities (Democratic Republic of Timor-Leste, 2023). As a member of the UNFCCC, Timor-Leste is required to develop strategies for climate mitigation, including REDD+ initiatives and a national forest monitoring system. These efforts require accurate and up-to-date spatial information on LULC.
Therefore, this study aims to develop a cost-effective and scalable framework for LULC classification using open-access satellite imagery and machine learning. The approach supports land monitoring and spatial planning in data-limited regions without relying on costly field surveys or proprietary datasets. Expected outputs include a 10 m resolution LULC map of Timor-Leste with seven land-cover classes and an anticipated overall accuracy of approximately 85%, as well as a transferable classification framework applicable to other data-scarce regions.
The research objectives of this study are:
- To develop and evaluate LULC classification models using open-access satellite embedding dataset from Google Earth Engine.
- To generate high-resolution LULC maps suitable for land monitoring and environmental applications in Timor-Leste.
- To validate LULC classification accuracy using high-resolution Google Earth imagery.
This study focuses on Timor-Leste, a young Southeast Asian country characterized by diverse landscapes and significant land use dynamics (Democratic Republic of Timor-Leste, 2023). The region has experienced considerable deforestation and land-use change in recent decades, making it a relevant case for LULC classification assessments in data-limited contexts. To capture the variability of land cover types, multiple areas of interest (AOIs) representing different ecosystems and terrain conditions will be selected across the country for model training and validation.
The scripts that download Google Satellite Embedding and ESA WorldCover dataset are at here, while Google Earth imagery requires to acquire manually from Google Earth pro and georeferencing in GIS software.
A labelled dataset was created from high-resolution Google Earth imagery using multiple Areas of Interest (AOIs) representing diverse land-cover types, terrain conditions, and ecosystems across Timor-Leste. To reduce labelling bias, two GIS and remote sensing interpreters independently annotated the imagery. The Forest/Non-Forest (FNF) model used two classes (forest and non-forest), while the LULC model employed six land-cover categories (water, wood, crops, built area, bare ground, and grassland).
Both FNF and LULC labels are manually labelled in GIMP and exported as PNG files and then georeferenced and merged through fnf_label and lulc_label as TIFF files. The labelled data were resampled from approximately 1 m to 10 m resolution to support model training and accuracy assessment using align_dataset.
Google Satellite Embedding (GSE) provides 64-dimensional, analysis-ready feature vectors at 10 m resolution derived from multi-source satellite imagery. These embeddings capture complex spatial and temporal patterns and were used as predictor variables for both FNF and LULC classification. The dataset was obtained from Google Earth Engine and temporally aligned with ESA WorldCover data.
The ESA WorldCover dataset provides global land-cover information at 10 m resolution derived from Sentinel-1 and Sentinel-2 imagery. It was selected as the primary source of training labels because it offers standardized, globally consistent land-cover classifications that align with the spatial resolution of the GSE dataset. In this study, the 2021 ESA WorldCover dataset was acquired through Google Earth Engine and reclassified into the target LULC categories. Tree cover and shrubland were merged into a single wood class, while uncommon or irrelevant classes (e.g., mangroves, wetlands, snow/ice, and moss/lichen) were excluded using recalssify_esa. The dataset was then aligned with the GSE data to support model training and prediction using align_dataset.
A supervised machine learning framework was developed to generate FNF and LULC maps from open-access satellite imagery. Both models employed Random Forest classifiers, with the FNF model trained from manually interpreted Google Earth imagery followed by the postprocessing to match the forest criteria (continuous area > 0.5 ha) and the LULC model trained using ESA WorldCover data. To reduce confusion between forest and shrubland, the LULC model classified six land-cover classes, combining both into a single wood category, which was subsequently refined using the FNF classification by combination. Model performance was evaluated using independent validation samples, and the final seven-class LULC map was assessed through stratified random sampling and visual interpretation of 30 m × 30 m plots.
Both models achieved strong performance, with overall accuracies of 86% for the FNF model and 84% for the LULC model. Most classes were classified accurately; however, water and bare ground showed lower precision and recall, likely due to seasonal variations between Google Earth imagery and the annual GSE dataset. In contrast, the wood and crops classes achieved consistently high accuracy. The built area class was more difficult to classify because small, scattered settlements are often generalized at 10 m resolution. Grassland was also frequently confused with wood and crops due to similar spectral characteristics.
▼ FNF model classification report
| precision | recall | f1-score | support | |
|---|---|---|---|---|
| Non-forest | 0.8517 | 0.8618 | 0.8567 | 28798 |
| Forest | 0.8710 | 0.8615 | 0.8663 | 31202 |
| Overall accuracy | 0.8617 | 60000 |
▼ FNF model confusion metrics
| True \ Pred | Non-forest | Forest |
|---|---|---|
| Non-forest | 24817 | 3981 |
| Forest | 4320 | 26882 |
▼ LULC model classification report
| precision | recall | f1-score | support | |
|---|---|---|---|---|
| Water | 0.74 | 0.99 | 0.84 | 71222 |
| Wood | 0.92 | 0.93 | 0.93 | 410356 |
| Crops | 0.89 | 0.92 | 0.91 | 127719 |
| Built area | 0.25 | 0.24 | 0.25 | 6507 |
| Bare ground | 0.51 | 0.37 | 0.43 | 54453 |
| Grassland | 0.66 | 0.54 | 0.59 | 98862 |
| Overall accuracy | 0.84 | 769119 |
▼ LULC model confusion metrics
| True \ Pred | Water | Wood | Crops | Built area | Bare ground | Grassland |
|---|---|---|---|---|---|---|
| Water | 70483 | 42 | 12 | 0 | 682 | 3 |
| Wood | 946 | 382688 | 3042 | 1559 | 6128 | 15993 |
| Crops | 2 | 1861 | 117620 | 7 | 324 | 7905 |
| Built area | 250 | 957 | 78 | 1590 | 3384 | 248 |
| Bare ground | 23641 | 3461 | 1178 | 2059 | 20101 | 4013 |
| Grassland | 285 | 25507 | 9994 | 1126 | 8462 | 53488 |
The preliminary results indicate that the proposed machine learning framework can produce reliable LULC maps for Timor-Leste using open-access datasets. The mapping resolution was improved from 30 m to 10 m, while forest and shrubland were successfully separated through the integration of FNF and LULC classifications. The following figure presents the resulting 2023 LULC map of Timor-Leste, comprising seven classes: water, forest, crops, built area, bare ground, grassland, and shrubland.
Ongoing.
Ongoing.
Cho, W., & Lim, C.-H. (2023). Simplified and High Accessibility Approach for the Rapid Assessment of Deforestation in Developing Countries: A Case of Timor-Leste. Remote Sensing, 15(18), 4636. https://doi.org/10.3390/rs15184636
Democratic Republic of Timor-Leste. (2023). Timor-Leste forest reference emission level modified submission. United Nations Framework Convention on Climate Change. https://redd.unfccc.int/media/tl_forest_reference_emissions_level_modified_submission_-_final_version_print.pdf
Huang, D., Zhou, Z., Zhang, Z., Dai, Q., Lu, H., Li, Y., & Huang, Y. (2025). Land Use/Land Cover Remote Sensing Classification in Complex Subtropical Karst Environments: Challenges, Methodological Review, and Research Frontiers. Applied Sciences, 15(17), 9641. https://doi.org/10.3390/app15179641
Pandey, P. C., Koutsias, N., Petropoulos, G. P., Srivastava, P. K., & Ben Dor, E. (2021). Land use/land cover in view of earth observation: data sources, input dimensions, and classifiers—a review of the state of the art. Geocarto International, 36(9), 957–988. https://doi.org/10.1080/10106049.2019.1629647
Brown, C. F., Kazmierski, M. R., Pasquarella, V J., Rucklidge, W. J., Samsikova, M., Zhang, C., Shelhamer, E., Lahera, E., Wiles, O., Ilyushchenko, S., Gorelick, N., Zhang, L. L., Alj, S., Schechter, E., Askay, S., Guinan, O., Moore, R., Boukouvalas, A., & Kohli, P.(2025). AlphaEarth Foundations: An embedding field model for accurate and efficient global mapping from sparse label data. arXiv preprint arXiv.2507.22291. https://doi.org/10.48550/arXiv.2507.22291
Zanaga, D., Van De Kerchove, R., Daems, D., De Keersmaecker, W., Brockmann, C., Kirches, G., Wevers, J., Cartus, O., Santoro, M., Fritz, S., Lesiv, M., Herold, M., Tsendbazar, N.E., Xu, P., Ramoino, F., Arino, O., (2022). ESA WorldCover 10 m 2021 v200. https://doi.org/10.5281/zenodo.7254221