Andrii Bidochko

• Updated: August 22, 2026
• 7 min read

Remote Sensing and Machine Learning-Based Analysis of Land Use and Vegetation Change in Dhaka District, Bangladesh

Direct Answer

The paper presents a satellite‑based, machine‑learning workflow that maps how Dhaka’s urban footprint, vegetation, and water bodies have shifted between 2019 and 2024. By combining Sentinel‑2 and Landsat‑8 imagery with Decision Tree, K‑Nearest Neighbors, and Random Forest classifiers, the authors deliver a high‑resolution, reproducible land‑use change map that quantifies a 59.5 % surge in built‑up area and a double‑digit loss of green and aquatic spaces.

Background: Why This Problem Is Hard

Rapid urbanization in megacities like Dhaka creates a moving target for planners, environmental agencies, and disaster‑response teams. Traditional ground surveys are costly, time‑consuming, and quickly become outdated. Remote sensing offers a bird’s‑eye view, yet extracting reliable land‑cover information from raw satellite pixels remains challenging for three reasons:

• Spectral ambiguity: Built‑up surfaces, bare soil, and certain crop types share similar reflectance signatures, leading to misclassifications.
• Temporal inconsistency: Cloud cover, sensor drift, and seasonal phenology cause gaps and noise across multi‑year datasets.
• Scalability of labeling: Supervised learning demands high‑quality, georeferenced training points, which are scarce in rapidly changing urban peripheries.

Existing approaches often rely on a single classifier or handcrafted thresholds for indices like NDVI or NDBI. While these methods can produce coarse maps, they struggle to capture nuanced transitions—e.g., a park converted into a mixed‑use development—without extensive manual correction.

What the Researchers Propose

The authors introduce a three‑tiered analytical framework that fuses high‑resolution optical imagery with a suite of supervised machine‑learning models. The core idea is to let each model specialize in a different aspect of the classification problem and then let the best‑performing model (Random Forest, in this case) dictate the final land‑cover map. The framework consists of:

1. Data ingestion layer: Sentinel‑2 MSI (10 m) and Landsat‑8 OLI (30 m) scenes are pre‑processed in Google Earth Engine (GEE) to produce cloud‑free composites for each year.
2. Feature engineering layer: Spectral indices—NDVI (vegetation health), NDBI (built‑up intensity), and NDWI (water presence)—are computed and stacked with raw bands to enrich the feature space.
3. Modeling layer: Three classic classifiers (Decision Tree, K‑Nearest Neighbors, Random Forest) are trained on a curated set of geolocated ground truth points collected via field surveys and high‑resolution Google Earth imagery.
4. Validation layer: Confusion matrices and Cohen’s kappa statistics assess each model’s reliability, guiding the selection of the optimal classifier for final mapping.

How It Works in Practice

The operational workflow can be visualized as a linear pipeline, but each component interacts through shared data artifacts in GEE. Below is a step‑by‑step description:

1. Image Collection & Pre‑processing

• Retrieve Sentinel‑2 and Landsat‑8 collections for the target years (2019‑2024).
• Apply atmospheric correction, cloud masking, and temporal mosaicking to generate a single, cloud‑free raster per year.

2. Index Computation

• Calculate NDVI = (NIR − Red) / (NIR + Red) to highlight photosynthetic activity.
• Derive NDBI = (SWIR − NIR) / (SWIR + NIR) to accentuate built‑up surfaces.
• Generate NDWI = (Green − NIR) / (Green + NIR) to isolate water bodies.
• Stack all indices with the original spectral bands, forming a 12‑band feature cube.

3. Training Data Curation

• Field teams collect GPS points for five land‑cover classes: built‑up, vegetation, water, bare soil, and mixed‑use.
• Each point is cross‑validated with high‑resolution Google Earth imagery to ensure label fidelity.

4. Model Training & Selection

• Decision Tree learns simple rule‑based splits, offering interpretability but limited depth.
• K‑Nearest Neighbors leverages proximity in the feature space, sensitive to noise and class imbalance.
• Random Forest builds an ensemble of decorrelated trees, reducing over‑fitting and improving generalization.
• All models are trained on 70 % of the labeled points; the remaining 30 % serve as a hold‑out test set.

5. Accuracy Assessment & Final Mapping

• Confusion matrices reveal per‑class precision, recall, and overall accuracy.
• Kappa statistics quantify agreement beyond chance, with Random Forest achieving the highest score (κ ≈ 0.92).
• The Random Forest classification is exported as a GeoTIFF for downstream GIS analysis.

What sets this pipeline apart is its end‑to‑end automation within a cloud‑native environment (GEE) and its explicit comparison of three distinct classifiers on identical data. The result is a reproducible, transparent methodology that can be re‑run annually with minimal human intervention.

Evaluation & Results

The authors evaluated the framework across two dimensions: classification performance and land‑use change quantification.

Classification Performance

Random Forest outperformed the other two models across all metrics, especially in distinguishing built‑up from bare soil—a critical capability for urban expansion monitoring.

Land‑Use Change Quantification (2019‑2024)

• Built‑up area: Expanded by 59.5 % (≈ 1,240 km²), driven primarily by peripheral infill and informal settlements.
• Vegetation cover: Declined by 8.46 % (≈ 210 km²), reflecting loss of peri‑urban farms and parklands.
• Water bodies: Shrank by 7.77 % (≈ 95 km²), largely due to encroachment of construction on floodplains.
• Land conversion pathways: The dominant transition was vegetation → built‑up, followed by water → built‑up.

These figures are not merely statistical; they map directly onto policy‑relevant pressures such as increased flood risk, heat‑island formation, and reduced carbon sequestration capacity.

Why This Matters for AI Systems and Agents

From an AI engineering perspective, the study showcases a production‑ready template for geospatial intelligence pipelines:

• Modular data‑centric design: Each stage (ingestion, feature engineering, modeling, validation) can be swapped out or scaled independently, mirroring micro‑service architectures used in enterprise AI platforms.
• Model selection as a decision‑making agent: The comparative evaluation acts like an autonomous agent that selects the most trustworthy model based on predefined metrics, a pattern applicable to any multi‑model ensemble system.
• Real‑time monitoring potential: By automating the workflow in GEE and exposing the output via APIs, downstream agents (e.g., alert bots, urban‑growth simulators) can ingest fresh land‑use maps daily.
• Explainability through feature importance: Random Forest provides per‑feature importance scores, enabling agents to justify why a pixel was labeled as built‑up—a valuable asset for compliance‑focused AI deployments.

Practically, city planners could embed the generated maps into a Enterprise AI platform by UBOS to trigger automated zoning checks, while environmental NGOs might connect the output to a Telegram integration on UBOS for community‑level alerts about encroaching construction on wetlands.

What Comes Next

While the framework delivers actionable insights, several avenues remain open for refinement:

• Temporal granularity: Moving from annual composites to seasonal or monthly analyses would capture short‑term dynamics such as monsoon‑driven flooding.
• Deep learning alternatives: Convolutional Neural Networks (e.g., U‑Net) could improve boundary delineation, especially in heterogeneous urban‑rural interfaces.
• Multi‑sensor fusion: Incorporating SAR (Synthetic Aperture Radar) data would mitigate cloud‑cover limitations and add surface‑roughness cues.
• Feedback loops: Deploying a citizen‑science portal where residents validate or correct classifications could create a continuous learning loop for the models.

Future research might also explore how the land‑use trajectories derived here feed into climate‑impact simulators or smart‑city traffic optimization engines. By linking the spatial outputs to agent‑based models of human mobility, planners could anticipate infrastructure demand before the physical expansion occurs.

References

Remote Sensing and Machine Learning-Based Analysis of Land Use and Vegetation Change in Dhaka District, Bangladesh

Illustration

Figure 1: End‑to‑end remote‑sensing and machine‑learning pipeline used to map Dhaka’s land‑use change.

Andrii Bidochko

CTO UBOS

Andrii Bidochko is an AI entrepreneur and researcher focused on AI agents, reinforcement learning, and autonomous systems. He writes about the technologies shaping the future of machine intelligence, from frontier models and agent architectures to real-world AI applications.