TDWG2026
Home / Talks / FORUM
talk · Friday 25 September · 09:30–⁠09:45 · FORUM

Evaluating Four Machine Learning Using Google Earth Engine for Analyzing LULC Dynamics of Sandwip Island, Bangladesh

Md Abrar Al Foysol · AI Driven Monitoring and Species Identification

Recording time 1:24:54–1:32:56Open on Vimeo ↗

Summaries and key moments are machine-produced. Check the recording for the speaker’s exact words.

The short versionRandom Forest in Google Earth Engine gave the most accurate land cover classification of the highly dynamic Sandwip Island, revealing large increases in built-up area and mangroves alongside strong shoreline erosion and accretion.

Overview

What this was about

In a pre-recorded talk, Md Abrar Al Foysol presented a study of shoreline and land use/land cover change on Sandwip Island in Bangladesh's dynamic Ganges-Brahmaputra-Meghna delta. Shoreline change from 1995 to 2025 was analysed with DSAS in ArcGIS, and land cover from 2000 to 2024/25 was classified in Google Earth Engine with CART, SVM, Random Forest and Gradient Tree Boost, assessed against 140 ground-truth points. Random Forest performed best (93.57% overall accuracy), and results show strong erosion and accretion, a large increase in built-up area and mangroves, and a decrease in barren land.

Why it matters. It shows how freely available satellite imagery and cloud-based machine learning can be used to monitor rapidly changing, understudied coastal environments and to compare classifier performance for such tasks.

Key ideas

In the room

  • Sandwip Island is shaped by heavy sediment loads, continuous erosion and accretion, and pressures from population growth, infrastructure, climate change and aquaculture, yet is understudied.
  • Landsat and Sentinel-2 imagery were used; shorelines were extracted in ArcGIS and DSAS used to compute EPR, NSM and LRR.
  • Onshore shoreline EPR ranged from -160.5 to +203.3 m/yr and offshore from -245.2 to +572.9 m/yr.
  • CART showed more mixed pixels, SVM some overgeneralisation, Random Forest the cleanest map with GBT close behind; SVM was best for built-up areas.
  • Overall accuracy with 140 ground-truth points (20 per class, seven classes): RF 93.57%, GBT 92.14%, SVM 90%, CART 87.14%.
  • Built-up area rose from about 18.1 to 60 km2 (2000-2024), barren land fell from about 101.1 to 23.5 km2, and mangroves rose from about 17.7 to 27.4 km2.
  • Accretion occurred mainly in the east and southeast, erosion in the west.
Jump in

Notable moments

In their words

Transcript

Automatically generated captions can contain mistakes, especially in names and technical terms. Times are relative to the room recording.

Read the transcript ↓
Loading transcript…
Keep exploring

Related talks.

See the map ↗
talk · 22 Sept · 12:15–⁠12:30 · SAL B

From satellite embeddings to Darwin Core: an open, ensemble-AI workflow for Prosopis juliflora mapping and reusable training-dataset generation

Satellite embeddings plus hyperspectral signatures and ground validation can map invasive Prosopis at scale and generate reusable, attributable training data.

Pranav Jha
talk · 24 Sept · 11:50–⁠12:05 · SAL C

From Sensor to Species: AI-Assisted Segmentation, Deep Learning Workflows, and Persistent Biodiversity Records from Earth Observation Systems

Persistent organism identifiers for remotely sensed individual trees could turn repeat Earth-observation imagery into longitudinal biodiversity records, but only if uncertainty and model provenance travel with every record.

Kit Lewers
talk · 24 Sept · 09:03–⁠09:17 · SAL B

Beyond Data Integration: Ecological and Semantic Interoperability in Coastal Digital Twins

Coastal digital twins expose a long list of unresolved land-sea interoperability gaps, from vertical datums and shifting shorelines to missing cross-domain ontologies.

Laura Slaughter
Programme abstract

Evaluating Four Machine Learning Using Google Earth Engine for Analyzing LULC Dynamics of Sandwip Island, Bangladesh

Md Abrar Al Foysol

Sandwip Island is a highly dynamic coastal environment experiencing continuous geomorphological changes driven by erosion, accretion, and sediment deposition. This study analyses shoreline changes from 1995 to 2026 using the Digital Shoreline Analysis System (DSAS) in ArcGIS and evaluates Land Use Land Cover (LULC) changes from 2000 to 2024 using Google Earth Engine and machine learning algorithms, including Classification and Regression Tree, Gradient Boosting Trees, Random Forest, and Support Vector Machines. A total of 140 ground-truth points representing seven LULC classes were used to assess model accuracy. Among the models, RF demonstrated the highest overall accuracy (93.57%), producing clearer boundaries and better separation of major land-cover classes. GBT achieved an accuracy of 92.14%, showing reliable classification across most LULC types. SVM reached a moderate accuracy of 90%, although some overgeneralization was observed in spectrally similar classes. CART showed the lowest accuracy (87.14%), with increased mixed-pixel occurrence in dense and heterogeneous regions. The findings reveal a gradual increase in the total area of Sandwip Island due to sediment deposition, contributing to the expansion of sandy areas (61.628 km²), mangroves (27.412 km²), vegetation (58.618 km²), cropland (31.587 km²), and built-up areas (60.031 km²). These results highlight the significant influence of natural coastal processes on land-cover dynamics and demonstrate the effectiveness of remote sensing and machine learning techniques for monitoring dynamic coastal environments. The study further suggests that integrating sediment transport modeling and tidal-current analysis with machine learning approaches could improve the accuracy of future coastal change assessments.

Submitted to the conference programme before the talk, so the talk itself may differ.