Canopy Disturbance Detection in Tanzania’s Rufiji Delta: Remote Sensing for Mangrove Conservation

Tanzania’s Rufiji Delta, Africa’s largest tidal mangrove wetland at 54,500 hectares, faces mounting deforestation pressure on one of the continent’s most ecologically significant coastal ecosystems. Omdena partnered with the UK FCDO and Tanzania’s National Carbon Monitoring Centre to build an AI-powered canopy disturbance monitoring system using satellite and radar imagery. The project delivered a validated near-real-time disturbance indicator and a working application ready for country-wide deployment.

The Problem: Deforestation at Scale in the Rufiji Delta

The Rufiji River delta is Africa’s largest tidal mangrove wetland and one of the continent’s most ecologically significant coastal ecosystems. Its 54,500 hectares of protected forest serve as a natural carbon sink, a coastal buffer against tsunamis and storm surge, and a habitat supporting seven genera of mangrove trees.

Tanzania ranks fifth globally in deforestation rate, despite holding Africa’s third-largest forest cover. In the Rufiji Delta, deforestation and land degradation threaten both the ecosystem and the communities that depend on it. Protecting and restoring these mangrove habitats sits at the centre of Tanzania’s national climate and conservation strategy.

The challenge is scale. The delta is a vast, dynamic landscape of coastal swamps, narrow creeks, and mangrove thickets, impossible to monitor comprehensively through field visits alone. Tanzania’s National Carbon Monitoring Centre (NCMC) needed a way to detect canopy disturbance continuously, at the pixel level, and direct limited field crews to the locations that matter most. A community-driven monitoring tool, one that local teams could operate and act on, was central to the project vision.

The Partnership: Omdena, UK FCDO, and Tanzania’s NCMC

Omdena partnered with the UK Foreign Commonwealth and Development Office (FCDO) and local partners in Tanzania to address this challenge. Over six months, the collaboration pursued three goals: build an AI solution for mangrove canopy monitoring in the Rufiji Delta, develop the AI capability of local engineers, and create a system the NCMC could operate and scale independently.

Over the course of the engagement, the initiative enhanced the AI capabilities of hundreds of local engineers, contributed to job creation in the sector, supported local entrepreneurs and startups in developing AI solutions, and produced two working prototypes ready for the next phase of deployment.

Project Objectives

Both phases of the project shared the same high-level objectives:

• Build an AI system to detect and monitor deforestation in Tanzania’s carbon-rich mangroves in the Rufiji Delta.
• Provide Tanzania’s National Carbon Monitoring Centre with up-to-date analysis on canopy disturbance in the pilot area.
• Pilot the system to test its effectiveness and identify refinements.
• Deliver recommendations for scaling the system country-wide, including data management and infrastructure requirements.

Phase 1 — Detecting Canopy Change from Space

Innovation Challenge: 50 Engineers, One Question

Omdena structured the work in two phases. The first was an Innovation Challenge: 50 machine learning engineers working in parallel to explore approaches for detecting canopy change from satellite data and identifying the type of activity driving the observed changes.

The challenge quickly revealed the primary constraint: reliable, labelled deforestation data for the Rufiji Delta did not exist at the required scale. The imbalance between forested and deforested pixels made supervised deep learning impractical, and pre-trained models from comparable forest environments did not transfer reliably to this coastal ecosystem. Both the Innovation Challenge and the Talent Project ran on an eight-week sprint schedule.

The Approach: NDVI as the Foundation

The team pivoted to the Normalised Difference Vegetation Index (NDVI), a well-established measure of canopy greenness computable directly from multispectral satellite imagery. The Innovation Challenge found that NDVI derived from either the Sentinel 2 satellite constellation or PlanetScope data from the NICFI programme was a reliable indicator of canopy disturbance, and required significantly less labelled data than deep learning approaches. The approach used a regression model to fit the multispectral data to the NDVI index, a technically appropriate choice given the data constraints.

The output was a web application that allowed users to identify areas of canopy change and mark locations for conservation or restoration. The NDVI-based approach is scalable across Tanzania: expanding coverage country-wide would require extending the region of interest, sourcing cloud-free imagery from two time points, and redesigning the data architecture to manage multiple projects simultaneously.

Phase 2 — Near-Real-Time Disturbance Indicator

Talent Project: Building the MVP

The second phase was a Talent Project: a smaller, focused team conducting a deep dive on the Innovation Challenge’s most promising finding and building it into a minimum viable product for the NCMC.

The team’s objective was to deliver a new data layer, a near-real-time canopy disturbance indicator, that would issue weekly or biweekly warnings when a contiguous area of the canopy was disturbed. This layer was designed to help the NCMC prioritise the deployment of its limited field crews across a vast and difficult-to-access area.

Multi-Sensor Data Fusion

The Phase 2 team extended the data pipeline beyond NDVI to incorporate three complementary sources: optical imagery from Sentinel-2, Landsat 7 and Landsat 8, and Synthetic Aperture Radar (SAR) data from Sentinel-1, processed using Google Earth Engine as the cloud computing platform.

Cloud cover is a persistent problem in the Rufiji Delta. Fusing optical and SAR data directly addressed this: SAR signals penetrate cloud cover, increasing the number of valid observations and stabilising the time series in a region where optical-only approaches frequently produce data gaps.

How the Disturbance Indicator Works

The indicator analyses each 30×30 metre pixel, mapped against known forest area using Hanzen canopy data, by comparing two signals to measurements from the same location 36 months earlier:

• Canopy greenness — derived from optical imagery (Sentinel-2, Landsat). A drop in greenness signals the transition from vegetation to bare soil.
• Forest structure — derived from SAR radar backscatter (Sentinel-1). A reduction in backscatter signals the removal of canopy structure.

When a contiguous area of 10 or more pixels shows a combined change in both signals, a disturbance warning is issued. The indicator was validated against global datasets from the Global Forest Watch (GFW) and the Global Mangrove Watch (GMW), showing good agreement with their canopy disturbance records in the absence of local ground truth data.

What Was Delivered: The Demo Application

Both project phases produced working web-based applications. The demo application consolidates the NCMC’s canopy monitoring results into an interactive platform built for operational use.

The application displays detected canopy disturbances on an interactive map of the Rufiji Delta. Users can zoom into specific areas of interest, view event-level detail on each disturbance, and apply a date filter to analyse disturbance patterns across any selected time period.

Zooming into individual events shows the spatial extent of each disturbance and its severity, giving field teams the precise location and scale information needed to plan site visits.

The planned next step for the application is an automated alert system. When a canopy disturbance event exceeds a defined threshold, the system will notify the relevant authorities by email, enabling rapid response before further damage occurs.

What Building an AI-Powered Canopy Monitoring System Taught Us

Both phases encountered constraints that are common to AI projects in complex, data-scarce environments. Each one points directly to what the next phase of this work should prioritise.

Multi-sensor fusion solves the cloud cover problem

Relying on optical imagery alone produced frequent data gaps in a region with persistent cloud cover. Adding Sentinel 1 SAR data, which penetrates cloud, increased valid observations and stabilised the time series. Any future deployment in tropical or coastal environments should build multi-sensor fusion into the data pipeline from the start.

Ground truth is the highest-value next investment.

The absence of labelled, on-the-ground deforestation data was the primary constraint on model performance in both phases. Validation relied on global datasets (GFW, GMW), which are good proxies but not a substitute for local ground truth. The next project phase should prioritise a structured ground truth collection programme, coordinated with the NCMC and local field teams.

Domain knowledge should shape technology selection from day one.e

The Innovation Challenge initially targeted deep learning approaches. Data limitations required a pivot to NDVI, a decision that ultimately produced a more reliable and scalable result. Involving domain experts (ecologists, NCMC analysts) at the problem definition stage, not just the validation stage, would surface these constraints earlier and reduce rework.

End-user involvement determines whether the tool gets used

Both projects developed applications without close involvement from the NCMC’s operational teams. Understanding how field crews actually plan site visits, what trigger conditions are actionable, and what interface features matter in practice would have shaped the product in ways that increase adoption. Future phases should bring end-users into the design process from the first sprint.

What’s Next

Omdena is establishing a Local Chapter in Tanzania to democratise AI usage and build sustained local capability. The deforestation monitoring project in the Rufiji Delta is being relaunched through this chapter, with three priorities:

• Ground truth programme — systematic on-the-ground data collection in partnership with the NCMC to enable fully supervised model training and remove the primary constraint on detection accuracy.
• Automated alerts — completing the planned email notification system so that canopy disturbance events above a defined threshold trigger immediate authority response, without requiring manual dashboard monitoring.
• Country-wide scaling — extending the pipeline from the Rufiji Delta pilot to cover Tanzania’s full forest estate, using the GEE infrastructure already in place and the architecture validated in Phase 2.

Key Achievements

Validated Near-Real-Time Disturbance Indicator

A pixel-level canopy disturbance indicator fusing Sentinel-2, Landsat, and Sentinel-1 SAR data, validated against Global Forest Watch and Global Mangrove Watch datasets. Produces weekly or bi-weekly disturbance warnings across the 54,500-hectare pilot area.

Working Demo Application

An interactive web application displaying canopy disturbance events on a map of the Rufiji Delta, with date filtering and event-level detail. Built for operational use by Tanzania’s National Carbon Monitoring Centre.

Scalable Data Pipeline

A Google Earth Engine pipeline combining three satellite data sources — designed for extension to additional regions and replication at country scale across Tanzania’s forest estate.

Foundation for the Next Phase

Both projects produced working codebases, validated approaches, and a clear specification for the next build — including ground truth requirements, alert system design, and country-wide scaling architecture.

About the Project

This case study documents work delivered by Omdena, a global applied AI organisation with more than 300 projects across 60+ countries, in partnership with the UK Foreign Commonwealth and Development Office (FCDO) and Tanzania’s National Carbon Monitoring Centre, as part of Omdena’s climate and environment programme.

If you are evaluating an AI monitoring system for environmental or conservation applications, contact Omdena to discuss what a scoped implementation would look like for your context.