Scroll to move forward
2026 · 3 projects · Sherbrooke
geogrid
A web platform that replays an underwater inspection: the ROV video and its position move together on the map.
The expert annotates what they see; each observation keeps its timestamp and GPS point, and the assessment exports to GeoJSON.
geogrid
A deep learning model that finds water chestnut (Trapa natans), an invasive aquatic plant, in aerial imagery.
geogrid
A deep learning model that detects white-tailed deer, to count populations and track how they change.
2025 · 5 projects · Algiers
Algerian Space Agency
A web application that puts the national address register in the hands of those who use it: search for an address, view it, work with it.
Algerian Space Agency
Contributed to the development of a QGIS plugin that produces building numbering and street signage plans.
Algerian Space Agency
Monitoring locust outbreaks with geospatial analysis and remote sensing, in support of locust control.
Algerian Space Agency
Detecting, monitoring and assessing forest fires with remote sensing, in support of field teams.
Algerian Space Agency
A platform that handles the full meteorological data life cycle: management, format conversion, archiving and publication.
2024 · 6 projects · Blida and Arzew
SARL Louifi Projects · Soumaa, Blida
Laying out the centreline of section 02 of the bypass with a Trimble R10 GPS in RTK mode.
SARL Louifi Projects · Sidi Moussa, Algiers
Detail survey with a Leica TS06 total station, for the planned Sidi Moussa bypass.
Louifi Group
Built the group’s two websites: SARL Louifi Projects, public and hydraulic works, and SARL Louifi Motors.
Space Applications Centre (ASAL) · Space Applications Centre
Final-year project: a web mapping platform for forest fire management, with FIRMS thermal anomalies and a multi-role alert system.
Space Techniques Centre (ASAL) · Space Techniques Centre
A pix2pix GAN model that raises the resolution of digital elevation models from 30 m to 15 m, for flood risk. Abstract accepted for an MDPI paper.
ENSGTS
A GIS-based multi-criteria analysis: apartments ranked against several criteria, weighted with the analytic hierarchy process.
2023 · 5 projects · Arzew
Space Applications Centre (ASAL) · Arzew, Space Applications Centre · 35.86° N / 0.31° W
Four techniques compared in Google Earth Engine: SVM, random forest, decision tree and NBR index thresholding.
Carried out at the Space Applications Centre, as part of the post-fire monitoring process.
ENSGTS · El Tarf · 36.77° N / 8.31° E
Change detection over the city of El Tarf between 2013 and 2017, using remote sensing.
To be written: the imagery used, the method and what the result shows.
ENSGTS · Arzew, ENSGTS · 35.86° N / 0.31° W
Static GNSS observation campaign, during a two-week GNSS internship at ENSGTS in June 2023.
To be written: the points observed, session lengths, processing and the accuracy achieved.
ENSGTS · Arzew, ENSGTS (Metz data) · 49.12° N / 6.18° E
Producing a topographic map of the city of Metz by photogrammetric plotting.
Two-week photogrammetry internship at ENSGTS, in June 2023.
ENSGTS · Arzew, ENSGTS
Producing a digital terrain model and a digital surface model from drone data, in Agisoft.
“Quarry and cadastre” sample dataset published on the Pix4D website.
2022 · 1 project · Arzew
ENSGTS · Arzew, ENSGTS · 35.86° N / 0.31° W
Precise levelling traverse, during a two-week internship at ENSGTS in June 2022.
To be written: the traverse, the misclosures and the accuracy achieved.
2026 · Sherbrooke
2026 · Inspection platform
After a dive, the expert is left with two things that don’t talk to each other: hours of video and a navigation log. geoRov brings them back together — image, track and annotation at the same second — and delivers a georeferenced assessment that any GIS can open.
An underwater inspection leaves two records, captured separately. On one side, the videos from the ROV cameras, several gigabytes per dive. On the other, a binary navigation log: USBL acoustic positioning coupled with the surface vessel’s inertial navigation system, writing position, depth and attitude fifty times per second.
Working out where the ROV was at minute 47 of a video means cross-checking timestamps by hand. And an observation recorded without its position is of little use: you know what was seen, but not where.
geoRov brings both into a single web interface. The video plays on one side, the dive track is drawn on the other, and a cursor follows the ROV on the map in exact step with the image.
When the expert spots something, they annotate it from the video: position and time are captured at the moment of the click, and the observation appears immediately in the list and on the map, without interrupting playback. At the end, the assessment exports to GeoJSON, ready to open as is in a GIS.
Before a dive can be explored, its data goes through an offline processing chain in Python. This is where synchronization happens: the camera logs carry the start and stop recording markers, and these are what split the continuous telemetry into segments matching the videos.
This is the core piece. A 50 Hz telemetry file for an hour of video is too heavy for a browser, so it is split up, and the interface only downloads the chunk for the minute currently playing.
Video time drives everything else. It selects the chunk to load, then the exact position within that chunk; this position feeds both the cursor on the map and the annotation form, which locks the GPS point at the moment of the click.
The server never loads an entire video into memory. It answers the player’s range requests: clicking anywhere on the seek bar takes the stream to that point without downloading what comes before. Playback is paced by a one-megabyte buffer — enough for a steady stream, small enough for several people to watch at once.
All disk reads and writes are asynchronous: saving a long assessment doesn’t freeze video streaming for others. And the server checks that every requested path stays within the data folder, refusing anything else.
Observation categories are configurable: they are edited from the interface and written to the mission configuration, so the review grid can be adapted to the type of inspection.
Saved assessments can be reloaded to be resumed or checked. The deliverable is a GeoJSON file: a format any GIS opens, with no bridge or conversion.
The platform is containerized: the interface is built, then served as static files by a lightweight server, while the API runs alongside in its own container. A single command builds and starts the whole stack.
The videos, however, don’t fit in a container — several hundred gigabytes. The physical folder holding them is mounted directly: it is the only setting to check before deploying on a new machine.
Two choices carried everything else: letting video time drive synchronization rather than a clock, and splitting the telemetry before it reaches the browser. Without them, nothing held up at 50 Hz.
To be written: the known limitation, and what I would do differently.