AI-Powered Automated Underwater Marine Debris and Anomaly Detection System using Side-Scan Sonar Imagery
Ministry of Earth Sciences (MoES) · Renewable / Sustainable Energy · Software
Visually distinctive with a concrete deliverable and almost no competition, but the debris imagery to train on essentially does not exist, so be explicit about what your model was actually trained on and make the false-positive filtering your real contribution.
What it actually is
Abandoned fishing nets keep catching and killing sea life for decades after they are lost, and the only practical way to find them on the seafloor is sonar. But that means someone staring at thousands of kilometres of grainy acoustic scans, where a net looks a lot like a rock formation. The ask is software that finds man-made objects in those scans automatically and reports exactly where they are.
What to build
A sonar analysis pipeline with a detection or segmentation model trained to find man-made objects — the statement names shipwrecks, pipes, cylinders and entangled netting — against natural seafloor texture, robust to the acoustic problems it enumerates including speckle, acoustic shadows, varying pixel resolution and dropouts from vehicle heave, pitch and roll; a confidence and noise filtering stage that suppresses the false positives natural rock clusters and shadows generate and attaches a calibrated score to every detection; a geotagging engine parsing sonar metadata and ping headers to emit a structured JSON or CSV report giving latitude, longitude, bounding dimensions and classification per hazard; and a dashboard where an operator uploads a raw sonar log, watches detections overlay the waterfall and downloads the report.
Smallest thing that wins the room
Upload a raw sonar waterfall, watch detections appear over the acoustic imagery with confidence scores, and open the downloaded report to show each hazard's coordinates matching where it appeared on the scan.
How crowded this one gets
A guess, projected from the 2025 statements — the last year where both the submission counts and the winners were published.
Quieter than 51% of the 226 · #111 of 226 by expected field
A normal-sized field. Your idea has to be good, not miraculous.
Why: central ministry statements sat below the average.
This is a guess, not a fact
Nobody has published 2026’s numbers yet. This is an analysed estimate from last year’s pattern, so please do not take it as the truth — check the live counter on the SIH portal before you decide anything. The range covers the middle half of likely outcomes, so one statement in two lands outside it. Entry closes at 500 ideas per statement, so no range goes past that — a statement that reaches the cap fills and shuts rather than drawing an unlimited crowd. The model reads only three things a team can see before choosing — software or hardware, the theme, and what kind of body posted it — and those explain about a quarter of the variation in last year’s field sizes (R² 0.25 on held-out statements). Trust the band more than the number, and the ordering more than either. It cannot see how good your idea is, which is the part that actually decides it.
The scores
The number is the shorthand. The line under it is the reason.
Acceptance potential
3/5Visually distinctive, a well-structured deliverable and a marine acoustics domain almost nobody will enter are all real advantages, but the training data for the actual target — ghost nets and debris — barely exists, so a NIOT judge will immediately ask what your model was actually trained to find.
Feasibility
3/5Public side-scan sonar imagery exists but it is thin and dominated by shipwrecks and aircraft rather than the ghost nets and debris this statement is actually about, so you will train on a few hundred images of the wrong target class and claim debris detection — the model is buildable, the specific capability is not well evidenced.
Innovation scope
3/5The four deliverable components are enumerated and candidate architectures are named, but the genuinely hard part — distinguishing an artificial anomaly from natural seafloor texture under speckle and shadow — is left entirely open and is where any real contribution sits.
Clarity
4/5The four modules are specified individually, the acoustic challenges are named precisely, the report format and its required fields are stated, and the edge-deployment preference is explicit, though no accuracy target or evaluation protocol is given anywhere.
Effort
HeavyAssembling and augmenting a sonar corpus, training a detector robust to acoustic artefacts, building a calibrated confidence and filtering stage, parsing sonar metadata for geotagging and constructing an operator dashboard is five pieces, and the data assembly will take longer than the model.
Demo-ability
EasySide-scan sonar imagery looks genuinely unfamiliar and striking, so detections appearing over an acoustic waterfall are visually arresting in a way an ordinary photograph is not, and the downloadable geotagged report closes the loop concretely.
In its favour
- Green flag: Sonar imagery is visually unfamiliar to almost every judge, so your demo carries a novelty that an ordinary image detection pipeline never will
- Green flag: Geotagged structured output in a specified format is a concrete, checkable deliverable — a judge can verify that the coordinates in your report correspond to what appeared on the scan
- Green flag: Marine acoustics attracts essentially no student teams, so the competitive field here will be very small
- Green flag: The confidence and false-positive filtering module is the module that actually matters operationally — a cleanup vessel dispatched to a rock formation is expensive — and building it seriously distinguishes you from teams that ship a bare detector
Against it
- Red flag: Public side-scan sonar corpora are small and dominated by shipwrecks and aircraft, not ghost nets, so training on what exists and presenting it as debris detection is a substitution a marine technology judge will spot immediately
- Red flag: Side-scan geometry is not photographic — slant range distortion, variable along-track spacing and the nadir gap mean an off-the-shelf detector applied to raw waterfalls learns artefacts, so geometric correction is a prerequisite rather than a refinement
- Red flag: Acoustic shadows are simultaneously the strongest cue for a raised object and the biggest source of false positives, and a model that keys on shadow shape will flag every boulder on the seafloor
- Red flag: No accuracy target or evaluation protocol is given, so you must define both and a judge may measure you against a different standard than the one you chose
What you will be writing
- YOLO or U-Net detection on side-scan waterfalls
- speckle denoising and acoustic shadow handling
- slant-range and ping-rate geometric correction
- calibrated confidence scoring with false positive suppression
- XTF ping header parsing for geotagging
- ONNX edge-optimised inference for AUV deployment
- Marine acoustics
- Underwater computer vision
- Ocean conservation
Prior art to read before you start
side-scan sonar object detection · acoustic anomaly discrimination from natural texture · geotagged detection reporting
Analysed by Claude Opus. Every score above is a judgment call with its reasoning attached — kindly cross-check this against the official statement on the SIH portal before your team commits to it.