# SHYENA (श्येन): low-altitude airspace awareness and counter-drone planning SHYENA helps border units see small drones before they cross, put sensors where coverage is missing, respond lawfully, keep evidence, and keep their own drones on their side of the line. It is defensive: detect, track, classify, alert, plan lawful non-kinetic responses by authorised agencies, and preserve evidence. It gives no weapon-targeting guidance. - Console: `/dashboard/shyena/` (overview, incidents, coverage planner, airspace and own drones, live replay, playbook, methods) - Product page: `/products/shyena.html` - Pipeline: `pipeline/trinetra/shyena/`, exports in `dashboard/data/shyena/` - Tests: `tests/test_shyena.py` (offline, fixtures in `tests/fixtures/shyena/`) ## 1. Users and the decisions they make | User | Decision SHYENA supports | |---|---| | Border units and sector headquarters | Where the line and the drop corridor lack coverage; which sensors close the gaps; how much warning a crossing gives | | State police anti-drone and checkpoint units | When and where drops happen; who is alerted at which level; how to recover payloads safely | | Own-drone operators | How far inside the border to set a hard and a warning geofence | | Airport and airspace authorities | Which aerodromes sit near the border; where counter-drone measures need coordination | | Analysts and researchers | What the official figures are, who counted them, and how reports disagree | ## 2. What it does, page by page - **Overview.** Headline figures with citations, a sector map (approximate border, district outlines, the 25 km border belt, aerodromes and advisory zone rings), drones recovered by year, 2025 reported to date, the Ministry of Defence 2025 breakdown, reported tactics, the three coverage layouts, simulated outcomes and the media-reported incident sample. - **Incidents.** Every official or press figure (43 figures from 18 reports, read 21 Oct 2026) with scope, period, to-date flag, source and a citation check, downloadable as CSV; sightings against recoveries; drone-linked drug cases and drugs seized (Ministry of Home Affairs, Rajya Sabha reply); the reported 22:00 to 04:00 peak; a media-reported incident sample from GDELT with district, outcome, payload, quantity and source links. - **Coverage planner.** Radar, EO/IR, RF and acoustic sensors on the 238 km Amritsar and Tarn Taran line. Three illustrative layouts; add, drag, edit and remove sensors; line coverage by type, layered and RF-independent coverage; gaps; drop-corridor coverage (0 to 10 km inside); warning time; a km-by-km strip; range assumptions with sources; physics calculators; the fewest sensors of each type that cover the whole line; export. - **Airspace and own drones.** Aerodromes within 100 km of India's land borders (52 neighbouring aerodromes within 50 km), advisory drone zones, the 25 km border belt, and a geofence setback calculator with open-sky, degraded and suspected-interference GNSS settings. Two stretches: Amritsar and Tarn Taran, and the convoluted India-Bangladesh line around Patgram. - **Live replay (simulated).** One seeded night: tracks coloured by belief, classification confidence over time, the evidence behind it, alerts (Previous and Next step through warning, danger and search alerts), human-approved responses, ground truth on request, and per-drone outcomes. The night downloads as a trial log in the form an operational log would take. - **Playbook.** Detect, classify, alert (with the police handoff for inland tracks), respond (authorised agencies, lawful non-kinetic means, a person decides), evidence, safe payload recovery, review, and test and certify equipment before fielding it (the Ministry of Home Affairs' Anti Rogue Drone Technology Committee). Each step names who acts and when it is done. Every claim is cited. - **Methods.** Sources and licences, formulas, thresholds, assumptions, limits, reproduction. ## 3. Data | Source | Use | Licence | Notes | |---|---|---|---| | BSF, Punjab Police, Ministry of Defence and Ministry of Home Affairs figures as reported by The Tribune, All India Radio, The Week (PTI), Swarajya, Unmanned Airspace, Indian Express (via SuperKalam), ETV Bharat, Millennium Post, Rozana Spokesman, Punjab Kesari (IANS) and the Press Information Bureau | Official totals and tactics | Figures from cited reports | Read 21 Oct 2026; each report re-checked for the quoted numbers on every run; second-hand summaries marked | | GDELT DOC 2.0 API | Media-reported incident sample, 2024 to date | Free with attribution | Quarterly windows; rate limited from shared addresses; failed windows listed | | OpenStreetMap via Overpass API | Border lines, district outlines, place names | ODbL 1.0 | Approximate; extract is ODbL | | OurAirports `airports.csv`, `runways.csv` | Aerodromes and runway geometry | Public domain | Airports only; heliports excluded | | Natural Earth 1:10m boundary lines | Distance from aerodromes to the border | Public domain | About 1 km accurate | | Peer-reviewed and government sources | Sensor range anchors, response and evidence practice | Cited | See section 5 | | SHYENA simulator | Night replay and layout comparison | Seeded (20261021) | Labelled SIMULATED everywhere | Every export has an entry in `dashboard/data/shyena/manifest.json` with rows, bytes, sources, retrieval time and whether it is simulated. Exports: `aggregates.json`, `events.json`, `border.json`, `airspace.json`, `coverage.json`, `sim.json`, `sim_summary.json`. ## 4. Key figures (as reported) - BSF Punjab Frontier drones intercepted: 21 (2022), 107 (2023), 294 (2024). For 2025, reports give 272 (to 30 Nov), 275 and 278 (reported 3 Dec) and 287 (reported Sep 2026). First quarter of 2026: 47. - Punjab Police and BSF combined: 313 drones and 385.119 kg heroin in 2025; 161 drones and 392.355 kg heroin from 1 Jan to 27 Jul 2026. - Ministry of Defence, western front, 2025: 791 intrusions, 237 brought down (72 with narcotics, 5 with ammunition, 160 with no payload). - Reported tactics: activity peaks 22:00 to 04:00; drones fly 3.5 to 10 km inside; drops in fields; about 500 g payloads; GPS-programmed routes; jammed drones fly home; anti-jam and satellite-linked models are appearing. - Ministry of Home Affairs (Rajya Sabha reply, July 2026): 298 of 305 drone-linked drug trafficking cases in India in 2025 were on the Punjab border; about 460 kg of drugs were seized in those cases (448 kg heroin, 9 kg methamphetamine, 3 kg opium). - Aircraft reported 465 GNSS interference or spoofing incidents from Nov 2023 to Feb 2025, mostly around Amritsar and Jammu. ## 5. Methods ### Border line OpenStreetMap border ways (admin_level 2, nation) are joined and oriented north to south, so India is on the left. Each vertex takes the district whose outline contains a probe 600 m inside India; gaps of up to 4 km are split between neighbouring labels. Check: the Punjab line measures 553.9 km against the official 553 km. The Amritsar and Tarn Taran stretch is 237.9 km (Amritsar about 130 km, Tarn Taran about 108 km). ### Airspace Zones follow the Ministry of Civil Aviation airspace map (24 Sep 2021): yellow from the ground up between 5 and 8 km of an operational airport's perimeter, green up to 200 ft (yellow above) between 8 and 12 km, and green up to 400 ft elsewhere unless designated red or yellow. The release gives no distance for red zones; SHYENA draws the area within 5 km as red because the yellow band starts there, and labels that as an inference to check against the Digital Sky map. The perimeter is approximated by the runway strip (centreline plus 150 m each side, 60 m beyond each end; ICAO Annex 14). The 25 km border belt follows the Ministry's 5 Aug 2021 statement and Punjab's 2022 district orders; the guideline text itself could not be retrieved from our network. The Digital Sky map is the authority. ### Coverage A sensor covers a point within its planning range (great-circle distance, sphere of radius 6,371,008.8 m). The line is sampled every 50 m; the drop corridor is a 500 m grid within 10 km on the Indian side. Metrics: coverage by type, any, layered (two or more types), RF-independent (radar, EO/IR, acoustic), redundant (two or more sensors), gaps, corridor coverage. Warning depth at a line point p with outward normal n, for a sensor at q with range R, v = q − p and a = n · v: depth(p) = max over covering sensors of a + sqrt(a² + R² − |v|²) warning = depth / 12 m/s Minimum sensor counts use an interval sweep over candidate positions every 250 m (100 m for acoustic) at default setbacks. Planar work uses a spherical transverse Mercator frame (Snyder 1987). The browser recomputes the same metrics and checks them against the pipeline on load. ### Sensor planning ranges | Type | Short | Typical | Long | Anchor | |---|---|---|---|---| | Radar | 1 km | 2 km | 5 km | X-band FMCW radar detected a Phantom 4 at up to 2 km (de Quevedo et al., EuRAD 2018); R ∝ σ^(1/4) | | EO/IR | 0.8 km | 1.5 km | 3 km | Johnson detection (2 px) of a 0.35 m target, 12 µm pixels, 100 mm lens: 1.46 km | | RF | 1 km | 3 km | 8 km | Free-space link budget (ITU-R P.525) with 10 to 20 dB margin; low-cost SDR 0.5 to 1.2 km (arXiv 2207.10795) | | Acoustic | 0.15 km | 0.3 km | 0.5 km | 294 m for the smallest UAS tested (Sedunov et al., IEEE HST 2018); ISO 9613-2 model gives 299 m at 30 dB(A) | CISA (Nov 2025) supports the ordering: radar generally longest, RF longer than EO/IR and acoustic, acoustic shortest; non-RF-emitting drones cannot be heard by RF detectors. ### Simulator One night, 18:00 to 06:00 IST, 2 s steps, seed 20261021, on the "full line" layout. Eight drone crossings (a stress test; the 2025 western-front average is about 2.2 a night) launched with weight 1.0 inside 22:00 to 04:00 and 0.35 outside; drop points 3.5 to 10 km inside; payload 0.5 kg (60 %), 1 kg (25 %), 2 kg (15 %); a quarter on GPS routes without a live link; 15 % anti-jam. Eighteen night-migrant bird tracks, thirty radar clutter events, ten RF false alarms, one friendly patrol drone. Detection probability P_d(r) = P_max / (1 + exp((r − 0.95 R) / (0.04 R))). Classification multiplies the odds by likelihood ratios (RF protocol match ×25, camera drone ×12, camera bird ×0.08, rotor micro-Doppler ×4, wing-beat ×0.3, purposeful flight ×3, wandering ×0.4, hover ×4), each counted at most three times. Alerts: watch, warning (P ≥ 0.5), danger (P ≥ 0.85 inside India with a camera or RF identification and a non-RF fix), search (hover of 16 s or more inside India). A person approves any response after 45 to 120 s; disruption works within an assumed 8 km of one of three illustrative response cells, never on GPS-route or anti-jam drones, and otherwise returns the drone home (75 %), lands it (20 %) or has no effect (5 %). Monte Carlo, 50 nights per layout (400 crossings each): drops that raise a search alert rise from about 20 % with fence-line layouts to about 65 % with a second row 6 km inside. Drop-area prediction (for the police handoff): with a track's first warning, an annular sector from its latest position, ±30° about its heading over the last 60 s, 3.5 to 10 km inside the line (the reported depth band, measured from the track's own depth). It uses only the track, never the planned drop point. Over 50 nights per layout, 77 to 79 % of drops fell inside an area that had been sent (median about 50 km²). Optimistic: simulated drones turn at most 25° at the line. ### Own-drone geofence hard fence = e + (v + w)(t_l + t_f) + [v²/2a + w·v/a] + w·h/v_d + m warning = hard fence + (v + w) × pilot reaction time Defaults (v 15 m/s, w 10 m/s, t_l 3 s, t_f 2 s, a 4 m/s², h 120 m, v_d 12 m/s, e 5 m, m 50 m) give about 350 m and 850 m. Open-sky GNSS error from GPS.gov (4.9 m); under suspected spoofing the fence cannot be trusted. ## 6. Safeguards (doctrine) - Events, places and conditions only. Headlines are not stored; sources whose URL looks like it names a person are dropped (the count is published); arrest stories are excluded. Military drone attacks during hostilities (May 2025) are dropped from the smuggling sample and counted apart. - No border outpost, troop position or person is mapped; security-force helipads are excluded. Aerodromes come only from public aeronautical data (OurAirports), including air force stations and advanced landing grounds, because own drones must keep clear of them. - Responses only through authorised agencies (MHA permission for counter-drone jammers; MoD for the defence forces; Cabinet Secretariat Jammer Guidelines 2025). Kinetic options are out of scope. - Software recommends; trained people decide; every decision is logged, including "no action". - Simulated data labelled SIMULATED in the UI and `"simulated": true` in JSON. Zones are advisory. ## 7. Limits and failure modes - Planning ranges are generic; terrain, vegetation (sugarcane, elephant grass along the Ravi), villages, fog and RF noise change them. Validate with a site survey. - The border line is approximate (tens of metres; more along shifting river channels). - Official figures come through the press; agencies and reports count differently. - The media sample depends on GDELT coverage and headline wording; failed windows are listed and shown as no data, never as zero. GDELT rate-limits shared addresses: in the 21 Oct 2026 build, 6 of 12 quarterly windows (Apr 2025 to Sep 2026) could not be fetched, so the sample covers Jan 2024 to Mar 2025 and Oct 2026 only. - Zone rings approximate perimeters and ignore temporary restrictions. - The simulator has no association errors, terrain masking or weather, and its likelihood ratios are assumptions. ## 8. Validation - Punjab border length from OSM: 553.9 km against the official 553 km. - Browser coverage math matches the pipeline to within 0.0003 on every preset metric. - Acoustic model reproduces the measured 294 m within 2 % at 30 dB(A) ambient. - Citation check: every cited report is fetched once per run and searched for the numbers and key words SHYENA quotes; 17 of 18 matched automatically and 1 (pib.gov.in, which refuses automated clients) was checked by hand. - 82 unit tests: geodesy against analytic values, transverse Mercator round trips, coverage chords, warning depth, greedy counts, headline parsing, clustering, privacy filter, zone geometry, simulator determinism and return-to-home behaviour, citation checks, export size limits and manifest merging. ## 9. Reproduce cd pipeline python3 -m trinetra.shyena run # fetch or reuse the cache python3 -m trinetra.shyena run --offline # rebuild from cached responses only python3 -m trinetra.shyena run --only coverage cd .. && python3 -m pytest tests/test_shyena.py -q ## 10. Next - Terrain and vegetation line of sight from a public elevation model. - Ingest real sensor logs and fit the likelihood ratios to them. - Feed SANGAM's fusion engine instead of the simplified tracker. - Pull zones from the Digital Sky API when access is granted. - Village-level incident geocoding with a published precision.