BHU-Rakshak is an environmental intelligence network: sensor nodes with on-device AI that detect landslides, floods, fires and pollution, and warn people before risks become disasters. They keep working without internet.
An early BHU-Rakshak landslide monitoring unit has been in pilot testing at an Indian Army camp in Assam since July 2026, running continuously with no downtime so far.
Troops carry it on convoy movements through the monitored stretch. Everything we learn there feeds straight into the next-generation node.
From the sensor on the slope to the alert on a phone: the seven building blocks of BHU-Rakshak, and where each one stands today. Download the 2-page brief (PDF)
Low-cost ESP32 nodes staked on site, each with its own sensors, logic and siren.
Every node learns its own site's normal, tracks trends and scores its own risk. No cloud is needed to decide.
One rule engine covers all seven hazards in the brief. Each alert is corroborated before it escalates.
Nodes report their GPS position, and a live map shows every node's risk level and latest readings.
Warnings reach people on site, in the control room and on their phones, each with a severity level and a confidence score.
Decisions happen at the edge. A base station keeps history, fits the failure forecast and exports data for analysis.
One base station serves many nodes, and the node's role is chosen on site, so a new slope or river costs a node, not a new system.
Every hazard the network watches runs through the same rule engine, with thresholds taken from Indian standards wherever one exists. Each value can be tuned per site.
Watches: Water level, rate of rise, rainfall
| Measure | Warning | Critical | Standard |
|---|---|---|---|
| Rate of rise | 5 cm per 10 min | 15 cm per 10 min | |
| Daily rainfall | 64.5 mm (heavy) | 115.5 mm (very heavy) | IMD |
| Danger height | Set for each river bank | Set for each river bank |
A river's danger height belongs to that bank alone, so there is no universal default. Until it is set, the system reports the height as unknown rather than guessing.
What people are toldMove to higher ground now. Do not cross moving water on foot or by vehicle. Knee-deep water moves a person.
Watches: Smoke, flame, temperature
| Measure | Warning | Critical | Standard |
|---|---|---|---|
| Smoke index | 30 | 60 | Site baseline |
| Flame | Detected | Detected at two nodes |
One node seeing smoke is capped at a warning. It becomes critical only when a second node agrees, so a cooking fire never triggers a district evacuation.
What people are toldMove away from the fire, crosswind, not uphill. Fire climbs a slope faster than a person can.
Watches: PM2.5, PM10
| Measure | Warning | Critical | Standard |
|---|---|---|---|
| PM2.5 | 91 µg/m³ (Poor) | 121 µg/m³ (Very Poor) | CPCB AQI |
| PM10 | 251 µg/m³ (Poor) | 351 µg/m³ (Very Poor) | CPCB AQI |
Bands follow the Central Pollution Control Board's National Air Quality Index, 24-hour averages.
What people are toldStay indoors, close windows and avoid outdoor work. Keep children, elderly people and anyone with breathing trouble inside.
Watches: Temperature, humidity
| Measure | Warning | Critical | Standard |
|---|---|---|---|
| Heat index | 41 °C (danger) | 54 °C (extreme danger) | |
| Air temperature | 45 °C | IMD heat wave |
Uses the heat index, not temperature alone: 40 °C at 20% humidity and 40 °C at 80% are very different for the human body.
What people are toldGet into shade, drink water and stop outdoor work in the middle of the day. Check on elderly neighbours living alone.
Watches: Carbon monoxide, nitrogen dioxide, ammonia
| Measure | Warning | Critical | Standard |
|---|---|---|---|
| CO | 50 ppm | 200 ppm | Exposure limits |
| NO₂ | 1 ppm | 5 ppm | Exposure limits |
| NH₃ | 25 ppm | 35 ppm | Exposure limits |
A dangerous gas concentration is a hard safety limit, so it can raise an alert on its own. Every other hazard needs corroboration.
What people are toldLeave the area crosswind, upwind if you can. Do not start vehicles or switch anything electrical on or off.
Watches: Turbidity, dissolved solids, pH
| Measure | Warning | Critical | Standard |
|---|---|---|---|
| Turbidity | 5 NTU | 25 NTU | BIS IS 10500 |
| Dissolved solids | 2000 ppm | 3000 ppm | BIS IS 10500 |
| pH | Outside 6.5–8.5 | Outside 5.5–9.5 | BIS IS 10500 |
Limits follow the Bureau of Indian Standards drinking-water specification.
What people are toldDo not drink or cook with this water. Boiling removes germs, not chemicals.
Watches: Tilt, vibration, soil moisture, rainfall
| Measure | Warning | Critical | Standard |
|---|---|---|---|
| Soil moisture | 35% (wet) | 45% (saturated) | |
| Tilt rate | 0.1° per 10 min | 0.1° per 10 min on saturated ground | |
| Forecast | Accelerating creep | Failure time from inverse velocity | Fukuzono |
Movement alone raises a warning. Movement on saturated ground is treated as a failing slope, and the forecast estimates the time left.
What people are toldMove off the slope and away from its base. Do not go back for belongings.
A one-minute 3D walkthrough of what happens when a monsoon hits a monitored slope, from the first raindrop to everyone reaching safety.
ALL SAFE
Sensor nodes on the slope measure tilt, vibration and soil moisture every second. Below them, the village goes about its day. Everything is within normal range.
BHU-Rakshak is a technology initiative developing intelligent, affordable and scalable solutions for environmental monitoring, disaster prevention and early warning.
We started in Darjeeling, where a single landslide can cut off NH-10, the only road linking Darjeeling and Sikkim to the rest of India. Landslide monitoring is our first application. The same platform is built to extend to floods, forest fires and infrastructure.
Developing practical technology for challenging environments.
Combining sensors, embedded systems and AI.
Building solutions designed for communities and critical infrastructure.
Four capabilities, one platform. Each deployment combines them to fit the site, from a single slope to a whole valley.
Continuous monitoring of environmental and geological conditions, from slope movement to soil saturation.
Technology that identifies dangerous changes early and raises a graded alert with time to act.
Long-range radio links that keep working where there is no mobile signal or internet.
Turning raw sensor data into clear information: risk levels, trends and forecasts.
Decisions are made on site, close to the ground being measured. A node keeps protecting its slope even if its link to the outside world is lost.
Nodes measure ground movement, vibration, moisture and air conditions around the clock.
Embedded firmware filters noise and learns what normal looks like at each site.
Trend analysis and established geotechnical methods estimate risk and time to failure.
Long-range radio carries readings to a gateway and base station, without cell towers.
Sirens on site and alerts on the dashboard, graded so people know how urgent it is.
Raw data never leaves the node. Each layer decides for itself and keeps working if the next one is cut off.
ESP32 · MPU6050 · GPS · soil, gas, smoke
ESP32 · LoRa 433 MHz
Raspberry Pi
bhurakshak.online
30 seconds of raw vibration is about 1,500 samples. What crosses the radio is a digest of 58 bytes or less, which is why the network runs on low-power LoRa with no mobile signal.
Choose a scenario to start the replay.
Replay from the BHU-Rakshak simulator, running the same risk engine as the device firmware.
Landslide monitoring in mountain regions is where we are deploying first. The same technology serves any site where conditions change and people need warning.
Environmental and slope monitoring in the Himalaya, where our first units are deployed.
Roads, bridges and critical infrastructure.
Monitoring hazardous and remote mining environments.
Early-warning systems for vulnerable settlements.
Environmental and fire-risk monitoring.
Monitoring in difficult-to-access environments.
Most monitoring assumes a mobile signal and a cloud server. The places that need warning most rarely have either.
Designed for difficult-to-access locations, with no mobile network needed.
Continuous monitoring. Each node warns locally even if its radio link is down.
Local processing and AI-assisted analysis, with alerts that need independent confirmation.
Designed to grow from a single device to multi-node networks under one base station.
Three parts that work together as one system: sensing in the field, communication on site, and a platform for the people responsible.
Environmental sensing and monitoring device, staked into the slope. Battery powered, with on-device risk scoring and a local siren.
How it worksCentral communication and monitoring unit. It collects every node's readings over long-range radio and corroborates alerts across nodes.
How it worksDashboard for monitoring devices, data and alerts. Public read-only view for communities, with a secure console for operators.
Open live dataEvery step keeps working if the next one fails.
Sensors continuously collect environmental data.
Embedded intelligence processes incoming information on the device.
Data is transmitted from remote locations over long-range radio.
Important changes reach the right people, on site and on the dashboard.
Built and tested for the conditions of the Eastern Himalaya.
The 1.2% figure comes from controlled data with a known failure time (24.7 min forecast against 25 min actual), not a real landslide. Field data from the monsoon is the next milestone.
Progress on the ground, in the lab and online.
bhurakshak.online now carries a public live view of deployed nodes, with a secure console for operators.
Open live dataThe redesigned node adds more sensing channels and on-device forecasting, and is being tested on the bench before field trials.
See the productsAn early version entered pilot testing with the Indian Army in Assam, running continuously on convoy routes.
Read moreNo. Nodes talk to the gateway over long-range LoRa radio, and the analysis runs on a base station at the site. Internet only matters if you want to watch the live data from somewhere else.
Partly, and we're specific about which part. The failure forecast uses inverse velocity, an established geotechnical method (Saito, 1965; Fukuzono, 1985) that works from physics rather than training data, so a new site is protected from day one. Each node also learns its own site's normal readings, and a classifier separates passing traffic from real ground movement.
It keeps scoring its own risk and can sound its own local alarm. The base station reports a silent node as silent rather than hiding it, so a missing reading is never mistaken for a safe one.
The highest alert level needs at least two independent channels to agree. A single sensor can raise a warning, but not an evacuation. In our truck-convoy test, the shaking was stronger than in a real landslide, yet the system correctly stayed safe.
In a controlled validation run, the forecast was 24.7 minutes against a true 25, an error of 1.2%. That was controlled data with a known failure time, not a real landslide. A full monsoon of field data is the next milestone.
Mostly the cost of a node. One base station serves many nodes, the analysis runs on site so there is no cloud bill per slope, and there is no per-site calibration, so scaling up doesn't need an engineer at every site.
An early version is in pilot testing with the Indian Army in Assam, running since July 2026. The next-generation node is in bench testing ahead of field trials.
Interested in deploying intelligent monitoring and early-warning solutions? We're looking for road authorities, district administrations and defence units to pilot with through the next monsoon.
BHU-RAKSHAK