UT San Antonio's off-grid flood node runs its warning model on the microcontroller

Validation accuracy answers a narrow question: how often the model's output matched the labels on data held back from its own training run. The article reports 98.82% as an initial assessment after training, and it does not split the remaining errors into missed floods and false alarms ^[9][18]. Those two failures cost different things. One leaves a low-water crossing unflagged. The other teaches a public works crew to ignore the radio.

The split matters here because water is rare in any sensor log. A classifier that answers "dry" almost always will score well on a set where most samples are dry, and a single accuracy figure cannot separate that behaviour from genuine detection. The measurements that would settle it are inundations caught per inundation that occurred, and false alarms per node-month of ordinary rain. Both numbers come out of deployment ^[10].

The sensor stack spreads each decision across several readings. Four optical water-level sensors sit at varying heights, with temperature, humidity, light and precipitation channels beside them ^[6]. A fouled lens is then one input among several. The article says this multimodal approach is far more reliable than traditional single-metric gauges ^[7].

The radio numbers set the deployment bill. Ten times the range covers a hundred times the area: a hub reaching 8 km in open ground covers roughly 201 square kilometres, while one held to 0.8 km by city streets covers about 2 ^[12]. The neighbourhoods that need block-level warning are the ones where a regional model misses a drainage channel or a campus access road going under while the surrounding blocks stay dry ^[15].

Power comes from an onboard chip that charges a battery and backup storage from ambient light, which the team says lets the node run indefinitely on a fraction of the energy a smartphone uses ^[13]. Pan, who directs UTSA's RISE Lab, put the reasoning for that plainly. "In many rural areas or coastal communities, power infrastructure can fail right when severe weather strikes," he said ^[4][3]. "Our system is an off-grid solution. It generates its own power, evaluates flood risk locally right on the device and sends timely warnings without needing external electricity or expensive network lines," Pan said ^[5].

The node's output is a lightweight radio packet to a local hub. There, custom server software aggregates several deployed nodes across a town or watershed and a second, larger AI model runs on the pooled data ^[17]. Emergency teams watch that hub from miles away ^[11]. Mimi Xie at UTSA and Hua Zhang and Wenlu Wang at Texas A&M University-Corpus Christi are on the team ^[16]. The commercial stations this is meant to displace depend on grid power or frequent battery replacement, and they go down during multiday storms and outages ^[14].