RiceRat AI is a low-cost rodent-inspired micro-robot designed for
greenhouses and high-bed farms. It moves close to the ground on
Medium-Tread Rubber Wheels and collects localized
field data through a Front RGB Camera, a
Retractable Slim Soil Moisture Probe for moisture
and temperature, Side Vent Sensors for air
temperature, humidity, and light, and an onboard
Control Board (ESP32/RPI). A
Top-mounted LED Light supports low-light
operation, while a
Flexible Signal Loop Antenna and a
Tail Antenna & Charging Port (Wi-Fi + Power)
keep the robot connected and rechargeable via a
Charging Dock (USB-C). A
Chassis Recovery Point allows easy manual
retrieval when needed. The system combines crop image analysis
with environmental sensing because stress is rarely captured by
one signal alone. By following repeated routes, it generates
consistent observations from the same rows and checkpoints.
The first version uses lightweight and explainable models —
YOLO, MobileNet, or
ResNet for crop image analysis, and
SVR for sensor-based moisture estimation. During
the pilot phase, fusion models such as
Extra Trees, Random Forest, or
XGBoost will combine image-derived features with
sensor readings once paired local data is available. Public
datasets bootstrap development, while the system adapts using
local pilot data labeled as dry,
normal,
warning, or
needs checking
— human-annotated categories that complement the computed
bands (Healthy,
Mild,
Moderate,
High,
Critical).
Preliminary experiments validate each module of the
stress-monitoring pipeline before pilot deployment.
Crop Stress Alert.
Four image classifiers were
. YOLO11s leads at 99.88 % accuracy on PlantVillage, while
MobileNetV3 Small offers a lighter alternative better suited to
the robot’s modest hardware — trading some accuracy for
significantly higher throughput during real-time patrol.
Open-set robustness.
In real fields, early stress symptoms may not match any labelled
disease class. To handle this, the pipeline adds an
: Mahalanobis-distance scoring on MobileNetV3 Small embeddings
achieves ROC-AUC = 97.26 % on PlantVillage and 87.15 % on Paddy
Doctor, letting the robot flag unseen abnormalities for human
review instead of force-fitting them into a known class.
Moisture Scout.
An
shows all models clustering near the mean-predictor level
(R² ≈ 0). This confirms what the proposal already assumes: public
remote-sensing proxies alone cannot capture local moisture
patterns — the real value of RiceRat lies in collecting
ground-truth probe readings along repeated patrol routes. The
pilot phase will replace these proxies with in-situ sensor data
where meaningful forecasting becomes feasible.
Sensor-to-map pipeline.
The current
is computed entirely from sensor signals (moisture, VPD,
temperature). Visual disease output feeds a separate alert flag on
the Risk Map Dashboard, and multi-modal fusion of both streams is
planned for the pilot iteration.
Decision-support extension.
Although yield prediction is not part of the robot’s first
operational loop, a
shows that the environmental features RiceRat collects also carry
long-term predictive value (Extra Trees R² = 0.9905 across 55 crop
types; rice-only LightGBM R² = 0.6425). This opens a path from
early-warning monitoring toward broader decision-support in future
versions.
How is the Stress Score (σ) calculated?
At each checkpoint the robot records four sensor readings —
soil moisture, air temperature, relative humidity, and soil
temperature — and combines them into a single
Composite Stress Score σ on a 0-to-100 scale.
The score is a weighted sum of three sub-scores, each
capturing a different type of crop stress
[4]:
σ = 100 × (0.55·Sm + 0.25·Svpd +
0.20·Stemp)
Every sub-score is clamped to the 0 – 1 range by a
clip(x) = min(max(x, 0), 1)
function, so that 0 always means no stress and 1 always means
maximum stress. The three components are described below.
-
Soil Moisture Stress (Sm) tracks
how far moisture has fallen from a well-watered baseline.
The formula
Sm = clip((70 − SM) / (70 − 40))
gives a score of zero when the soil reading SM sits at or
above 70 %, meaning the crop has adequate water. As SM drops
toward 40 % the score rises linearly until it reaches its
maximum. These thresholds are drawn from the FAO-56 water
stress coefficient
[4], adjusted for
rice's particular sensitivity to water deficit
[5].
-
VPD Stress (Svpd) measures how
strongly the atmosphere draws moisture from leaves.
Vapour-pressure deficit (VPD) is first computed from air
temperature and relative humidity
[4], then mapped to a
stress level with
Svpd = clip((VPD − 1.0) / (2.3 − 1.0)).
Below 1.0 kPa the canopy is comfortable and the score stays
at zero. Above 2.3 kPa stomatal conductance in rice drops
sharply and the score saturates at one
[6].
-
Temperature Stress (Stemp)
blends an air and a soil component through
Stemp = 0.7·Sair + 0.3·Ssoil. Air stress is zero within the 20 – 35 °C comfort zone and
rises linearly on either side
[7], expressed as
Sair = max(clip((20 − Tair) / 10),
clip((Tair − 35) / 5)). Soil stress follows the same pattern with a narrower safe
window of 20 – 34 °C, based on Yoshida's rice physiology
thresholds [8], giving
Ssoil = max(clip((20 − Tsoil) / 5),
clip((Tsoil − 34) / 6)).
Soil moisture receives the largest weight (55 %) because water
deficit is the primary yield-limiting factor in rice
[5], followed by VPD (25
%) and temperature (20 %). The resulting σ maps to five action
bands:
0–20 Healthy
20–40 Mild
40–60 Moderate
60–80 High
80–100 Critical
Four public datasets used to bootstrap early model training and
validate prediction pipelines before pilot data collection.
Four image classifiers benchmarked for disease / stress recognition
across PlantVillage (15 classes) and Paddy Doctor (10 classes),
ranked by accuracy.
Open-set anomaly detection on MobileNetV3 Small embeddings, trained
on healthy images only, ranked by ROC-AUC.
Early baseline for soil-moisture forecasting on the Mendeley
dataset, ranked by RMSE.
Yield regression on the Kaggle crop dataset — all 55 crop types and
rice-only subset, ranked by RMSE.
