Soil Moisture SK4
Description
International Soil Moisture Network (ISMN) site within Canada’s Real-Time In-Situ Soil Monitoring for Agriculture (RISMA) network in Saskatchewan, station 4. From 2014–2021, 15-minute in-situ record comprising multi-sensor soil moisture and temperature at six depths (0–5 cm to 150 cm) and collocated meteorology. The site has frequency-domain reflectometry sensors (Stevens HydraProbe II) installed at depths of 5 cm, 20 cm, 50 cm, 100 cm, and 150 cm below the surface. An additional surface measurement of the 0–5 cm layer is also recorded, effectively providing six distinct depth readings (surface 0–5 cm, 5 cm, 20cm, 50 cm, 100 cm, 150 cm). The station also measures air temperature, precipitation, relative humidity, wind speed, and wind direction at a nearby weather station, and it monitors battery voltage for maintenance. Data from this station are logged at 15-minute intervals, providing a high-resolution time series from 2014 through 2021. This data was used to research on "Multi-Scale Soil Moisture Prediction Using Recurrent Neural Networks with Temporal Attention."
Files
Steps to reproduce
We implemented a transparent data preprocessing and feature engineering workflow to prepare the raw sensor data for model training. Sensor Fusion: The raw data were first cleaned to handle missing or implausible readings caused by telemetry faults or power outages. Multiple co-located probes at the same depth were fused by averaging valid values (excluding zeros or gaps), which reduces random noise and sensor bias. The same approach was applied to soil temperature sensors. Meteorological variables—air temperature, rainfall, relative humidity, wind speed, and wind direction—were retained as individual features. Antecedent Aggregation: To incorporate short-term dependencies and cyclic patterns, the 15-minute data were aggregated to hourly resolution. Using the hourly dataframe, we computed several rolling aggregates: a 24-hour cumulative rainfall (capturing daily precipitation), 24-hour means for air temperature, wind speed, and relative humidity, and a 3-day rolling mean of the target soil moisture to smooth local fluctuations. These engineered features capture antecedent wetness, thermal inertia, and atmospheric influences relevant to soil dynamics. Seasonal and Diurnal Cycles: We added cyclical encodings of time to represent repeating patterns: sine and cosine transformations for day of year (period 365) and hour of day (period 24). This allows models to learn seasonal and daily rhythms without artificial discontinuities. Physical State Features: To account for unreliable dielectric readings during frozen conditions, we derived a binary feature, *is_frozen*, set to 1 when shallow soil temperature (<5 cm) drops below 0 °C. This informs the model that moisture readings may be physically constrained. Wind Direction Encoding and Scaling: Wind direction (in degrees) was decomposed into sine and cosine components for continuity. Missing values were filled using linear interpolation for short gaps and 24-hour rolling means for longer ones. For longer seasonal gaps, values from the same day of the previous year were used. Finally, all features were normalized with Min–Max scaling (0–1) to stabilize neural network convergence.
Institutions
- Acadia University