IoT-based Low-Cost Time-Domain Reflectometry for Granular Soils: Sand-Specific Moisture Calibration and Machine-Learning Sand Identification

Soil is not just dirt; it is a complex mixture of solid particles, water, and air, and the balance between these ingredients dictates how plants grow, how foundations hold, and how water moves through the ground. For engineers and farmers, knowing exactly how much water is in the soil is critical, but measuring it without digging up the earth has long been a challenge. One established method involves sending a high-speed electrical pulse through the ground and listening for how it bounces back. Because water slows down this pulse more than dry sand or rock does, the time it takes for the signal to return reveals the soil's moisture content. However, this technique has a significant flaw: the standard formulas used to interpret these signals were developed for fine, clay-like soils. When applied to sand, which behaves very differently, these universal formulas often fail, leading to inaccurate readings that can cause irrigation systems to overwater or construction projects to underestimate ground stability.

A team of researchers at Rajshahi University of Engineering and Technology in Bangladesh set out to fix this problem specifically for sandy soils. They built a low-cost, internet-connected sensor system capable of measuring not just moisture, but also temperature and other soil properties. Instead of relying on a single, one-size-fits-all formula, they tested five different types of natural sand, each with slightly different grain sizes and textures. By collecting thousands of measurements in a controlled laboratory setting, they created a unique mathematical relationship for each specific type of sand, linking the raw electrical signal directly to the actual weight of water in the soil. Their work demonstrated that by tailoring the measurement tool to the specific material it is testing, they could predict moisture levels with an accuracy of 96.5 percent, a significant improvement over generic methods.

Beyond simply measuring water, the researchers discovered that the sensor data contained enough hidden information to identify the type of sand itself. They trained computer programs, known as machine learning models, to recognize the unique "fingerprint" of each sand sample based on its moisture and temperature readings. When the system was tested on an unknown sample, it successfully identified the correct sand type nearly 90 percent of the time. This is a crucial step forward because it means a sensor could potentially tell a farmer or engineer not just how wet the ground is, but also what kind of ground they are dealing with, without needing to take a physical sample to a lab.

The study began by gathering five distinct sand samples from the field. The researchers carefully analyzed the physical properties of each, measuring the size of the grains and how they were distributed. They found that while all five samples were classified as poorly graded sands—meaning they were mostly uniform in size—they still possessed subtle differences in their grain structure. To test the sensors, they prepared hundreds of soil specimens, varying the water content from very dry to quite wet. For every single specimen, they recorded the raw electrical signal from the sensor and compared it against the true moisture content, which was determined by drying the soil in an oven and weighing it. This process generated a massive dataset of nearly 1,900 paired readings, providing a solid foundation for their analysis.

When the researchers plotted the sensor readings against the true moisture values, a clear pattern emerged. For each type of sand, the relationship between the signal and the water content followed a straight line, but the slope of that line was different for every sand sample. One sand type required a much larger change in the signal to indicate a rise in moisture, while another responded more sharply. This finding proved that a single formula cannot work for all sands. When they tried to use one universal equation for all five types, the error in their predictions was noticeably higher. However, when they used a specific equation for each sand type, the predictions became remarkably precise, with errors dropping to less than three percentage points in most cases.

The team then took this approach a step further by asking if the sensor could identify the sand type without any prior knowledge. They fed the moisture and temperature data from the sensors into three different machine learning algorithms. These programs learned to distinguish between the five sand types by recognizing patterns in how the moisture and temperature changed together. The results were impressive: the system correctly identified the sand type in nearly 90 percent of the cases. The computer was particularly good at distinguishing the sands that had distinct temperature ranges, but it occasionally confused the three sands that shared similar warm temperatures. This suggests that while the current method is highly effective, adding more data points, such as electrical conductivity, could make the identification even more reliable.

To ensure their findings held up in the real world, the researchers tested the calibrated device on fresh soil samples taken from the field. They compared the sensor's moisture readings against the results from standard laboratory oven-drying tests. The device matched the laboratory values with an average accuracy of 96.5 percent, confirming that the specific calibration equations developed in the lab worked perfectly in practical, outdoor conditions. Furthermore, they demonstrated the full power of their system by taking an unknown soil sample, running it through the identification process, and successfully determining its type. Once the system identified the sand, it could instantly provide a detailed profile of its engineering properties, such as grain size and density, which are essential for construction and agricultural planning.

This work bridges the gap between simple sensor data and complex soil science. By moving away from generic assumptions and embracing the specific characteristics of different materials, the researchers have created a more reliable way to monitor the ground. Their integrated framework, which combines a low-cost sensor with smart data analysis, offers a scalable solution for anyone who needs to understand soil conditions quickly and accurately. Whether for managing irrigation in agriculture or ensuring the stability of a building foundation, the ability to know both the moisture level and the type of soil in real time represents a significant leap forward in how we interact with the ground beneath our feet.