How Machine Learning Is Transforming Soil Sampling and Soil Mapping

‍

In this week’s Passion Academy, we revised 3 scientific articles about soil sampling and soil mapping for geographical applications.

‍

Soil sampling is about finding the minimum number of sample points while still getting a good understanding of an area. Soil mapping is about using those points to predict soil properties in places where no samples were taken. These methods are widely used in agriculture, environmental studies and in climate research to monitor land and better understand soils.

‍

The articles explore different machine learning techniques used for these problems: clustering methods & genetic algorithms for sampling, Random Forests for soil mapping, and Convolutional Neural Networks for analyzing historical satellite images. These approaches help reduce time, effort and cost. This is a great example of how data science and geography work together

‍

‍

The Core Problem: Where Should You Sample?

Soil data is often collected to measure attributes such as:

• Organic matter
• Soil carbon
• Nutrient levels
• Moisture content
• Soil type

The challenge is deciding where these sampling points should be placed.

‍

A naive approach would simply spread sampling points evenly across an area. While straightforward, this often creates two major problems:

• Too many sampling points increase operational costs
• Poor placement can miss important variation in soil conditions

‍

Large agricultural regions, environmental monitoring projects and land development initiatives all face this challenge.

‍

The goal is simple: Collect as little data as possible while learning as much as possible.

‍

‍

Why Geography Matters

‍

A major principle behind modern soil sampling comes from something called the Third Law of Geography:

‍

"The more similar environmental conditions are between two locations, the more similar their geographical properties are likely to be."

‍

In practical terms: If two locations have similar rainfall, elevation, vegetation and climate conditions, their soil properties may also be similar. This allows researchers to make smarter sampling decisions rather than treating every location equally.

‍

‍

Adaptive Sampling

One approach highlighted in the session was:

‍

Adaptive Uncertainty Guided Stepwise Sampling (AUGSS)

‍

The goal is to reduce uncertainty while using as few samples as possible.

‍

The process works by:

• Selecting initial sampling locations
• Training a prediction model
• Identifying areas with high uncertainty
• Collecting new samples only where additional data is most valuable

‍

This repeats until:

• A sampling budget is reached
• Or prediction improvements begin to plateau

‍

Rather than sampling blindly, the system continuously learns where new information matters most.

‍

Models used in this process include:

• Classification and Regression Trees (CART)
• Random Forests (RF)
• Support Vector Regression (SVR)

‍

‍

Genetic Algorithms for Smarter Sampling

‍

Another interesting approach explored was: Genetic Algorithm Sampling (GAS)

‍

This method balances two competing goals:

• Spatial coverage
• Environmental representativeness

‍

Rather than adding samples step by step, GAS searches for optimal sampling strategies using evolutionary optimisation techniques.

‍

It was tested in flat terrain environments and compared against methods such as:

• K Nearest Neighbours
• Fuzzy C Means Clustering Sampling
• Random Forest models

This is particularly useful when trying to optimise large scale land surveys.

‍

‍

Soil Mapping After Sampling

‍

Collecting samples is only half the challenge. Once physical measurements are collected, researchers still need to predict conditions across the rest of the landscape. This is known as soil mapping.

‍

Machine learning models help interpolate between known sampling points and estimate values in areas where no physical testing has occurred.

‍

These predictions can be used to map:

• Soil organic carbon
• Nutrient density
• Soil classification
• Agricultural suitability

‍

Without interpolation, organisations would need significantly more expensive physical sampling.

‍

‍

Using Satellite Data to Improve Predictions

‍

One of the most interesting developments is combining soil sampling with historical satellite data.

‍

Researchers explored using:

• Satellite imagery
• Vegetation “greenness” over time
• Elevation data
• Annual precipitation data
• Other environmental variables

‍

These datasets were combined with physical soil measurements to predict soil organic carbon at regional scale.

‍

In one study:

• ‍733 physical samples collected across multiple years were used as ground truth data.
• Deep learning models were then trained to improve prediction accuracy.

‍

This creates a much richer picture than relying purely on physical sampling.

‍

‍

Why This Matters

‍

Better soil mapping has major implications for industries like agriculture and environmental science.

‍

It can help organisations:

• Reduce unnecessary sampling costs
• Improve crop planning
• Monitor environmental degradation
• Measure carbon storage
• Improve land management decisions

‍

And the efficiency gains are significant.

‍

Smarter sampling methods can reduce required sampling points by around 30% while maintaining similar levels of information quality. That is a meaningful operational improvement when working across large geographic areas.

‍

‍

The Bigger Picture

‍

This is a good example of where machine learning creates value outside traditional AI conversations. By combining spatial reasoning, optimisation algorithms and environmental data, machine learning is helping organisations understand physical environments faster and at lower cost.

‍

As climate monitoring, agriculture and land management become increasingly data driven, this type of work will likely become far more important.

‍