photo by Malherbe Rossouw, South Africa
Soil health is our health. Here we hope to improve our tools of assessing soil health and make it widely available.
Infrared spectroscopy has traditionally provided great scientific insight. Now it can improve the quality of life of farmers and their ecosystems.
Data Collected between 2009 and 2013 by African Soil Information Service (AFSIS) partnered research centers in Africa.
Sponsored by:
Infering nutrients through Infrared Spectroscopy
Thousands of soil samples have been both scanned with (dry testing) and tested in the lab (wet testing), for a more complete soil profile. The goal is to predict the more detailed nutrient profile by using the coarser, but more efficient and affordable, infrared methods.
3 Research Centers Across Africa: CROPNUTS, ICRAF, and RRES. We worked primarily with CROPNUTS due to them having a large data set available.
Each research center performed dry testing (infrared and xray scanning) and wet testing (chemical extraction and solutions) on soil sampled from the same sights across the continent.
The best predictions came from this infrared scanning tool. It measured a broader range of spectra.
The Bruker HTS-XT
more information on Bruker HTS-XT
The infrared scanner can be configured in a variety of ways, in order to target different spectra ranges:
The data includes scannings from a variety of infrared methods. Different tools are better at perceiving specific ranges in the infrared spectrum. Running a predictive model on the various tools helped decipher which particles it was better at perceiving.
For soil nutrient molecules near the soil surface, all the infrared tools performed decently. For aspects of the soil harder to depict physically, such as soil eletrical capacitance due to insoluable salt buildup, certain tools perform better than others.
A NIH research paper considers this more closely:
The same data transformed by taking the derivative of the numbers (helps the signal appear through the noise, though the usefulessness is debatable and should be crossvalidated with different transformations on different soil samples)
These are the values that we're trying to predict. The available values vary from lab to lab. For this project, we focused on common necessary macronutrients in the soil.
The same data transformed by taking the natural log (this step helps the computer keep track of tiny, infinitesimal numbers)
Considering geodata in the model. Attempting transfer learning on neural networks using data from other continents.
Best ecological practices of agriculture can be measured and their efficacy can be proven. We can use a cheap and quick method of soil health assessment such as a infrared scanning, made widespread in smartphones.
What quality of infrared scannings could a smartphone camera provide? Another next step is to find such examples and 'mask' the available infrared scannings we're working with to resemble that quality and spectral range.
My hope is this project begins what could become a predictive model robust enough to inform cruder handheld infrared scannings. Such a model would consider global and local datasets, and determine weights based on that location.
Such a device/application could democratize soil health and demystify the process of knowing your soil's nutrient profile.
Another horizon of interest would be considering ways to easily assess soil microbiology; the living side of the soil health equation.