High Resolution Soil Nutrient Maps Using Machine Learning
As part of its mission, Vital Signs collects soil samples from points scattered across the countries that it works in – filling critical gaps on soil nutrients, agricultural suitability, and land degradation. These soil samples are then analyzed in the World Agroforestry Center laboratory in Nairobi to yield data on soil properties including particle size, pH, nutrient availability and nutrient content. To date, Vital Signs has collected 5,969 soil samples and has had 3,714 analyzed by the lab. Here is the breakdown by country:
|Collected Soil Samples||Lab Processed Soil Samples|
These soil samples fill critical data gaps for soils scientists and researchers worldwide. Recently, the lab-analyzed samples were used by the International Soil Reference and Information Center (ISRIC) in combination with soil samples from other initiatives like AfSIS and EthioSIS, and One Acre Fund to generate maps of soil nutrient content across the continent. These maps were created using ensemble machine learning techniques like random forests and gradient boosting using the soil samples from Vital Signs and other projects as training data. Geographic data on known soil nutrient co-variates like land cover, precipitation, lithology, and vegetation was used to predict nutrient availability at fine spatial scales across sub-Saharan Africa for organic Carbon, total (organic) Nitrogen, total Phosphorus, and extractable Phosphorous, Potassium, Calcium, Magnesium, Sulfur, Sodium, Iron, Manganese, Zinc, Copper, Aluminum and Boron. Cross validation found that all of the nutrients were predicted significantly except for Sulfur, Phosphorus and Boron.
It is anticipated that these maps will contribute to targeting agricultural investments and interventions, as well as targeting restoration efforts and estimating yield gaps. This could be especially beneficial in Uganda, where the government gives farmers seeds as part of “operation wealth creation.” This operation could benefit from information on crop suitability, to ensure these agro-enterprises are situated on soils that are favorable for specific crops. These maps can also contribute to mapping degradation hot-spots for guiding the AFRI100 forest landscape restoration initiative for which Uganda and Rwanda have made pledges.
The results have recently been accepted for publication in the journal Nutrient Cycling in Agroecosystems. Final Maps of the data outputs are available below.