Digging deep from above: satellite-guided investigation of soil fertility and crop yield response
Low quality soils are the source of multiple problems in global smallholder agriculture. A lack of plant essential nutrients — such as nitrogen, phosphorous, potassium, boron, and zinc — in the soil directly leads to plant deficiencies and reduced size and nutrient content of the food grains. Additionally, degraded soils often limit the effectiveness of fertilizers, as the applied nutrients become unavailable to the crop due to adverse conditions in the soil. Ineffective fertilizers become unprofitable for smallholders, discouraging their use and thus preventing the replenishment of soil nutrients during annual cultivating. This cycle of soil degradation and low fertilization drives low food production and poverty in developing agricultural systems across the world. In a recent paper, we show how satellite data can help to break this cycle, by allowing one to rapidly assess which nutrients are most limiting in a region.
Despite numerous greenhouse and field trials demonstrating the impact of soils on crop yields and fertilizer response, the adoption of soil fertility management and soil-specific fertilization regimes is low. In many cases farmers are unable to obtain accurate and timely measurements of their own soils, or do not have access to fertilizer blends that work well for their soils. Conducting local field measurements and fertilizer trials to provide this information for a given subregion is expensive, time consuming, and infeasible to complete for the almost 500 million smallholder farmer households in the world. However, recent advances in (1) big data quality and availability related to the soil-agricultural system and (2) geospatial and statistical inference techniques may offer a solution.
For example, several interest groups like ISRIC and Africa Soil Information Services have aggregated soil data from historic national surveys and public research and performed spatial interpolation to create and share gridded soil data products for many regions. Additionally, high-resolution (10–30 meter) multispectral satellite data, along with the programming tools and computing power to work with it, is now widely accessible thanks to cloud-based platforms such as Google Earth Engine.
In the paper, we leverage this new generation of data and techniques to investigate soil and wheat yields in Nepal. This work is done in collaboration with the Nepal branch of The International Maize and Wheat Improvement Center (CIMMYT) and their Nepal Seed and Fertilizer Project (NSAF), which aims to introduce integrated soil fertility management and teach best practices for crop and soil management to Nepali farmers.
Members of CIMMYT collected a dataset of over 10,000 government soil samples in Nepal, to which they applied machine learning techniques to predict soil properties from satellite-measured topographical and climatological features. This model was used to generate 250-meter resolution grids of soil properties for all of Nepal’s agricultural regions. We use both the point and gridded soil data in our analyses, as well as survey data on typical management and fertilization practices in the region from approximately 1,700 households. Crop yields were predicted at 10-meter resolution using satellite measures from throughout the growing season (see paper for more details).
We first look at our estimated crop yields as a non-linear function of soil parameters. Figure 1 shows the partial dependence of changes in yield on each of the measured soil properties. We find a significant positive response of yield to increasing organic matter content up to approximately 2.2%, and a significant yield optimum at 0.67 ppm zinc with yields decreasing on either side of this value. Overall, yields can be increased by between 1 to 2.5% of median yield by raising organic matter and zinc from suboptimal to optimal conditions. We did not find any significant response to other soil properties in this region.
The analytically determined thresholds we find in the previous figure can be applied to our soil grids to map soil deficiencies across Nepal. In Figure 2 we map croplands deficient in organic matter (OM < 2.2%) and/or zinc (Zn < 0.66 ppm) using 250-meter soil grids. We find that about 70% of croplands in Nepal are experiencing at least one of these deficiencies. By revealing the spatial distribution of these deficiencies, we can begin to inform management interventions targeted at alleviating specific soil limitations in each locality.
Finally, we assess the performance of two different nutrient fertilizers across soils of differential quality. To do so we developed a soil quality index (SQI), which rates a given soil profile from 0 (poor) to 1 (good) in quality. We then stratified the data by SQI and analyzed the response of yield to fertilizers in each stratum using linear regression analysis (Figure 3).
We find a stronger response of yield to nitrogen fertilizer in low SQI soils compared to high SQI soils, and the reverse for zinc fertilizer. We suspect that due to the correlation of nitrogen content to organic matter content and the driving influence of organic matter in soil quality, the low SQI soils are nitrogen-deficient and the high SQI soils are less so. Thus, the addition of nitrogen is more beneficial in low SQI scenarios, while the addition of zinc makes little impact if nitrogen is still limiting.
The synthesis of satellite data with large ground datasets on soil and fertilizers has the potential to overcome the practical limitations of field trials and unveil soil-fertilizer-yield relationships at unprecedented scales. We’ve demonstrated the ability to look at yield responses to individual soil features, to map soil deficiencies with high accuracy, and to compare the effectiveness of fertilizers in varying soil quality. These techniques can easily be applied to other countries and regions, and the results motivate further work in the dissemination of soil knowledge and locally tailored fertilizer blends to smallholder farmers and precision monitoring of related agricultural and food security outcomes.
About the author:
Jake Campolo is a Ph.D. candidate at Stanford’s Department of Earth
Systems Science.
