Why a Fertilizer Dashboard for Kenya?

In addition to the training datasets, QED also used a combination of publicly available high-resolution satellite imagery datasets, including Sentinel-2 which is a public satellite run by the European Space Agency. The satellite data is collected every five days (since 2015) to train the models to infer patterns and correlate with crop cycles between satellite imagery and the training datasets. The results are then extrapolated across Nigeria. Using high-performance computing, QED trained and evaluated more than 50,000 different models to produce cropland maps with an accuracy and precision of 85%.  

However, even with the level of detail provided by the Sentinel satellite imagery, the main rice and maize growing seasons occur during the rainy season, which is also the period when the availability of cloud-free imagery is scarce. With transfer learning, QED leveraged modeling and data gleaned from the mapping of other countries. Having the basic model built upon previously developed models allowed the team to spend more effort fine-tuning the models to suit local conditions.⁶

An outstanding question is the ability to differentiate land under fallow from land under active cropping. Attempting to model fallow lands is difficult, as fallow lands may be hard to differentiate from natural lands. Historically, cropland maps have rarely accounted for fallow cropland but, when factored in, the inaccuracies may be substantial. To improve crop yield estimates more broadly, it is vital to estimate the proportion of fallow cropland. QED will continue to review the literature on possible fallow land types and incorporate these types into the field validation surveys used by ground teams in Nigeria.

In addition to satellite-based verification, QED has been working with Nigerian agricultural experts for further validation of the maps, drawing on the expertise of on-the-ground experts and field staff. 

From April 2020, QED began collecting ground data in North Central Nigeria. During the wet season, the region grows mostly maize and rice. QED provided field tools, including mobile phones that were pre-authenticated and pre-loaded with data collection applications; navigation software; mobile data; and communication tools. In the field, surveyors followed predefined waypoints and collected a balanced mix of locations of croplands and non-croplands (e.g., roads, natural vegetation, fallow fields, buildings, etc.).

The survey tool contained data on land use class, present and historical fertilizer use, and the number of seasons under cultivation (if classified as cropland). For data classified as croplands, surveyors were required to include a photograph of the field for verification purposes. Using digital collection tools allowed for near real-time monitoring and feedback for the team. Data from more than 1,000 locations across the North  Central region of Nigeria was collected within one month. 

Cropland Statistics

Statistics have been generated for cropland areas under production across the various geopolitical regions and states in Nigeria (see table below). These statistics include total land area, estimated cropland area, and estimated cropland percentage, which can be compared against statistics from other sources. A further breakdown on regional and state statistics can be viewed via an interactive map.⁷ 

Cropland Area Comparison Nigeria: 2020 vs. 2021

QED has updated the previous 2020 cropland map produced for Nigeria in order to reflect cropland estimates in 2021 and review any changes between the two years. To derive multi-year cropland predictions, QED employed Transfer Learning (TL), an AI technique which is used in scenarios with a large existing dataset and trained models and a smaller but related dataset from which new models are trained. Rather than collecting enough new data to build new models, the existing data and models are leveraged to develop the new models and are augmented with smaller sets of new data to provide additional data for the model to learn. In this case, QED used the cropland model developed for 2020 and adapted it to work with the dataset for 2021. Using this technique, the updated cropland map for Nigeria from the 2021 season was produced with slightly higher accuracy than from 2020.Developing the ability to map croplands across multiple years provided additional insights into national crop production dynamics. We estimate that arable land across Nigeria between 2020 and 2021 experienced a net increase of 1.34%. While numerically small as an absolute percentage (given the size of Nigeria), 1.34% equates to 12,225.19 km, roughly half of the total land area of Rwanda.These results demonstrate promise for more cost-effective updating of national-scale maps, if there is demand to regenerate them on a yearly subscription basis across multiple African countries.

Highlight: Comparisons estimate an increase of 1.34% in Nigeria’s national cropland production equal to +12,225.19 km.

We also noticed that many inter-annual differences occurred around the borders of natural areas. Many bandits (e.g., Boko Haram) use the forested areas as home bases. There appears to be consistent correlations between the regions known to have banditry and corresponding reductions in cropland.