Searching for Positive Deviants Among Cultivators of Rainfed Crops in Niger
By Andreas Gluecker, Project Manager, GIZ Data Lab,
with Erik Lehmann, Data Scientist, GIZ Data Lab,
and Esther Barvels, GIS and Remote Sensing Specialist, GIZ.
© PromAP / Eco-consult: Hama Bouka.
In every community, there are individuals or groups with uncommon behaviors who, while having access to similar resources, find better solutions to challenges than their peers. Finding these so-called positive deviants and promoting their solutions is referred to as the Positive Deviance approach. Building on this, the Data Powered Positive Deviance (DPPD) initiative combines traditional and non-traditional data to identify and understand positive deviance in new ways. Contributing to this initiative, the GIZ Data Lab, the UNDP Niger Accelerator Lab, GIZ PromAP Niger, and the University of Manchester are conducting a pilot in Niger to look for farmers who are particularly successful in cultivating rainfed (i.e., without additional irrigation) cereal crops pearl millet and/or sorghum. This blog post provides a brief, step-by-step presentation of the pilot, including identification and validation of potential Positive Deviants.
DPPD Niger Pilot
Sustained agriculture in Niger is under tremendous pressure, as climate change and the reduction of rainfall affect crop cycles. Produce is of lesser quality as crops grow in a drier climate, which aggravates food insecurity in Niger and other countries in the Sahel region. More than 4 million people in the Sahel region are food insecure, and 80% of lands are at risk of degradation (WFP). In Niger, 1.7 million people are estimated to become food insecure in 2021.
These developments also challenge the cultivation of pearl millet and sorghum, Niger’s principal, rainfed subsistence crops. Pearl millet and sorghum are predominantly cultivated on fragile and degraded land by small-scale family farmers in an arid climate. Farmers are affected by erratic rainfall distribution at the beginning of the rainy season, when they need to find the right timing for sowing. Their agricultural activities are further challenged by a rapidly changing agrarian context: increasing pressure on and degradation of natural resources (soil, water, and biodiversity) due to rapid population increase, conflict between livestock breeders and farmers over access to land, water, and biomass, and insecurity in several regions of the country.
Within this context, we set out to find positive deviants among farmers who cultivate sorghum and pearl millet and achieved consistently higher yields than other farmers cultivating these crops under similar conditions. In doing so, we aim to find, understand, and leverage already existing and successful local practices in rainfed agriculture, which might help design interventions tailored to different contexts. To find positive deviants, we used already existing earth-observation and administrative data. Please find the code and data used in our DPPD Niger GitHub.
Area of Interest and Unit of Analysis
In our analysis, we focused on the southern region of Niger, where sorghum and pearl millet are predominantly cultivated. We applied a two-step approach to identify potential positive deviants (PD). First, we identified positively deviant communities where yields of pearl millet and sorghum were consistently and considerably higher than those obtained in other communities within the same agricultural zone. In the second step, we manually zoomed into those communities to identify the plots and areas that contributed to the communities’ better performance. In subsequent field research, we will try to interview the farmers of those plots to learn about any uncommon but successful strategies, practices, or interventions that can be shared within their community.
We accessed data on the locations of settlements in the southern region of Niger from the Humanitarian Data Exchange (2018) and estimated the boundaries around those communities using a technique called Voronoi polygons. To only focus on communities with agricultural activity within their boundaries, we used the Copernicus Global Land Cover Layers (2019) to include only communities that contained areas with “managed or cultivated vegetation,” an indicator for agriculture. This way, we found 12,093 communities for our data analysis.
Homogeneous Grouping
When applying our analysis to the southern part of Niger, we are looking at a large geographical area, a mosaic of agricultural zones with specific structural characteristics, like temperature and soil. These zones might pose specific challenges and interact differently with farmers’ practices on the ground, possibly affecting their success. One way to account for this diversity is to group various zones within the region with similar structural characteristics that can affect performance. In this case, the characteristic is the agricultural yield of rainfed crops. We then identify positive deviants compared to other communities within the same agricultural zone. This way we reduce the likelihood that higher performance is based on structural factors as opposed to effective agricultural practices.
While we usually group these zones ourselves, for this pilot, we were lucky to find a recent classification of agricultural zones by Hauswirth et al. (2020) which divides the agricultural area of Niger into 17 zones of specific biophysical (e.g., soil properties), climatic, and land-use characteristics, as well as ecological (e.g., fauna and flora) and environmental dynamics (e.g., changes in natural vegetation). They further provide information on natural resources (e.g., access to water), demographic and social characteristics (e.g., population density), agricultural practices, and major opportunities and constraints for agricultural development in each zone.
Below: Communities were grouped according to these agricultural zones (areas in different colors).
Selection and Measurement of Performance
In applying DPPD, we are leveraging existing digital data sources as indirect measures, or proxy measures, of performance. This is particularly valuable in contexts where data are scarce or unavailable. In Niger, we used crop biomass as a proxy for the agricultural yield of sorghum and pearl millet. One common way to estimate crop biomass is to use remote-sensing-based vegetation indices (see e.g., the pilot in Indonesia). We used the Soil-Adjusted Vegetation Index (SAVI), as it is suitable for arid regions with little vegetation. To compute the SAVI, we accessed Sentinel-2 data.
Ideally, crops are harvested when their expected yield is highest. In this case, this means when biomass for sorghum and pearl millet have reached their maximum. We, therefore, used the maximum SAVI value, indicative of maximum biomass, as our measure of performance.
To increase the accuracy of our approach, we analyzed SAVI values only for areas within communities classified as “managed or cultivated vegetation” by the Copernicus Global Land Cover Layers (2019). Additionally, we calculated the SAVI values only for the rainy season from June to September, when sorghum and pearl millet can be cultivated as rainfed crops. By limiting the period to the rainy season, we reduced the probability of accidentally measuring the biomass of other crops grown on the same plot, such as cowpea. We further analyzed the rainy seasons of several years (2018–2020) to reduce the effect of single events that might have had a short-term impact on yield but that are not representative of long-term performance.
Identifying Positive Deviants
Positive deviants are those who outperform their peers under similar circumstances. Applied to this pilot, positive deviants are communities that achieve considerably higher biomass for sorghum and pearl millet (represented in higher average maximum SAVI values) than other communities in the same agricultural zone. Three steps are necessary to find those communities: (1) get an estimation of the average maximum SAVI for each community, (2) calculate the deviance (i.e., residuum) of the maximum SAVI of each community from this estimation, and (3) identify those communities whose (positive) deviance from the estimation is the strongest in its respective agricultural zone.
(1) We estimated the maximum SAVI value for the rainy seasons during 2018–2020 for each agricultural zone using predictor variables such as land cover, soil characteristics, temperature, precipitation, and soil moisture, among others. We used an ensemble of three statistical models for this estimation: a regression model to estimate maximum SAVI individually for each agricultural zone, a boosting tree algorithm to capture non-linear effects on maximum SAVI, and a neural network. The three models could explain 63% (boosting tree), 60% (neural network), and 46% (regression) of the variance in maximum SAVI, respectively. We merged these results into one joint prediction, to reduce model-specific biases and to make our estimation less vulnerable to outliers.
(2) We calculated the residuum, the difference between the maximum SAVI value in each community and the estimated maximum SAVI value of their respective agricultural zone from Step 1. We then standardized each residuum using the mean maximum SAVI value and standard deviation of the corresponding agricultural zone. Calculating the residuum allows us to identify positive deviants relative to their environment, making sure their higher performance is not due to a more favorable environment of crop cultivation. A larger, positive residuum indicates biomass that is higher than what we would have expected, while a negative residuum indicates a smaller than expected biomass.
(3) We compared the residuum of the maximum SAVI for the years 2018–2020 of all the communities across all zones to each other. Positive deviants were those communities whose residuum positively deviated from the norm, that consistently performed better than expected compared to communities in the same agricultural zone. In this case, this meant having a residuum that positively deviated at least two standard deviations from the mean residuum of all communities.
Based on this approach, we identified 180 communities that were performing particularly well, which we then validated and examined further. In the next step, we also identified the plots within those communities that seemed to have contributed to the strong performance.
Validating and Exploring Positive Deviants
Once communities were identified, we used high-resolution satellite imagery and structural variables to manually validate and explore each community with local domain experts. This way we could exclude any community that was falsely identified as a positive-deviant community. False identification can occur due to a lack of data on factors affecting yield or the lack of accuracy in the data. For instance, we found communities where pastures, bushes, and trees were inaccurately classified as agricultural plots. As bushes and trees tend to score high on SAVI, this led to some communities being falsely identified. Furthermore, as we were looking for rainfed plots, we excluded communities close to permanent and temporary water sources and urban centers, as both increase the likelihood of irrigation and the cultivation of other crops, such as irrigated legumes.
Another added benefit of exploring potential positively deviant communities in detail is the possibility of uncovering the influence of structural variables. While this is explorative, we found, for example, that higher-than-average precipitation in the early months of crop cultivation occurred repeatedly among these communities. This is in line with previous findings on the importance of precipitation in the early months of crop cultivation. Systematically analyzing how structural variables in better-performing communities deviate from those in average or poorly performing communities could help uncover more intricate relationships. This could help inform more targeted interventions, like providing resources to balance unfavorable structural characteristics (e.g., poor soil quality, lack of precipitation).
Finally, manually inspecting the better performing communities helps uncover the plots within each community that might have considerably contributed to the high performance of that community. In other words, these plots might belong to farmers who, through uncommon practices, have achieved higher yields: positive deviants. As mentioned in the beginning, it is ultimately these farmers who we hope to learn from.
We are currently finishing the validation and exploration of the communities we found and are preparing the field research. In the field research, we will try to contact and interview the farmers to learn about their practices and how they developed them within their environment. Due to security reasons, we will focus our field research on the Tahoua, Dosso, Zinder, and Maradi regions.
Moving Forward
While this approach is still experimental and results from the field are to be obtained, there are several ways in which this method could be used.
First, valuable and scalable practices that we might learn from better-performing farmers could be disseminated among other farmers in the same or similar regions. Beyond scaling practices, this method applied in Niger might also be transferable to other countries with similar practices, challenges, and environments. Most of the data we used were open and readily available remote-sensing data and might also be available for neighboring countries Mali, Chad, and Burkina Faso.
Second, this method can provide useful information for interventions and policies tailored to contexts and target groups. It can help uncover structural drivers of performance, like precipitation in the early months of crop cultivation or soil characteristics and how their influence varies across zones. It could help target specific groups, such as those performing poorly given their respective context and who might benefit most from interventions.
Third, Data Powered Positive Deviance can help evaluate and assess the impact of interventions already in place. Those individuals or groups known to have benefited from interventions should outperform others and appear, with higher likelihood, as positive deviants in the analysis. Additional field research could then explore why they responded better to the interventions, how contextual factors contributed to this, and whether certain interventions were more effective in specific contexts.
About
The Data Powered Positive Deviance Initiative (DPPD) was established on the belief that lessons on how to tackle complex sustainable development challenges are best learned from the people who face those challenges every day. It is with this mindset that the GIZ Data Lab, the University of Manchester, and the UNDP Accelerator Labs are conducting a series of pilots in different countries and domains to uncover effective, locally developed practices and innovation as a response to development challenges.
With special thanks to Andreas Pawelke, Basma Albanna, Daouda Sangare, Damien Hauswirth and Assane Boukar for their insightful comments and contributions.
Contact
GIZ Data Lab: Andreas Gluecker / Catherine Vogel: datalab@giz.de
University of Manchester: Basma Albanna:
basma.albanna@manchester.ac.uk
UNDP Niger Accelerator Lab: Assane Boukar: assane.boukar@undp.org
UNDP Accelerator Labs: Jeremy Boy: jeremy.boy@undp.org
GIZ PromAP Niger: Damien Hauswirth: damien.Hauswirth@eco-consult.com
