Large-scale yield gap estimation with multi-source remote sensing data

While cropland expansion was for a long time considered as the main strategy to increase food production across sub-Saharan Africa, it is now widely acknowledged that this strategy is not sustainable due to strong land constraints and environmental concerns (Zabel et al., 2019). In this context, the reduction of existing yield gaps on current available cropland, is considered as a promising pathway to meet the current and future food demand of an increasing population (Cassman and Grassini, 2020). Yield gap (YG) is defined as the difference between the potential yield (Yp) and the actual yield (Ya) observed in farmers’ field in a given context. Sub-saharan Africa is a region suffering from large YG mainly due to nutrient limitations resulting from inherently low soil fertility and low use of external inputs (Affholder et al., 2013). 

Currently the Global Yield Gap Atlas (https://www.yieldgap.org/ , GYGA) is a unique tool that provides information on the yield gaps of major crops aggregated at national, climate zone or reference weather station levels using a combination of ground data, crop model simulations and climate data at coarse spatial resolution through a bottom-up approach (Grassini et al., 2017). 

Despite the huge step forward made by GYGA, several improvements are needed to up-scale location-specific estimation of the YG at a large scale while keeping sufficient levels of details to be in line with the high heterogeneity observed in the sub-Saharan Africa smallholder context. Hence the core of the improvements of the large-scale assessment of YG will rely on the use of data that are able to cover large areas while being precise enough to catch yield variability.

The main objective of the PhD research study is to assess YG at large scale for rainfed cereals in SSA using up-to-date multi-source remote sensing and recent advances in geospatial technologies while benefiting from the analytical capabilities provided by the Google Earth Engine cloud computing platform. We propose to focus on East Africa for three reasons (1) The region experiences one of the highest level of food insecurity in the continent (FAO et al., 2020), (2) whilst also suffering from the largest yield gaps in rainfed crops (Mueller et al., 2012), (3) the existence of large dataset on crop monitoring through different initiatives (e.g. TAMASA, CGIAR Bigdata Platform).

First, the upscaling of location specific estimates of yield gaps to larger areas is done using a bottom-up approach based on technology extrapolation domains (TEDs) that are assumed to delineate zones that are relatively homogeneous in terms of crop and soil management options and hence where crops perform similarly (Andrade et al., 2019). However the TED delineation relies only on biophysical data at a coarse spatial resolution. This PhD proposes to expand the recent work conducted on agricultural landscape unit segmentation (Bellón et al., 2017; Ndao et al., 2021) to improve the TED delineation by using an object-based image analysis applied on a combination of Sentinel-2 time series with a set of recent ancillary data at high spatial resolution on biophysical and socio-economic conditions (e.g. input and output prices and farm size).

Second, a crucial issue in YG assessment is the estimation of the Ya. Ya are currently estimated using national crop-specific yield statistics available for administrative units, farm surveys or expert appraisal. These estimates are however subject to high uncertainties (Carletto et al., 2015) and or highly labor demanding to be monitored over large areas. Multi-source remote sensing has already been used successfully to estimate Ya in African smallholder landscapes from local (Leroux et al., 2020) to national scale (Jin et al., 2019). This PhD research will rely on the spectral, spatial and temporal richness and the time depth (from 2016 onwards) provided by the Sentinel constellation (Sentinel 1 and 2) to estimate Ya. The estimation of Yp will be done using a platform of spatialized multi crop process-based models. To ensure a large-scale assessment of YG, the calibration of the models and then the Ya and Yp estimation will be done for each TED.

Lastly, GYGA provides only information concerning the overall YG while YG can be disaggregated into two components (Lobell et al., 2007) : (1) a transient YG caused by inconsistent yield limiting factors (e.g. pest attacks) and (2) a persistent YG caused by consistent factors (e.g. management practices). Adding information regarding the relative contributions of persistent and transient factors to overall YG can help to provide a more realistic view of the potential to narrow current YG through improvement in management practices. The Lorenz Curve and the Gini Coefficient, a methodology derived from the Economics, will be introduced for analyzing the YG inter-annual distribution, and discriminate between persistent YG and transient YG, using output from step 1 and step 2.

Funding category: Autre financement public

CNES - Cirad

PHD title: Doctorat en Sciences Environnementales

PHD Country: France