Using remote sensing to assess the effect of trees on millet yield in complex parklands of Central Senegal

Highlights

• Millet yield in parkland systems were estimated with high resolution remote sensing images.

• Satellite-based information on parkland structure improves millet yield estimations.

• Yield spatial variability is mainly driven by soil nutrients.

• Trees benefit to crop for a landscape woody cover of ~30–40%.

• Advances in earth observation open new avenues for parkland systems monitoring.

Abstract

Agroforestry is pointed out by the Intergovernmental Panel on Climate Change report as a key option to respond to climate change and land degradation while simultaneously improving global food security (IPCC, 2019). Faidherbia albida parklands are widespread in Sub-Saharan Africa and provide several ecosystem services to populations, notably an increase in crop productivity. While remote sensing has been proven useful for crop yield assessment in smallholder farming system, it has so far ignored the woody component. We propose an original approach combining remote sensing, landscape ecology and statistical modelling to i) improve the accuracy of millet yield prediction in parklands and ii) identify the main drivers of millet yield spatial variation. The parkland of Central Senegal was chosen as a case study. Firstly, we calibrated a remote sensing-based linear model that accounted for vegetation productivity and tree density to predict millet yield. Integrating parkland structure improved the accuracy of yield estimation. The best model based on a combination of Green Difference Vegetation Index and number of trees in the field explained 70% of observed yield variability (relative Root Mean Squared Error (RRMSE) of 28%). The best model based solely on vegetation productivity (no information on parkland structure) explained only 46% of the observed variability (RRMSE = 34%). Secondly we investigated the drivers of the spatial variability in estimated yield using Gradient Boosting Machine algorithm (GBM) and biophysical and management factors derived from geospatial data. The GBM model explained 81% of yield spatial variability. Predominant drivers were soil nutrient availability (i.e. soil total nitrogen and total phosphorous) and woody cover in the surrounding landscape of fields. Our results show that millet yield increases with woody cover in the surrounding landscape of fields up to a woody cover of 35%. These findings have to be strengthened by testing the approach in more diversified and/or denser parklands. Our study illustrates that recent advances in earth observations open up new avenues to improve the monitoring of parkland systems in smallholder context.

Introduction

Scientific and political spheres agree on the need to foster the inclusion or upholding of trees in agricultural systems, in order to tackle the social and environmental dimensions of the sustainable development goals (SDGs, United Nations, 2016). Agroforestry, i.e. the combination of trees and crops or pastures on the same piece of land (Nair, 1993) has been acknowledged as an option to respond to climate change and land degradation (IPCC, 2019).

In sub-saharan Africa, around 40% of people in rural areas live in landscapes with more than 10% tree cover, often agroforestry systems (Zomer et al., 2014). In semi-arid West Africa, traditional parklands are characterized by the deliberated retention of trees on agricultural land (Boffa, 1999) due to the socio-ecosystem services they provide (Sinare and Gordon, 2015). Parklands contribute to the conservation of natural resources and biodiversity, and improve soil fertility and agricultural productivity (Baudron et al., 2019; Bayala et al., 2014; Duriaux Chavarría et al., 2018; Peltier, 1996). Trees compete with crops for resources but they can improve nutrient cycling, soil moisture retention and microclimate (e.g. Kho et al., 2001; Sida et al., 2018).

Studies on the impact of tree on crop productivity were generally conducted at tree-scale where crop performance under tree crown was compared with crop performance in a control area without tree influence (Bayala et al., 2015). Tree density in West African parklands is often very high and some tree species can influence crops beyond their crown (sometimes more than 100 m²/tree, Sileshi, 2016). Finding a control area without tree influence can thus be challenging, which can bias the quantification of trees influence on crops. In addition, parklands are composed of combinations of tree species with different densities and spatial arrangement. Synergies or antagonisms occur between trees and trees effect on crop performance is not likely to be additive. The direction and magnitude of the impact of trees on crop productivity depends on the dominant tree and crop species, and management practices. For instance, nitrogen-fixing Faidherbia albida, was found to improve millet and wheat yield (Bayala et al., 2012; Kho et al., 2001; Louppe et al., 1996; Sida et al., 2018) but not groundnut yield (Louppe et al., 1996). In Burkina-Faso, millet performed better under Adansonia digitata than Parkia biglobosa, the latter being a shading-tree (Sanou et al., 2012). The presence of Grevillea robusta in maize and wheat fields decreased fertilizer use efficiency while the presence of F.albida improved it (Sida et al., 2019).

Though parklands have been the focus of researches for several decades, few studies have tackled the question of the landscape-scale effect of parklands on crop productivity. Research in Ethiopia on the effects of F.albida on barley yields according to different land use systems (Hadgu et al., 2009), and agricultural productivity along a forest-agriculture gradient (Baudron et al., 2019; Duriaux Chavarría et al., 2018) are rare example.

The inter-connection of social, environmental and economic challenges as committed by the SDG calls for systemic and integrated approaches in which landscape scale is particularly appropriate to inform decision making (Reed et al., 2016). Remote sensing provides physical measurements of temporal and spatial development of agroforestry systems (e.g. structure, biomass). It could help account for tree-crop interactions and the resulting impacts on crop productivity. Current statistical models establish relationships between remote sensing vegetation productivity indices and in-situ yields measurements or national agricultural statistics. Until recently, crop growth monitoring and crop yield mapping in smallholder agriculture have relied mainly on low spatial resolution images covering large areas (Leroux et al., 2016, Leroux et al., 2019; Maselli et al., 2000; Mkhabela et al., 2005; Rasmussen, 1992). However, in agroforestry parklands across sub-Saharan Africa, accurate estimates of crop yields are hampered by landscape fragmentation, fields being often smaller than one hectare (Fritz et al., 2015). Diversity in soil conditions, crop management and tree conservation practices further amplifies inter and intra-field yield variability. New satellite or low-cost nanosatellite sensors with high spatial resolution (≤ 10 m) and high revisit frequency (< 2 weeks) are more suited to these complex and spatially variable agricultural systems. These new sensors open unprecedented opportunities to predict and map crop yield in smallholder context. A promising crop yield mapping at field level have been obtained for East and West African farming systems using Sentinel-2, Sentinel-1 and PlanetScope data (Burke and Lobell, 2017; Jin et al., 2017, Jin et al., 2019; Lambert et al., 2018). However these studies masked out trees to capture ‘pure cropped pixels’ (Lambert et al., 2018) and thus masked-out the crop below tree crown and neglected the influence of the tree on crops beyond its crown projection. Though promising, these approaches have usually failed to fully reproduce the wide variability in observed crop yield in farmer fields in sub-Saharan Africa (e.g. Jin et al., 2019; Lobell et al., 2019).

Combining information on vegetation productivity and parkland structure derived from high spatial resolution, satellite images offers the opportunity to capture the variability in crop yield in parkland systems and to identify where and how crop productivity could be improved. Remote sensing have been extensively used to identify and analyzed yield gap (i.e. the difference between observed actual yields and water-limited yields) (e.g. Jain et al., 2017; Jin et al., 2019; Löw et al., 2017; Zhao et al., 2015). In Kenya, Jin et al. (2019) explained more than 70% of maize yield variability by edaphic drivers using remote sensing, crop process-based modelling and machine learning. In parkland systems, analyzing drivers of yield spatial variability could help assess relevant opportunities to optimize parkland management.

The main aim of this study was to assess the role of trees in explaining spatial variations in millet yield in a case-study agroforestry parkland dominated by Faidherbia albida, in the Groundnut Basin of Senegal. To do so, we used high spatio-temporal resolution images (Sentinel-2, PlanetScope and RapidEye) and ground-observations. More specifically, we addressed three questions: (i) Does information on parkland structure (i.e. number of trees per field, tree density, and percentage of tree cover) help improve the accuracy of millet yield prediction in parklands of central Senegal? (ii) What are the main drivers of the predicted spatial variability in millet yield?, and (iii) What is the relative influence of trees compared with the other identified drivers?

We thus propose an original approach combining remote sensing, field data and statistical modelling. This approach was tested for an agroforestry parkland dominated by Faidherbia albida, in the Groundnut Basin of Senegal.