On 9 December 2019 in Montpellier, Julien Sarron defended his thesis research on the spatialised estimation of mango yields in Senegal. Numerous results, publications and tools will follow, including a digital tool aimed at the mango sector, which is currently in the development-transfer phase, providing numerous openings for Julien, who has just been recruited by CIRAD to continue his research in Emile Faye’s team at HortSys.
What were your objectives when you began your thesis?
In the context of food insecurity in West Africa, improving methods for monitoring agricultural production is crucial. Fruit trees have an important role in people’s diets, but their yields are low. To address development challenges, we need to identify the causes of this. My thesis subject was applied to mango production, with the goal of producing tools for estimating mango yields at different spatial scales: the tree, the orchard and the region. In order to identify the factors underlying the variability in yields observed, I focused on 30 orchards in the Niayes region near Dakar, divided into three types of cropping systems: intensive (monoculture), extensive, and also diversified, in other words agroforestry.
Agroforestry, cropping systems, remote sensing, artificial intelligence, etc. Your research covers several areas that combine agronomy and digital tools?
Yes, I am an agronomist by training, and I had already worked on the variability of annual crop yields, but fruit trees were new to me. I had to immerse myself in research on tropical arboriculture to understand the factors explaining heterogeneity in mango yields, with questions linked to tree ecophysiology, but also to cropping systems, which are highly varied in West Africa. For example, the mango agroforestry system is complex and particularly interesting.
I also explored drone imagery and artificial intelligence (machine learning, neural networks) with a dual objective: building a land use map and estimating mango yields at the tree level, in and between orchards.
Drone imagery is interesting for research in agroforestry, in order to compare cropping systems…
At the orchard level, the land use map enables us to identify, for each orchard, the position of trees, their species, and even their variety in the case of mango trees. To develop this map, we classified drone images with eCognition software. We trained an algorithm to distinguish mango trees from other objects present in fields: other trees, such as citrus and cashew trees, vegetables, bush vegetation, soil, tracks, buildings, etc. We then modelled the production of each mango tree by cross-referencing the land use map with information on the structure of trees, in other words their height, their projected area and their volume, which was also obtained by drone imagery.
- The modelling and spatialisation of production reveal yield variabilities in and between orchards: the spatial distributions of tree characteristics and yields appear to be more heterogenous in extensive and diversified orchards than in monoculture orchards.
- Drone imagery is interesting from an agronomy/agroforestry perspective. A lot of information on orchards has been acquired: varietal composition, number of trees, area under vegetable crops, etc. A large number of measurements can be made and compared in order to estimate the impact on yield and to compare cropping systems. For example, we see that diversified agroforestry orchards have a higher average yield per tree than intensive orchards..
A drone would not be suited to the constraints of small producers, whereas a smartphone would
At the tree level, we have developed a tool to analyse digital images through model assisted neural networks to estimate mango yields prior to harvesting. No drones are used here, as they would not be suited to the constraints of small producers, whereas a smartphone, for example, would. So, for experiments, we used a high resolution digital camera to take photos of each tree and its mangoes. In order to detect the number of mangoes per tree, with the support of our colleagues from the AMAP joint research unit, we chose algorithms already used by an Australian team for monoculture orchards. Our added value is that we trained the neural network to detect three varieties of mango trees. These are the most common varieties found in the Niayes region, with highly varied crop conditions. Mangoes are hidden within trees, so I developed a statistical calibration model to link the number of visible fruits detected by the algorithm to the actual number of fruits. It is calibrated for these three varieties of mango trees, but can be used elsewhere if recalibrated. I used the yield gap concept to calculate the gap between the achievable yield of a mango tree, defined by its structure, its age and its variety, and its actual yield. In the 30 orchards monitored, for the three varieties and the three cropping systems found in the region, we noted the effects of the mango variety and the cropping system on yield gaps. We also identified the effects of certain practices. For example, irrigated mangoes have on average 30% smaller gaps between their achievable and actual yields in relation to non-irrigated mango trees. This neural network + calibration model combination is the basis of the PixFruit tool currently being developed.
At the regional level, in other words the Niayes production basin, I tackled the regional agricultural study: what are the factors that determine yield? What are the yield variabilities between orchards? I am currently analysing the results to compare orchard performances. We can already note that tree density is a key variable in determining yield in intensive systems, and that trees produce more in diversified systems. Climate, cropping practices and crop diversity are also explanatory factors of yield variability, but their effects remains to be explored. In the future, the idea would be to have a model to extrapolate yields at the regional level, in order to estimate the overall production of orchards in the basin.
And after your thesis?
I have just been recruited by Cirad, and will be able to continue my work within the same team, in the HortSys unit, this time in Montpellier. At the moment, I am writing new articles on my thesis research. For PixFruit, we are about to begin the validation of models at the orchard level, the development of a regional model and the development of the tool. I am also going to tackle new research questions, such as the temporal dynamics of fruit production (flowering, maturity, etc.). Theses aspects were discussed little in my thesis, as there was already a lot to do on mature fruits. I would also like to determine whether the methodological concept developed for mangoes can be applied to other species of fruit, for example lychees in Madagascar. To do so, several research collaborations will continue for regional remote sensing aspects, with AMAP for the supervision of internships and within the PixFruit project team (see box).
Contact: julien.sarron [AT] cirad.fr – Tél. +33 4 67 59 37 14 – See also: Homepage &publications
PixFruit: towards a decision support tool for the mango sector in Africa
PixFruit mobilises digital agriculture tools to manage data on fruit crop yields. Producers will be able to estimate yields from their trees, in real time and offline.PixFruit mobilises digital agriculture tools to manage data on fruit crop yields. Producers will be able to estimate yields from their trees, in real time and offline.
To ensure a successful transfer, a consortium brings together CIRAD (HortSys and AMAP units), the French-Moroccan start-up SOWIT (informative data, decision support tools) located in West Africa, and the Orange Labs laboratory (R&D) in Grenoble, which provides expertise in design thinking and telecommunications.
The PixFruit platform will be operational in 2021. Initially, it will inform all actors in the mango sector about mango yields for the year in Senegal and Côte d’Ivoire.
Contact: Emile Faye, Cirad HorSys, emile.faye [AT] cirad.fr