Know more

Exploiting multi-year high-resolution Sentinel image timeseries for mapping fallow practice in West Africa

I am an Agronomic engineer, graduated by the Public University of Navarre (UPNA) at Pamplona, Spain. During my training, I’ve done several exchange programs and got known more about remote sensing field, as well as crop modelling. There are many reasons for  starting a PhD, but in my case, the main one was the possibility to learn more about a field I’m interested in. My current work intends to develop a strategy for automated supervised land use mapping —with main focus on fallow practices— employing multi-year temporal series satellite imagery and machine learning/deep learning techniques.

Remote sensing is a field in which I’ve  always been interested; development of new technologies has made it a useful tool in many disciplines, particularly in Agriculture. This PhD offered me the perfect opportunity to learn more about multispectral image analysis and machine learning techniques. My past experiences on remote sensing combined with my agronomic training and my unique set of skills have made me be one step closer to innovation.

• Starting date: 1st December 2020
• University : AgroParisTech, Montpellier
• PhD school: GAIA
• Scientific field: Sciences de la Terre et de l’eau
• Thesis management: Agnès Bégué, UMR Tetis, Cirad
• Thesis supervisors: Raffaele Gaetano, UMR Tetis, Cirad - Louise Leroux, UPR Aida, Cirad
• Funding: #DigitAg – Cirad
• #DigitAg : Axe 5 : Fouille de données, analyse de données, extraction de connaissances, Challenge 8 : Développement agricole au Sud, Axe 6 : Modélisation et simulation (systèmes de production agricole), Challenge 1 : Le challenge agroécologique, Challenge 4 : Des productions animales durables, Challenge 6 : La gestion des territoires agricoles

Keywords: Tropical farming systems, fallows, land use, remote sensing, Sentinel, radar and optical imagery, multi-year time series

Abstract: Fallow mapping in West Africa is essential to accurately assess agricultural systems and
its contribution to food security and agro-ecological sustainability of current practices,
and yet the available mapping methodologies are not adapted to the environmental and
cropping conditions encountered when addressing tropical smallholder agriculture. In this
doctoral thesis, we explore different mapping strategies based on supervised classification
techniques and making use of Sentinel-2 imagery and rainfall data as input, as well as
multiple years of in-situ data to map fallow land at local scale in a Soudanian site in
Burkina Faso (Koumbia) between the years 2016 and 2021. Results show that ”tradi-
tional” machine learning based mapping approaches are not sufficient for detecting fallow
land under the given pedoclimatic conditions, resulting in very low accuracy figures (e.g.,
F1-scores below the 0.2 mark). Most promising results were obtained when following a
trajectory analysis approach, where a series of methodological adaptations had to be done
to exploit annual data in a multi-year oriented manner. In this last case we reformulate
the mapping problem to target non-active agricultural land (NAAL) as whole, obtain-
ing F1-score ranging from 0.75 to 0.92 values when validating against complete (no data
gaps) reference data set. Our results show that strategies that incorporate multiple years
of spectral data in their learning process as a potential viable approach, where fallow
land is not described by current status of land surface (i.e. land cover) but rather by the
changes of it along the period that encircles the moment in which crop inactivity begins.
However, results also indicate that the spatial application scope might be limited, with
an augmentation of model uncertainty in areas where no ground truth data is available,
highlighting the need to incorporate unsupervised approaches for enhanced extrapolation.
On the other hand, more explicit multi-year strategies, where temporal analysis is dele-
gated to model classifiers yielded marginally better results than annual direct mapping
strategies, yet performances obtained do not reach satisfying results, with top average
F1-score reaching the 0.44 mark. Methodological development is still required for both
(a) exploiting more efficiently and direct manner multi-year data, and (b) building more
cost-efficient unsupervised solutions that could be tested in areas with a reduce amount
of ground truth data.

Jury compound

• Bernard Tychon (Rapporteur), Universitè de Liège
• Laurence Hubert-moy (Rapporteur), Universitè Rennes 2
• Thierry Bonaudo (Examinateur), Agroparistech
• Inbal Becker (Examinateur), University of Maryland
• Martin Brandt, University of Copenhagen
• Raffaele Gaetano (Co-encadrant), Cirad

Papers & Communications

Papers in international journals

Castro Alvarado, E., Bégué, A., Leroux, L., & Gaetano, R. (2023). A multi-year land use trajectory strategy for non-active agricultural land mapping in sub-humid West Africa. International Journal of Applied Earth Observation and Geoinformation, 122, 103398. https://doi.org/10.1016/j.jag.2023.103398

Acts of conference

Castro Alvarado, E. (2022, May 23). Fallow ID: Characterization and mapping of fallow fields in West-Africa study case using Sentinel-2 [Oral presentation]. Living Planet Symposium 2022, Bonn, Germany. https://earth.esa.int/living-planet-symposium-2022-presentations/

Contact : nzcstr [AT] gmail.com- Tél: 07.54.19.65.75