hal-05719413 Mapping the Biophysical Sensitivity of Agrosystems in the Central Sahel Using (…)

Climate extremes and their trends in several parts of the world have exacerbated the biophysical susceptibility of agrosystems in semiarid regions. In the Sahel region, this is reflected in a gradual deterioration in the climatic resilience of agrosystems and their ecosystem productivity, with potential losses in cereal yields of up to 27% by 2050. To meet current and projected challenges in terms of food security, resilience of rural communities and the sustainability of agricultural systems, a better understanding of the biophysical sensitivity of the spatial units that make up an agrosystem is crucial. In previous studies, the mapping of the biophysical sensitivity of spatial units was approached by using anomalies in vegetation condition indicators to quantitatively approximate the response of agrosystems to climatic water deficits. However, due to the complex interactions of factors affecting the biophysical susceptibility of agrosystems, anomalies in a single component may not be sufficient to assess biophysical sensitivity. The aim of this study is therefore to propose a new approach to mapping biophysical sensitivity by considering gradients in the magnitudes of changes in anomalies of biophysical variables in place of simple anomalies. To this end, time series data of the vegetation index, surface temperature, vegetation primary productivity and soil moisture from NOAA/AVHRR and TerraClimate products were used to assess the biophysical sensitivity by comparing the performance of four machine learning models. Methodological steps included analysis of significant trends obtained by linear regression of index using the lm() function in Rstudio. Next, models were implemented using the caret package with a data partitioning of 70% of the model training and 30% of its testing. The results suggest a spatial variability in biophysical sensitivity. The highest is obtained in croplands and lower in other land-use types and random forest model stood out for its high ability to predict biophysical sensitivity classes, with an ROC= 94%, while the ROCs of support vector machine, K-nearest neighbors, and naive Bayes were 78.6%, 72.03% and 77.6%, respectively. Finally, analysis highlights the importance of combining multi-indicator change amplitudes to obtain an overall assessment of the biophysical sensitivity of agrosystems, particularly with random forest.