Using UAV and geostatistics to upscale crop yield in heterogeneous agro-silvo-pastoral system

The beneficial effect of Faidherbia albida on the yield of certain associated crops has been demonstrated for long and is often characterized by a distance-decay pattern. While several approaches have tested the spatial extent assessment of tree influence, none of them has been designed either to capture the directional variations or to address the park effect at the landscape scale. Recently, Roupsard et al. (2020) proposed an approach based on multispectral (MS) UAV (Unmanned Aerial Vehicle) imagery and geostatistics to bridge this gap. In the present study, we extended their study by proposing a new application of their approach to groundnut crop and validation for millet crop. In addition, we tested improvements of the method, using several MS images along the crop cycle.

In a typical F. albida parkland (Niakhar, Senegal), groundnut and millet under-crops of agroforestry plots of approximately 1-ha and 2-ha respectively, have been harvested. On each plot, groundnut and millet traits were measured at three different positions from six F. albida trees (under crown "S"; crown edge "B" and far from the crown "H"). We found that F. albida improves the haulm yield of the groundnut crops under its crown by about 50%. However, unlike its strong effect on millet, it does not significantly affect the groundnut pod yield. Through geostatistical analysis of multi-spectral, centimetric-resolution images obtained from the UAV flights carried out during the wet season, we observed that F. albida affects the groundnut NDVI signal up to 9.8-m and the NDVI of millet up to 18-m. We found statistically significant, positive correlations between groundnut pod yield and MSAVI2, NDVI (r² = 0.73; RMSE = 9.8) first, and between groundnut haulm yield and MSAVI2 (r² = 0.85; RMSE = 5.81). For millet, the multiple linear regression model is able to explain 74% of the millet yield variability (RMSE=20.48) using millet+weed MSAVI2 and NDVI. We used the regression model to upscale groundnut pod and haulm yield maps at the whole-plot scale. Compared with groundtruth, the error was only by 8% and 13% for groundnut pod and haulm yield, respectively. Using a geostatistical proxy for the sole crop, the crop-partial Land Equivalent Ratio (LERcp) was estimated at 1.02 for pod yield and 1.05 for haulm yield.