Abstract
In recent decades, remote sensing has been shown to be useful for crop cover monitoring over smallholder agricultural landscapes, such as agroforestry parklands. However, the fraction of green vegetation cover (FCover) has received little attention. Indeed, the collection of FCover ground data representative of the within-field heterogeneity is time-consuming. Thus, this article aims to bridge this gap by proposing an original methodological framework combining FCover data derived from unmanned aerial vehicle (UAV) and Sentinel-2 (S2) images for estimating millet FCover at the landscape scale in an agroforestry parkland of the groundnut basin of Senegal during the 2021 and 2022 cropping seasons. UAV-based FCover was computed over a 3 m x 3 m grid using a thresholding approach for six dates over the cropping seasons and then used as ground observation for the upscaling of millet FCover at the landscape scale with S2 data. Various spectral vegetation indices and textural features were derived from S2, and several modeling approaches based on machine learning algorithms were benchmarked. Our results showed that the modeling approach using the full-time series in combination with a random forest algorithm was able to explain 73% (root mean square error = 12.13%) of the UAV-FCover variability after validation in 2021 and 2022. In addition, UAV images are suitable for consistent monitoring of millet FCover over heterogeneous agricultural landscapes by training S2 satellite images. To further check its robustness, this approach should be tested for different crops and practices across a variety of agricultural landscapes in sub-Saharan Africa.