Improving Pearl Millet Yield Estimation From UAV Imagery in the Semiarid Agroforestry System of Senegal Through Textural Indices and Reflectance Normalization

Abstract

Enhancing food security in the Sahel through nature-based solutions is urgent given population growth, resource scarcity and climate change. Traditional agroforestry parklands are a farmer- and nature-based widespread form of ecological intensification which randomly integrates trees into crop fields. While most studies estimating crop yields in agroforestry have been conducted in controlled experimental settings, few have addressed the inherent variability in such highly heterogeneous systems. Thus, the purpose of this study is to benefit from a UAV-based proxy-sensing and machine learning approach to address the variability of pearl millet grain yield, according to the distance to randomly distributed trees in a traditional agroforestry system dominated by Faidherbia albida (i.e. groundnut basin of Senegal). 21 vegetation indices (VIs), 32 normalized difference texture indices (NDTIs) derived from multispectral drone images, and normalization variables for radiative conditions were used with yield data collected in 15 plots (around 1 ha each) and subplots ( $15~m^{2} $ each) displayed at 3 distances from the tree over five cropping seasons (2018–2022). In this context, the optimal phenological stage was determined for predicting pearl millet grain yield, which proved to be the pre-heading period. This period was used as the basis for our machine learning model training dataset in the subplots. Two models, Random Forest (RF) and Gradient Boosting Machine (GBM) were compared by combining VIs, NDTIs and normalization variables. GBM was the best-performing model, explaining 78% of observed pearl millet yield variability over five years in the subplots, with a RMSE of 16 g. $m^{-2} $ . This study revealed that NDTIs calculated from red and green bands were more influential for yield estimation than those based on near-infrared. These results were subsequently used to predict yield in all plots, resulting in a mean relative error of 17.5% between yields estimated by the farmers and GBM-estimated yields. This approach represents a pathway to assessing the withinfield yield variability in highly heterogeneous agroforestry plots and to demonstrate, quantify and optimize tree benefits for ecological intensification.

Keywords

• Agroforestry
• drone
• heterogeneity
• multispectral
• machine learning
• upscaling