Sugarcane yield estimation through remote sensing time series and phenology metrics

Abstract

Accurate agricultural yield prediction is a fundamental tool for sustainable agricultural planning and to ensure food security in regions critically affected by climate change and extreme weather events. Existing regression-based crop yield estimation approaches typically rely on a specific set of predictor variables, but have not been compared systematically. This paper demonstrates and compares the utilization and the combinatorial use of three different sets of object-based predictors for sugarcane yield estimation through the agricultural monitoring platform ag|knowledge which utilizes earth observation data of the Sentinel-2 satellites, captured between 2018 and 2019 for a study area of about 10,000 hectares in Ethopia. We compare several regression models using a range of different predictor variables, such as (i) multi-temporal variables (i.e., parcel-based vegetation index time series), (ii) time series descriptors (i.e., phenological metrics) and (iii) spatio-temporal variables. We achieve R² scores of up to 0.84 for the estimation of sugarcane yield and up to 0.82 for the estimation of sugar quantity through Random Forest regression, based on the combinatorial use of all predictor variables. Our experiments demonstrate that dimensionality-independent phenological metrics achieve good yield estimation results which could be a very useful variable set for model transfer and domain adaptation.