Prediction of pasture yield using machine learning-based optical sensing: a systematic review

Abstract

Accurate and reliable predictions of biomass yield are important for decision-making in pasture management including fertilization, pest control, irrigation, grazing, and mowing. The possibilities for monitoring pasture growth and developing prediction models have greatly been expanded by advances in machine learning (ML) using optical sensing data. To facilitate the development of prediction models, an understanding of how ML techniques affect performance is needed. Therefore, this review examines the adoption of ML-based optical sensing for predicting the biomass yield of managed grasslands. We carried out a systematic search for English-language journal articles published between 2015-01-01 and 2022-10-26. Three coders screened 593 unique records of which 91 were forwarded to the full-text assessment. Forty-three studies were eligible for inclusion. We determined the adoption of techniques for collecting input data, preprocessing, and training prediction models, and evaluating their performance. The results show (1) a broad array of vegetation indices and spectral bands obtained from various optical sensors, (2) an emphasis focus on feature selection to cope with high-dimensional sensor data, (3) a low reporting rate of unitless performance metrics other than R2, (4) higher variability of R2 for models trained on sensor data of larger distance from the pasture sward, and (5) the need for greater comparability of study designs and results. We submit recommendations for future research and enhanced reporting that can help reduce barriers to the integration of evidence from studies.