Predicting within‐field cotton yields using publicly available datasets and machine learning

Abstract

Early detection of within-field yield variability for high-value commodity crops, such as cotton (Gossypium spp.), offers growers potential to improve decision-making, optimize yields, and increase profits. Over recent years, publicly available datasets have become increasingly available and at a resolution where within-field yield prediction is possible. However, the viability of using these datasets with machine learning to predict within-field cotton lint yield at key growth stages are largely unknown. This study was conducted on two cotton fields, located near Mungindi, New South Wales, Australia. Three years of yield data, soil, elevation, rainfall, and Landsat imagery were collected from each field. A total of 12 models were created using: (a) two machine learning algorithms: random forest (RF) and gradient boosting machines (GBM); (b) three growth stages: squaring, flowering, and boll-fill; and(c) two different amounts of variables: all variables and the optimal variables determined by a recursive feature elimination (RFE). Results showed a strong agreement between predicted and observed yields at flowering and boll-fill when more information was available. At flowering and boll-fill, root mean square error (RMSE) ranged between 0.15 and 0.20 t ha −1 and Lin's concordance correlation coefficient (LCCC) ranged between 0.50 and 0.66, with RF providing superior results in most cases. Models created using the optimal variables determined by the RFE provided similar results compared to using all variables, allowing greater model accuracy and resolution with targeted sampling. Overall, these findings indicate significant potential of publicly available datasets to predict within-field cotton yield and guide decisionmaking in-season.

A machine learning modeling framework forTriticum turgidumsubsp.durumDesf. yield forecasting in Italy

“…In GBM, weak learners (variables) are sequentially converted to strong learners to decrease bias from highly correlated variables. In contrast, RF generates models (trees) from a random subset of the training data and averages all trees to reduce the variance (Leo et al., 2021).…”

“…The use of machine learning (ML) approaches can capture nonlinear relationships and has shown a strong ability to predict crop yield (Paudel et al., 2022a; Shendryk et al., 2021) when compared with traditional linear approaches (Filippi et al., 2019a; Shahhosseini et al., 2020). Furthermore, among the widely used algorithms, random forests (RF) and gradient boosting machines (GBM) have shown high predictive ability and improved accuracies (Leo et al., 2021). Artificial neural networks are also widely used in predictive modeling and soil mapping (Schillaci et al., 2021).…”

Assessment of hail damages in maize using remote sensing and comparison with an insurance assessment: A case study in Lombardy

“…Although this saturation may be partially contained when selecting the bands to be analysed, the best alternative is to use three-band indices (Verrelst et al, 2013(Verrelst et al, , 2015. To overcome the saturation problem, new VIs have been developed (Fadaei, 2020;Talukdar et al, 2020;Leo et al, 2021). Verrelst et al (2015) evaluated many vegetation indices generated from Sentinel-2 data and found that the best indices matched the three-band indices according to the normalised formula (ρ560-ρ1610-ρ2190) / (ρ560 + ρ1610 + ρ2190).…”

An integrated feature selection approach to high water stress yield prediction

“…Leo et al. ( Leo et al., 2021 ) employed RFE within the packing method to choose spectral indices. This led to an improved prediction accuracy for the model when assessing the correlation between predicted and observed scores.…”