Comparison of Machine Learning Techniques in Cotton Yield Prediction Using Satellite Remote Sensing

• DOI: 10.20944/preprints202112.0138.v2
• Publication Date: 2021-12-10T08:06:10Z
• Authors: Francielle Morell...; Nayane Jaqueline ...; Danilo Tedesco; Elizabeth Haruna ...; Franciele Morlin ...; Leticia Bernabe S...; Getulio Freitas S...; Glauco Souza Rolim; Luciano Shozo Shi...; Rouverson Pereira...
• Keywords: Hyperparameter; Growing season; Multilayer perceptron; Stepwise regression; Predictive modelling; Perceptron

The use of machine learning techniques to predict yield based on remote sensing is a no-return path and studies conducted farm aim help rural producers in decision-making. Thus, commercial fields equipped with technologies Mato Grosso, Brazil, were monitored by satellite images cotton using supervised techniques. objective this research was identify how early the growing season, which vegetation indices algorithms are best at level. For that, we went through following steps: 1) We observed 398 ha (3 fields) eight (VI) calculated five dates during season. 2) Scenarios created facilitate analysis interpretation results: Scenario 1: All Data (8 5 = 40 inputs) 2: variable selected Stepwise regression (1 input). 3) In search for algorithm, hyperparameter adjustments, calibrations tests performed performances evaluated. 1 had metrics all study, Multilayer Perceptron (MLP) Random Forest (RF) showed adjusted R2 47% RMSE only 0.24 t ha-1, however, scenario predictive inputs that generated throughout season (approx. 180 days) needed, so optimized prediction tested VI each field, found among VIs, Simple Ratio (SR), driven K-Nearest Neighbor (KNN) algorithm predicts 0.26 0.28 ha-1 5.20% MAPE, anticipating low error ±143 days, important aspect requiring less computational demand generation when compared MLP RF, example, enabling its as technique helps yield, resulting time savings planning, whether marketing or crop management strategies.