Optimal models under multiple resource types for Brix content prediction in sugarcane fields using machine learning

Abstract

In this paper, multispectral imageries were combined with estimated crop height, age of sugarcane, and weather data (rain event) to predict Brix content in sugarcane fields across multiple sugarcane fields grown with different crop management techniques and under tropical climate over four consecutive growing seasons (2017–2021). Four machine learning regressors were used to train Brix predictive models: (i) lasso, (ii) support vector machine (SVM), (iii) decision tree, and (iv) gradient boosting regression. We compared the model performance in terms of accuracy and practicability based on five different resource types, which used a combination of 10 vegetation indices: normalized difference vegetation index (NDVI), ratio vegetation index (RVI), chlorophyll index green (CIgreen), chlorophyll index Rededge (Cirededge), photosynthetic vigor ratio (PVR), plant pigment ratio (PPR), Red edge NDVI (NDRE), nitrogen reflectance index (NRI), green leaf index (GLI), and the simple inverted reflectance green; and combinations of 10 vegetation indices with estimated crop height, sugarcane age, and rain event information. Overall, gradient boosting regression gave the highest accurate models, followed by SVM, decision tree, and lasso regression. The highest accuracies obtained by models from each resource type are R2 0.54 (RMSEP 3.28 oBrix, RPD 1.5) for a resource that used only a combination of vegetation indices (CIrededge, GLI, PPR, and simple inverted reflectance green), R2 0.67 (RMSEP 2.87 oBrix, RPD 1.7) for a resource that used a combination of CIrededge, GLI, PPR, and simple inverted reflectance green with estimated sugarcane height data, R2 0.66 (RMSEP 2.97 oBrix, RPD 1.7) for a resource that used a combination of CIrededge, PPR, and GLI with the age of sugarcane data, R2 0.71 (RMSEP 2.81 oBrix, RPD 1.8) for a resource that used a combination of CIrededge, NRI, PPR, and simple inverted reflectance green with rain event data, and R2 0.75 (RMSEP 2.81 oBrix, RPD 1.8) for a resource that used a combination of CIrededge, PPR, GLI, and simple inverted reflectance green with estimated sugarcane height and rain event data. The model composed of vegetation indices, estimated sugarcane height, and rain event data is suitable for quantitative prediction, while the other four models are appropriate for screening prediction.