Sorghum grain yield estimation based on multispectral images and neural network in tropical environments

Abstract

The application of machine learning and remote sensing techniques to estimate crop yield has garnered significant attention due to their ability to analyze large volumes of data and combine various inputs for enhanced results. This study aimed to optimize sorghum grain yield prediction in tropical conditions over two seasons by integrating vegetation indices and soil elevation data to calibrate Artificial Neural Networks (ANNs). We developed and implemented ten ANNs with a Multilayer Perceptron architecture using the Keras library, incorporating vegetation indices (CIgreen, SR, VARI, WDRVI) from the PlanetScope platform and soil elevation data from a harvester machine. The general M2 model (CIgreen + SR + VARI + WDRVI + soil elevation) achieved the highest performance with an R2 of 0.89 and RMSE of 0.22 t ha⁻¹ at 30 days after sowing (DAS). In 2019, the M9 model (CIgreen + SR + WDRVI + soil elevation) performed best at the same growth stage with an R2 of 0.82 and RMSE of 0.27 t ha⁻¹. Conversely, in 2020, the M3 model exhibited the best performance at 120 DAS, with an R2 of 0.84 and RMSE of 0.15 t ha⁻¹. These results highlight the variability in model performance due to environmental factors, plant growth dynamics, and the suitability of specific indices at different growth stages and years. Despite that the general model performed similarly to the growth stage-specific models, suggesting its potential applicability across different conditions and cultivars for estimating sorghum yield in tropical environments.

Keywords

• Artificial intelligence
• Multilayer perceptron
• Vegetation index
• Precision agriculture
• Remote sensing