Generalization of peanut yield prediction models using artificial neural networks and vegetation indices

Le résumé fourni par la source

The prediction of crop yield is vital for the management and decision-making processes in agriculture. Techniques such as Remote Sensing (RS) and Artificial Neural Networks (ANN) emerge as potential tools for predicting these agronomic parameters. Therefore, the objective of this study was to combine RS data in ANN models to remotely and anticipatively predict peanut yield. The experiment was conducted in eleven commercial fields, divided into six fields in the 2020/21 season and five in the 2021/22 season. The input data for the development of the models were vegetation indices (EVI, GNDVI, MNLI, NLI, NDVI, SAVI, and SR) derived from high-resolution satellite images on five dates, from one to thirty days before the start of the peanut harvest. The Vegetation Index (VI) data from the 20/21 season were inserted into Multilayer Perceptron (MLP) and Radial Basis Function (RBF) Artificial Neural Networks (ANNs) for the calibration. Subsequently, the generated equations were applied to the fields of the subsequent season for generalizing and recalibration of the models using the dataset from both seasons. Both networks proved capable of making predictions using the VIs as input, both in validation and recalibration, where an improvement in the precision and accuracy of the models was observed. The validation of the models demonstrated a high potential for generalizing the variability of peanut yield in new fields. The MLP network presented the best results in this study, with an MAPE of 9.3%, thirty days before harvest and a determination coefficient of 0.80. The VIs that stood out the most as input were EVI, SAVI, and SR. The use of RS combined with ANN is a powerful tool for predicting peanut yield and assisting the farmer in crop management. The results obtained highlight the importance of developing predictive models for peanut yield over the years, taking into account the interaction between genotypes and environments to enhance model robustness. Furthermore, it is essential that these models be applicable in new areas, as demonstrated by this work, which evidenced good generalization across distinct locations, even under varying management practices and cultivars.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.