Coffee yield prediction using high-resolution satellite imagery and crop nutritional status in Southeast Brazil

• Published in: Remote sensing applications Vol. 33; p. 101092
• Main Authors: Zanella, Marco Antonio, Nogueira Martins, Rodrigo, Moreira da Silva, Fábio, Carvalho, Luis Carlos Cirilo, de Carvalho Alves, Marcelo, Fim Rosas, Jorge Tadeu
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.01.2024
• Subjects: Biennial yield effect; Coffee farming; Remote sensing; Vegetation index; Biennial yield effect; Vegetation index; Coffee farming; Remote sensing
• ISSN: 2352-9385; 2352-9385
• DOI: 10.1016/j.rsase.2023.101092

Timely and accurate prediction of arabica coffee yield is pivotal to support harvest planning and to increase the profitability in the coffee industry. Currently, yield estimates are made by harvesting coffee fruits in a few plants on the field, whose actual yield is then extrapolated to the entire production area. These estimates are time-consuming and considered subjective, demonstrating the need for faster and more accurate methodologies to predict the coffee yield. Therefore, this study aimed to (1) compare the effectiveness of using spectral bands and VIs versus foliar nutrient content for yield prediction, as well as to (2) evaluate the feasibility of combining these variables as input for developing more robust prediction models. For that, an experiment was set up in a 22-ha commercial field of arabica coffee located in Southern Minas Gerais State, Brazil. Data collection was performed using a georeferenced regular grid with 64 sampling points (57 × 57 m). The spectral variables were composed of five bands (RGB, redEdge, and NIR) and six VIs derived from the RapidEye satellite imagery. Images from 10 dates were acquired throughout the 2011–2012 and 2012–2013 crop seasons. The foliar nutrient analyses were composed of nitrogen, phosphorus, potassium, calcium, and magnesium. Lastly, the coffee yield (60 kg-bag ha−1) was determined by manual harvesting of the coffee fruits in the five plants within each sampled point. The average value of the five plants was considered as the yield of the sampled point. Then, prediction models based on simple and multiple linear regression were developed using as input the results of a principal component analysis, which was carried out considering the dataset from each image acquisition date. Overall, the regression models based only on the spectral data enabled the prediction of the crop yield up to 9 months before the harvest date with R2 and RMSE values ranging from 0.04 to 0.69 and from 36.45 to 18.43 bags ha−1 for both seasons. In addition, when the foliar nutrient contents were combined with the spectral bands and VIs, the accuracy was improved with R2 and RMSE varying from 0.57 to 0.72 and from 23.78 to 19.39 bags ha−1. Finally, the study demonstrated that the time-consuming methodology can be replaced or supported by alternative approaches, especially, when using high-resolution satellite imagery.