Índices de vegetação obtidos por RPA e redes neurais artificiais na determinação de áreas de produção de sementes de soja (Glycine max (L.) Merrill)

Abstract

The constant search for technologies in the field is related to increased production capacity. Consequently, the use of ARPs has emerged as a significant option in precision agriculture. Furthermore, high-quality seeds are essential for the implementation of crops with high productive potential, as the seeds have high added value, carrying biotechnologies and genetic characteristics that allow them to reach the highest productive potential. Therefore, monitoring during seed production is crucial to ensure the quality of the seeds that will be produced. The general objective of this work was to monitor the seed fields through remote sensing with the use of ARPs, using vegetation indices, to verify in the seed production fields which areas should be destined to the production of better-quality seeds. The experiment was conducted in two fields, an area of commercial production of soybeans (Glycine max (L.) Merrill) in the 2022/2023 agricultural year, located in the district of Piquirivaí in the municipality of Campo Mourão, Paraná, Brazil. Samples were collected from multiple locations, categorized as high, medium, and low NDVI environments, based on an ARP flight conducted on stage R7 on March 6, 2023. The study was conducted in a 2x3 factorial randomized block design (DBC) (seed fields, environments), with 5 replications and using Artificial Neural Networks (ANNs). A statistical difference was observed when considering the fields together, showing greater germination in high NDVI environments compared to medium and low NDVI environments. When analyzing each field separately, we noticed that in Field 01 the high NDVI stands out, while medium and low NDVI there is no statistical difference. For Field 02, the high NDVI environment differs from the low NDVI environment but does not differ statistically from the medium NDVI. The Artificial Neural Networks after 10 rounds for each sample, in Field 01, the most effective training was R4, with R² of 0,9621, RMSE of 1,5916 and BIAS of 0,133. In Field 02, after 10 rounds for each sample, the most effective training was R7, with R² of 0,9654, RMSE of 1,8619 and BIAS of 0,400. Remote sensing with a multispectral sensor embedded in ARP proved to be a valuable tool for monitoring the development of soybean crops. Artificial Neural Networks (ANNs) stood out for their accuracy and precision in predictions, with robust and statistically significant correlations between Vegetation Indices and seed germination, especially in stages R4 and R7. This study confirms that ARP images are viable for selecting ideal areas in soybean seed production, providing valuable insights for future research and practical applications.