Imagens multiespectrais para diferenciação de plantas de alface inoculadas com fungo solubilizador de fosfato sob diferentes doses de fósforo

Abstract

Advanced technologies such as remote sensing stand out in the monitoring of various agricultural crops. However, few studies have investigated the potential of low-cost multispectral sensors for the discrimination of treatments involving the application of biological inputs, especially in lettuce (Lactuca sativa) cultivation. The objective of this study was to evaluate the potential of multispectral images obtained from low-cost sensors, combined with machine learning techniques, for discriminating treatments applied to lettuce, considering different forms of Aspergillus niger inoculation and variations in phosphorus fertilization doses. The experiment was conducted in 2021, following a 6 × 3 factorial design, with six A. niger inoculation methods (no inoculation; seed treatment with conidial suspension; granular inoculation with 4 × 10² and 4 × 10⁶ conidia g⁻¹ at transplanting; and inoculation with conidial suspension at 4 × 10² and 4 × 10⁶ conidia mL⁻¹ at transplanting) and three phosphorus doses (0, 50, and 100%). During plant development, images were captured using an RGB camera and a MAPIR Survey 3W multispectral camera on days 5, 12, 19, 26, and 33 after transplanting. For treatment discrimination, mean reflectance values from the spectral bands were extracted and vegetation indices were calculated. Classification was performed using machine learning algorithms in the Weka software, considering three distinct scenarios. The results demonstrated superior performance of the Random Forest model, especially when combining attributes from both RGB and RGN sensors, achieving overall accuracies of 65.35%, 79.57%, and 94.24% in the respective scenarios. This study reinforces the potential of spectral imaging for agricultural crop monitoring and management.