Characterization of establishment and forage mass prediction of Cynodon spp. pastures through remote sensing

Author(s)

Rigles Maia Coelho

• Document type: Master's Dissertation
• Press: Piracicaba.
• Institution: Universidade de São Paulo (USP). Escola Superior de Agricultura Luiz de Queiroz (ESALA/BC)
• Defense date: 2024-07-18

Examining board members

Carlos Guilherme Silveira Pedreira; Peterson Ricardo Fiorio

• Advisor: Carlos Guilherme Silveira Pedreira

Abstract

Grasses of the genus Cynodon spp. can be utilized for haymaking or grazing and exhibit high quality compared to other tropical grasses. However, most cultivars are propagated vegetatively, increasing the operational cost of their implementation. During the establishment phase, propagules undergo various stresses, including water availability and competition with invasive species during canopy formation. For established pastures, monitoring forage mass is essential for forage planning and proper adjustments in animal stocking rates. Based on this, the objective of this dissertation was to propose a machine learning-based method for monitoring the establishment of Cynodon spp. and to enhance the method for estimating forage accumulation based on Beer\'s law and radiation use efficiency using time series data. For monitoring establishment, we developed and tested a supervised classification tool based on support vector machine (SVM) and random forest (RF). These algorithms were trained with the classes \"Cynodon\", \"Exposed soil\", and \"Invasives\". As a cross-check, we used a visual scoring method for the three aforementioned classes. The correlation between both machine learning methods and the visual method was above 90% for all classes. The establishment curves followed the same trend in both methods. The main difference between the supervised classification methods was in processing time and model generation, with RF being 60% more efficient. The Jiggs cultivar was more efficient in covering the soil during establishment compared to Tifton 85. For predicting forage mass and accumulation rate, we adapted a model based on Beer\'s law. The adaptation involves estimating the leaf area index via remote sensing, with this value predicted using a time series parameterized with seven years of data. The model was implemented to predict forage mass 28, 59, and 91 days ahead. The prediction error was close to 691 kg.ha-1. The results indicate the efficiency of both proposals for the genus in question. (AU)

• FAPESP's process: 22/10653-0 - Implementation of algorithms based on digital image processing and support vector machine to characterize the establishment of Tifton 85 and Jiggs grass pastures from RPAS images
• Grantee: Rigles Maia Coelho
• Support Opportunities: Scholarships in Brazil - Master