Aplicação de redes neurais convolucionais na estimativa da altura de pastagens nativas e cultivadas

Abstract

Pasture-based livestock production requires an accurate estimation of forage availability to enable precise animal management, ensuring good results in the production process and preventing issues such as overgrazing or undergrazing. The literature shows the use of vegetation height to estimate forage availability in pastures, as it correlates with the amount of dry matter available. As an advantage, height is easily measured in the field and requires minimal procedural guidance, although it demands physical labor when measured directly. This work proposes a solution for measuring vegetation height in native and cultivated pastures using computer vision techniques. Specifically, the inference system processes an image of the vegetation through convolutional neural networks, each trained for a specific vegetation type, to estimate its average height. The pasture species used in the models presented in this work are native vegetation from the southern Brazilian grasslands, ryegrass, and Sudan grass. However, the same technique applies to any other type of vegetation. While convolutional neural networks are not new, their specific application for quantifying vegetation height in pastures represents a novel approach as far as our literature review shows. The benefits of this approach include the development of a data collection application and its integration with the H-Pasture app, which provides recommendations for animal entry and removal based on the average pasture height, enabling more accurate decision-making in pasture management. Automating this process can help reduce operational costs and increase production efficiency, promoting more sustainable management of natural resources. The results obtained the following correlation coefficients and mean absolute errors: 0.8949 and 1.5236 cm for native pastures, 0.96 and 1.3791 cm for ryegrass, and 0.81 and 7.9795 cm for Sudan grass, respectively. These data highlight the applicability of the approach in pasture-based livestock production activities, especially in the management of native pastures or ryegrass.