Identificação Automática de Áreas de Cana-de-Açúcar e Árvores Isoladas Utilizando Segmentação Semântica

Abstract

Manual orthophoto segmentation has limitations such as high costs, low scalability, and susceptibility to errors. Semantic segmentation is a crucial task for agricultural monitoring but faces challenges like class imbalance.The main objective of this work is to investigate semantic segmentation methodologies for the automatic identification of sugarcane areas and the mapping of isolated trees in orthophotos.One of the most significant contributions was the comparative evaluation of state-of-the-art models, such as SegFormer and InternImage, against more traditional and widely used convolutional architectures, like U-Net and DeepLabV3+. This analysis provided a comprehensive overview of the performance of different approaches in agricultural image segmentation.Furthermore, the study delved into the impact of class imbalance, a significant technical challenge in agricultural domains. To mitigate this issue, pre-processing strategies were proposed and investigated, including Crop-Focused sampling and Boundary-Zone sampling.The research also included an analysis of the effect of patch size on segmentation performance, addressing the trade-off between capturing global context and preserving local details. This demonstrated the viability of automating essential mappings, such as identifying planting areas and isolated trees, which shows direct gains in productive efficiency and a reduction in operational costs.These contributions, collectively, validate the capability of semantic segmentation solutions for automated monitoring, offering a path toward the implementation of more efficient and sustainable agricultural practices.