de Mattos Teixeira Carlos André, Alves de Araujo Thabatta Moreira, Cardoso Evelin, Costantin Filho Marcos Antonio, Weyl Costa João, Lisboa Frances Carlos Renato
Instituto de Tecnologia, Universidade Federal do Pará, Belém, Pará, Brazil.
Departamento de Computação, Centro Federal de Educação Tecnológica de Minas Gerais, Divinópolis, Minas Gerais, Brazil.
PeerJ Comput Sci. 2025 Jun 16;11:e2917. doi: 10.7717/peerj-cs.2917. eCollection 2025.
Visual inspection is essential to ensure the stability of earth-rock dams. Periodic visual assessment of this type of structure through vegetation cover analysis is an effective monitoring method. Recently, multispectral remote sensing data and machine learning techniques have been applied to develop methodologies that enable automatic vegetation analysis and anomaly detection based on computer vision. As a first step toward this automation, this study introduces a methodology for land cover segmentation of earth-rock embankment dam structures within the Belo Monte Hydroelectric Complex, located in the state of Pará, northern Brazil. Random forest (RF) ensemble models were trained on manually annotated data captured by a multispectral sensor embedded in an uncrewed aerial vehicle (UAV). The main objectives of this study are to assess the classification performance of the algorithm in segmenting earth-rock dams and the contribution of non-visible band reflectance data to the overall model performance. A comprehensive feature engineering and ranking approach is presented to select the most descriptive features that represent the four dataset classes. Model performance was assessed using classical performance metrics derived from the confusion matrix, such as accuracy, Kappa coefficient, precision, recall, F1-score, and intersection over union (IoU). The final RF model achieved 90.9% mean IoU for binary segmentation and 91.1% mean IoU for multiclass segmentation. Post-processing techniques were applied to refine the predicted masks, enhancing the mean IoU to 93.2% and 91.9%, respectively. The flexible methodology presented in this work can be applied to different scenarios when treated as a framework for pixel-wise land cover classification, serving as a crucial step toward automating visual inspection processes. The implementation of automated monitoring solutions improves the visual inspection process and mitigates the catastrophic consequences resulting from dam failures.
目视检查对于确保土石坝的稳定性至关重要。通过植被覆盖分析对这类结构进行定期目视评估是一种有效的监测方法。近年来，多光谱遥感数据和机器学习技术已被用于开发基于计算机视觉的自动植被分析和异常检测方法。作为实现这种自动化的第一步，本研究介绍了一种对位于巴西北部帕拉州的贝洛蒙特水电综合体中的土石坝结构进行土地覆盖分割的方法。随机森林（RF）集成模型是在由无人驾驶飞机（UAV）搭载的多光谱传感器捕获的手动标注数据上进行训练的。本研究的主要目标是评估该算法在分割土石坝方面的分类性能以及不可见波段反射率数据对整体模型性能的贡献。提出了一种综合的特征工程和排序方法，以选择代表四个数据集类别的最具描述性的特征。使用从混淆矩阵得出的经典性能指标评估模型性能，如准确率、卡帕系数、精确率、召回率、F1分数和交并比（IoU）。最终的RF模型在二值分割中平均IoU达到90.9%，在多类分割中平均IoU达到91.1%。应用后处理技术对预测掩码进行细化，分别将平均IoU提高到93.2%和91.9%。当作为逐像素土地覆盖分类的框架时，本工作中提出的灵活方法可应用于不同场景，是实现目视检查过程自动化的关键一步。自动化监测解决方案的实施改善了目视检查过程，并减轻了大坝故障导致的灾难性后果。
