The influence of sampling on deep learning for complex urban land use and land cover classification using RPA imagery:

a case study of Conjunto Paulo VI (Belo Horizonte, MG)

Abstract

This study evaluated the efficiency of deep learning for land use and land cover classification using very high spatial resolution imagery of the Conjunto Paulo VI neighborhood, in Belo Horizonte (MG), Brazil. From an orthomosaic acquired by a remotely piloted aircraft (RPA), 48% of the neighborhood area was manually vectorized, defining 15 classes, of which 70% of the samples were used for training and 30% for accuracy assessment. Several scenarios with different sampling proportions were tested, ranging from 10% to 70% of the training data. Classification was performed through semantic segmentation with the U-Net architecture (ResNet34 backbone) in ArcGIS Pro, and performance was assessed using confusion matrices and the Overall Accuracy, Kappa, F1-Score, and IoU metrics. Models with lower sampling rates showed lower performance in identifying less represented classes, whereas those trained with larger sample sizes overfitted, compromising accuracy. The model trained with 50% of the samples achieved the best performance across general and per-class metrics, considering the human effort of vectorization and the efficiency of the training epochs. These results indicate that it is possible to optimize sample labeling in urban environments, showing that an intermediate data volume can ensure high accuracy while avoiding redundancy and overfitting.