Evaluación de redes neuronales convolucionales para la extracción de linderos prediales como herramienta para la actualización catastral

Abstract

The training and validation process of machine learning models follows an iterative cycle composed of three main activities: input enlistment, model optimization, and results analysis. Enlistment includes data selection and preparation, while optimization focuses on the choice of architecture and the adjustment of model parameters, and finally, results analysis involves the calculation of metrics for model evaluation, where in the case of a non-optimal evaluation, an iterative optimization process must be performed to reach performance thresholds. In the study, the use of pre-trained models is proposed for the extraction of visible physical property boundaries in rural soil in Cundinamarca, for which 2 orthoimages corresponding to the municipalities of Tabio and Subachoque were chosen as inputs, which have a Ground Sample Distance less than 50 centimeters. In the case of Tabio, the roads, drains, walls and fences present in the official basic cartography of the municipality were used as elements of the boundary class, and for Subachoque, these elements were manually digitized. 145 tests were performed between the SegNet, Unet and ResNet models with tile datasets of 3 different sizes; 256x256, 512x512 and 2557x1887. Likewise, the window size, stride, batch size and number of epochs parameters were varied, in search of the optimal combination in terms of execution time and training and validation metrics. The values obtained for the evaluation metrics show that the UNET architecture achieves the best performance, as it stood out in the F1-Score and Kappa index values for both municipalities. However, as in any automated task, it is necessary to consider practical factors such as minimum machine requirements and execution time versus these parameters for the conventional capture of boundaries. Finally, the importance of the quality of the data set to guarantee the accuracy of the results is evident, therefore, improvement is suggested both in the selection and in the debugging of the input vector and raster data.