CACTU: Complemento de actualización urbana para la clasificación de imágenes de alta resolución mediante redes neuronales convolucionales.

Abstract

Throughout this document, the development of a complement that allows the identification of physical changes in non-urbanized urbanizable lots is exposed, by means of multi-temporal comparison and coverage classification between two high-resolution images on free geographic-type software such as QGIS. , which allows any user to generate a raster file that defines the probability of constructive changes in a particular area during the period between the images by loading a comparison base image and an image from a later time. , specific case of the application, those areas that went from being an undeveloped lot to presenting some type of construction or vice versa. This plugin performs image processing through supervised learning with convolutional neural networks, where a previous training phase is carried out, based on 60% of the total samples duly labeled in the defined classes (Built and not built). ), to subsequently verify it with the 40% of the remaining samples, generating a correspondence validation of the class that it predicts, with respect to the previously labeled class. Subsequently, with the model already trained, the classification of an image with a later date than the one used in the training is carried out, predicting the probability of whether a pixel corresponds to a construction or not. Finally, the plugin performs a subtraction between the two classifications, the input and the predicted one, thus generating a cover change probability raster. Once the input is generated, the user will be able to make the right decisions regarding the work plan, allowing the establishment of optimal areas and routes for gathering information in the field, focusing resources on areas where changes in real estate dynamics are detected. .