Diseño de un prototipo para clasificación automática de imágenes satelitales mediante algoritmos de machine learning

Abstract

In this document, a prototype is designed and implemented that, using machine algorithms learning allows automatic land cover classification for classes predefined; Additionally, the training set is presented, this will allow, In the future, said prototype can be versioned so that new training data, new preprocessing steps, different algorithms or different combinations of hyperparameters can be taken into account at the time of training and thus avoid The performance of the model decreases over time. To achieve this result, a selection of areas is made in Colombian territory and in different times (years 2021 and 2022) so that a set of data can be obtained multitemporal and spatially distributed, these areas are chosen taking into account that have little presence of clouds (although areas that have them are chosen, so that the model learn to characterize them correctly, as well as the shadows generated). Later,14 For each image, a series of unsupervised models were trained which identify different clusters present in each image. Following this, a reclassification of the previously mentioned clusters to predefined classes. Then, for each image and with the In order to slightly reduce the amount of data due to issues associated with hardware, a stratified random sample to ensure that all classes of all images are present in the final training set. Then a model is built and trained proposed, the hyperparameters of the model are found using group validation, where each image is a group. The validation of the hyperparameterized model is carried out on images from 2023 which were not taken into account in the modeling process. Finally, the hyperparameters of the model that maximize the effectiveness of the model have been obtained. The prototype is developed, which is deployed in an open repository for free consumption.