Avaliação Comparativa de Modelos de Segmentação Semântica de Nuvens em Imagens de Satélite

Abstract

The presence of clouds in satellite images poses significant challenges in various fields, including meteorology, agriculture, and environmental monitoring, due to the need to distinguish clouds from other elements present in the images. Technological advance ments in remote sensing and the widespread availability of satellite images have provided valuable insights into the terrestrial environment. Accurate detection and subsequent understanding of cloud presence in satellite images are crucial not only for scientific rese arch but also for practical applications that rely on precise environmental analyses. The main objective of this work is to develop a comparative analysis between threshold-based models and deep learning-based models for semantic segmentation of clouds in satellite images, aiming to identify the most effective approach to map and understand the spatial distribution of clouds to aid, for example, in the use of techniques for removal. Clouds can obscure objects on the Earth’s surface and cause difficulties in various remote sensing applications. The choice of this work was to explore both a threshold-based approach, using a model where multiple thresholds are applied in multiple bands of the images, depending on the selected satellite, and deep learning-based models. Four deep learning models - U-Net, FPN-Net, and YOLO - were selected and analyzed, which were trained without weights, along with a model like Cloud-net with its pretrained weights, and a threshold-based model known as Fmask. Furthermore, it is important to mention that the 38-Cloud database was used for training the models. Due to the high number of images with null pixels, some modifications were proposed, including the removal of these images and the application of data augmentation techniques. Another training step in cluded the use of early stopping. Two experiments were conducted in this work: one involving the trained and evaluated models using two public datasets: 38-Cloud and 95- Cloud, employing metrics such as Intersection over Union (IoU), precision, recall, and accuracy. The other experiment involved the use of test images from 38-Cloud, which were used in the Cloud-net tests. After applying the images to the model, all images that were cropped from the original image were placed back, and the metrics were applied to the generated mask. The results indicate that deep neural network models outperfor med threshold-based techniques, highlighting the remarkable performance of Cloud-Net, which achieved an IoU of 78.50% and an accuracy of 96.48%. This analysis underscores the fundamental importance of employing deep learning models in cloud segmentation in satellite images, providing valuable and applicable insights across a wide range of areas of interest.