Skip to content
The Chat Archive
Unspecified NVIDIA model (HelpSteer2)· nvidia NVIDIA HelpSteer2

refine this text for a scientific paper: Sentinel-2 satellite imagery has become…

  1. User

    refine this text for a scientific paper: Sentinel-2 satellite imagery has become essential to monitor and manage natural resources at scale due to its global coverage, recurrence of acquisition and spatial resolution. However, satellite data is heavily affected by weather phenomena such as the presence of clouds, haze, and snow, which limits its applicability. This drawback is partially alleviated by the recurrence of data acquisition, that increases the probability of finding pixels free from such weather effects, hence enabling the replacement of affected areas with temporally close data. Nonetheless, applying this method at scale requires to identify affected pixels accurately and automatically. To that end, we leveraged Artificial Intelligence, Computer Vision, and month-long Sentinel-2 time series to compose cloud-free mosaics in an automatic manner. To do so, we increased spatial resolution of all 20- and 60-meters bands of each Sentinel-2 scene in a time series to 10 meters by applying the super-resolution CNN model DSen2. The semantic segmentation CNN model HRNet was then applied on super-resolved scenes to detect pixels covered by haze, snow, and clouds, as well as their shadows, generating binary invalid pixels masks. These masks were then used to calculate pixel-wise weights of each scene in the time series according to proximity to invalid pixels and the difference between scene’s acquisition date and the central date of the month. Once weights are computed, a pixel-wise weighted average of all scenes in a time series is computed for all 12 bands, resulting in a 10m resolution cloud-free Sentinel-2 mosaic. Results show that this methodology can reliably compose cloud-free mosaics given enough scenes in the time series, while preserving pixel information of temporally close scenes and avoiding artifacts that could hamper the applicability of this method.

  2. Assistant

    Sentinel-2 satellite imagery has become a crucial tool for monitoring and managing natural resources at a large scale due to its global coverage, frequent data acquisition, and high spatial resolution. However, satellite data is often affected by various weather phenomena, such as clouds, haze, and snow, which limit its applicability. The frequent data acquisition of Sentinel-2 partially mitigates this issue by increasing the probability of finding pixels free from such weather effects, enabling the replacement of affected areas with temporally close data. Nonetheless, accurately and automatically identifying affected pixels is essential for applying this method at scale.

    To address this challenge, we leveraged Artificial Intelligence, Computer Vision, and month-long Sentinel-2 time series to compose cloud-free mosaics in an automatic manner. We first increased the spatial resolution of all 20- and 60-meter bands of each Sentinel-2 scene in a time series to 10 meters by applying the super-resolution CNN model DSen2. The semantic segmentation CNN model HRNet was then applied on super-resolved scenes to detect pixels covered by haze, snow, and clouds, as well as their shadows, generating binary invalid pixels masks. These masks were then used to calculate pixel-wise weights of each scene in the time series according to proximity to invalid pixels and the difference between the scene's acquisition date and the central date of the month.

    Once the weights were computed, a pixel-wise weighted average of all scenes in a time series was computed for all 12 bands, resulting in a 10m resolution cloud-free Sentinel-2 mosaic. Results show that this methodology can reliably compose cloud-free mosaics given enough scenes in the time series, while preserving pixel information of temporally close scenes and avoiding artifacts that could hamper the applicability of this method.

    In conclusion, the proposed methodology accurately and automatically identifies affected pixels in Sentinel-2 satellite imagery and composes cloud-free mosaics, enabling more reliable and efficient monitoring and management of natural resources at scale.

Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.