Deep Interpolation of Remote Sensing Land Surface Temperature Data with Partial Convolutions.

Area of Science:

• Earth Science
• Remote Sensing
• Environmental Monitoring

Background:

• Land Surface Temperature (LST) is crucial for various applications, but remote sensing data (e.g., MODIS LST) are often obscured by clouds.
• Existing statistical interpolation methods struggle to integrate local knowledge and dependencies for accurate LST gap-filling.

Purpose of the Study:

• To develop a novel approach for interpolating remote sensing LST data by incorporating local ground-site air temperature measurements.
• To address the limitations of current methods in handling cloud-induced data gaps and improving LST data utility.

Main Methods:

• A two-step method was employed: first, learning LST from air temperature at ground stations, and second, using a U-Net deep learning architecture with partial convolutions for remaining interpolations.
• This approach integrates local meteorological data with advanced deep learning techniques for enhanced spatial and temporal LST reconstruction.

Main Results:

• The proposed method achieved a 44% improvement in Root Mean Square Error (RMSE) compared to state-of-the-art statistical interpolation techniques.
• The approach successfully provided 100% data coverage, even in areas with no valid LST measurements, by leveraging air temperature data.

Conclusions:

• The novel interpolation technique significantly enhances the accuracy and completeness of remote sensing LST data.
• This gapless, high-resolution LST data enables fuller utilization of valuable remote sensing resources for environmental studies.