A MULTI-SCALE SEGMENTATION APPROACH TO FILLING GAPS IN LANDSAT ETM+ SLC-OFF IMAGES THROUGH PIXEL WEIGHTING
Monitoring changes on Earth’s surface is a difficult task commonly performed using multi-spectral remote sensing images. The absence of surface information in optical images due to the presence of cloud, low temporal resolution and sensors defects interfere in analyses. In this context, we present an approach for filling gaps in imagery mainly caused by small clouds and sensor defects. Our method consists of an adaptation from an existing method that uses spatial context of close-in-time images through the use of the most frequent value obtained using multiscale segmentation. Our method uses the pixel proportion contained in each segment to fill missing values. We applied the gap-filling methodology on three dates containing simulated images from Landsat7 using Landsat8 images. We validated the method by introducing and filling artificial gaps, and comparing the original data with model predictions. The developed approach surpassed Maxwell et al. (2007) gap-filling method for all bands, presenting a minimal R2 of 0.78. Our method proved to enhance the Maxwell et al. (2007) gap-filling method, while also asymptotically maintaining the algorithm cost. It also allowed image texture to be conserved on reconstructed images. This characteristic enables narrow features, e.g., as roads, riparian areas, and small streams capable of being detected on the filled images. Based on that, further object-based approaches can be used on images filled using this methodology, demonstrating its capacity to estimate Earth’s surface data.