Landsat8 training data for deep learning model

This documentation introduce the image patches dataset for Landsat 8 cloud and cloud shadow maksing used in a paper in review: Hankui Zhang, Dong Luo, David Roy, A learning attention network algorithm (LANA) for accurate Landsat-8 cloud and shadow masking, Remote Sensing of Environment The dataset contains a total of 16,861 image patches each with 512×512 30m pixels. The patches were derived from (1) 27 Landsat 8 images annotated by USGS and refined by the experts from South Dakota State University (SDSU), (2) 69 SPARS image subsets each with 1000×1000 30m pixels, (3) 4 images annotated by the experts from SDSU. These images and image subsets are randomly distributed on different land cover types (Barren, Forest, Grass/Crops, Shrubland, Snow/ice, Urban, Water, and Wetland) and across different season (Spring, Summer, Fall, and Winter). Each patch consists of three patch image files: (i) 8 bands 30 m Landsat-8 top of atmosphere (TOA) reflectance patch image stored in numpy file, (ii) annotated cloud and cloud shadow mask patch image stored in GeoTiff file, and (iii) USGS Fmask derived cloud and cloud shadow mask patch image stored in GeoTiff file. They are named as hhhLC08_sssssssssssssssssssssssssssssssssssxxxx.yyyy.ddd.tif/npy example 046LC08_L1GT_215071_20130601_Y.3705.2068.CLD.tif where hhh (e.g., 046) is a three-digit internal ID used by SDSU to identify which image the patch is from, ‘sssssssssssssssssssssssssssssssssss’ (e.g., _L1GT_215071_20130601_Y.) is the Landsat image ID made by USGS, ‘xxxx’ and ‘yyyy’ are four digit numbers (e.g., 3705 and 2068) indicate the start column and row locations in the image of the patch, and ‘ddd’ is the indicator to identify one of the three patch image types: TOA = (i) 8 bands 30 m Landsat-8 top of atmosphere (TOA) reflectance patch image. *.TOA.npy: npy file (need python numpy package to open it) that has shape (8,512,512). The order of the first dimension (i.e., 8) is: Aerosols, blue, green, red, NIR, SWIR1, SWIR2, Cirrus. They store relfetance×10000 as int16. data type. CLD = (ii) annotated cloud and cloud shadow mask patch image The patch for the annotated cloud and cloud shadow mask (CLD.tif) has four unique pixel values : 128: clear 192: thin cloud 64: cloud shadow 255: cloud FMS = (iii) USGS Fmask derived cloud and cloud shadow mask patch image The patch for the annotated cloud and cloud shadow mask (FMS.tif) has three unique pixel values (thin cloud is not derived by Fmask): 128: clear 64: cloud shadow 255: cloud There is no filled pixels in the derived patches as we designed this on purpose to avoid their impact on model training. It contains "L8_training_data_readme", zip files "PATCH.L8.patch512.step256.release000to058" and "PATCH.L8.patch512.step256.release062to102".