SAPAD: SSMI/SSMIS Precipitation Artifact & Regular-Orbit Patch Dataset

This dataset supports the study "A Sensor-Adaptive Incremental Learning Framework for Artifact Detection in Satellite Precipitation Data." It provides a curated, machine-learning-ready collection of surface-precipitation image patches drawn from passive-microwave (PMW) retrievals, labeled to distinguish physically implausible artifacts from regular precipitation fields. Source retrievals were produced by the Goddard Profiling Algorithm (GPROF V07, 2021v1) and obtained from the NASA Goddard Earth Sciences Data and Information Services Center (GES DISC). Two conically scanning instruments aboard the Defense Meteorological Satellite Program (DMSP) are covered: the Special Sensor Microwave Imager (SSMI, DMSP F08) and the Special Sensor Microwave Imager/Sounder (SSMIS, DMSP F16, F17, and F18). Anomalous orbits were manually identified, curated, and archived by personnel at the NASA Precipitation Processing System (PPS). Each record is a 70 × 70-pixel patch extracted from a Level-2 orbit. At the nominal 12.5 km pixel spacing of these products, a patch corresponds to a surface footprint of roughly 875 km × 875 km. Artifact patches were obtained with an overlapping sliding-window protocol (23-pixel stride, ~66% overlap) and manually audited; regular patches were extracted on a non-overlapping grid. Every patch stores the retrieved surface rain rate (mm/h), its geolocation, a binary land/ocean mask, and provenance metadata (source file, timestamp, orbit identifier). Missing/invalid pixels are encoded with a fill value (approximately −9999). Artifacts are grouped into six morphological classes defined in the accompanying paper: Lines, Spots, Bands, Landmask, Smoothness, and Mosaic. Regular patches are labeled "Good." The dataset is deliberately assembled to reflect the strong class imbalance of the real problem (artifacts are rare relative to nominal orbits), and it preserves the chronological orbit-level train/validation partition used in the study so that results can be reproduced exactly. Contents (four DataFrames, CSV): ArtifactDF_SSMI.csv: 599 artifact patches from 20 SSMI (F08) orbits, April 1988 to January 1989. Classes present: lines (312), spots (287). ArtifactDF_SSMIS.csv: 760 artifact patches from 19 SSMIS orbits (F16, F17, F18), September 2006 to May 2025. Classes present: bands (335), mosaic (158), spots (83), smooth (75), lines (75), landmask (34). RegularDF_SSMI.csv: 6,185 non-artifact ("Good") patches from 167 SSMI (F08) orbits, August 1987 – December 1991. RegularDF_SSMIS.csv: 10,148 non-artifact ("Good") patches from 193 SSMIS (F16) orbits, January 2015 – June 2017. The dataset is intended for direct ingestion into computer-vision pipelines for training and evaluating sensor-adaptive artifact-detection models, and for benchmarking anomaly-detection methods on operational PMW precipitation products. Preprocessing scripts, patch-generation code, and trained-model outputs are available in the companion software repository (see Related works).