A Sensor-Adaptive Incremental Learning Framework for Artifact Detection in Satellite Precipitation Data

Andres F. Monsalve; Hernan A. Moreno; Christian D. Kummerow

Source-linked AI summary

A Sensor-Adaptive Incremental Learning Framework for Artifact Detection in Satellite Precipitation Data

Andres F. Monsalve, Hernan A. Moreno, Christian D. Kummerow

arXiv:2609.01514v1Sep 1, 2026physics.ao-phcs.CVstat.AP

TL;DR

The paper addresses inadequate validation of rapidly expanding satellite precipitation data, where artifacts can compromise near-real-time products. It develops a sensor-adaptable patch-based framework using transfer learning and scarce human labels, and evaluates it on SSMI and SSMIS data. The framework separates regular from artifact-containing orbits and supports iterative, explainable refinement.

Problem
Rapid growth in SmallSat precipitation observations has outpaced validation, while artifacts can introduce significant errors into precipitation fields.
Method
A patch-based convolutional framework uses transfer learning and incrementally ingests human-labeled image patches to classify precipitation-field anomalies.
Results
The framework successfully identifies and classifies passive-microwave precipitation artifacts in data-scarce SSMI and SSMIS settings.
Takeaways & Limitations
The framework provides sensor-adaptable, interpretable artifact detection for near-real-time satellite precipitation quality control.
Takeaways & Limitations
Joint and Warm-Start approaches rely on geometrically compatible sensors, and transfer across divergent scanning geometries remains untested.

Abstract

from arXiv · show

Historically, retrieving rainfall data from satellite imagery has been the domain of space agencies. However, in recent years, the development of cheaper, more compact satellites (SmallSats) capable of detecting rainfall proxies has led to a significant increase in private-sector initiatives for satellite launch and surface precipitation products. This rapid growth has yet to be matched by data validation efforts. Consequently, the need for a robust tool to detect anomalies in near-real-time data before it is disseminated to the public has become critical. In this paper, we present the development of an anomaly-detection system to identify artifacts in global satellite-based rainfall products. The developed framework leverages pre-trained computer vision models and incorporates scarce human-labeled data to detect specific anomalies. Our proposed anomaly detection strategy is tested on data from the Special Sensor Microwave Imager (SSMI) and the Special Sensor Microwave Imager/Sounder (SSMIS). Results demonstrate the efficacy of our approach at separating regular orbits from artifact-containing orbits for each satellite, with performance comparable to state-of-the-art in-place methods. Additionally, the framework offers explainability and the capacity for iterative refinement following false-positive or false-negative classifications.

I. INTRODUCTION

The expansion of SmallSat precipitation observations increases the need for quality control because undetected radiometric anomalies can produce erroneous precipitation fields. The paper introduces a sensor-adaptable, patch-based framework that incrementally learns to identify artifact morphologies from scarce human-labeled data.

Motivation: SmallSat constellations lower the cost and simplify the launch of additional precipitation-observing satellites.Standardized, Lego-like assembly and non-exclusive launches support rapidly expanding satellite constellations.
Motivation: Undetected Brightness Temperature anomalies can propagate into significant errors in calculated precipitation fields.CubeSat missions often lack the rigorous quality-control oversight available to major missions such as GPM.
Artifact problem: Artifacts may exhibit physically implausible geometric patterns even when individual pixel intensities remain within realistic ranges.The paper illustrates structured, parallel, intermittent patterns in an SSMIS orbit off Argentina and Chile.
Study scope: The study validates the framework on SSMI and SSMIS data to separate regular orbits from artifact-containing orbits.The article describes the data, framework, results, discussion, and conclusions in successive chapters.
Proposed framework: The framework uses a transfer-learning-initialized deep convolutional network incrementally updated with human-labeled image patches.It spatially isolates and classifies morphological anomalies while adapting to sensor-specific distortions.

II. DATA

The study uses GPROF-derived precipitation data from SSMI and SSMIS satellites, supplemented with manually curated anomalous orbits. Orbit-level scarcity creates class imbalance, while sliding-window decomposition expands localized training examples.

Data sources: The dataset contains SSMI data from DMSP F-8 and SSMIS data from DMSP F-16, F-17, and F-18.SSMI records cover 1987–1991, while randomly sampled SSMIS orbits cover 2006–2024.
Data sources: GPROF V7 precipitation datasets were obtained from NASA GES DISC and supplemented with PPS-curated artifact cases.The curated labeled dataset includes regular and artifact classes for both instruments and is publicly deposited on Zenodo.
Data generation: SSMI and SSMIS convert measured microwave radiances into precipitation estimates using the Bayesian GPROF algorithm.GPROF compares measured Brightness Temperatures with simulated rain-profile databases.
Data representation: Each orbit is stored as an HDF5 or netCDF file containing surface rain rates with latitude and longitude coordinates.The cross-track widths are 128 pixels for SSMI and 180 pixels for SSMIS.
Data challenges: Artifact orbits are statistically infrequent relative to regular observations, producing a heavily imbalanced dataset.The study summarizes orbit counts by regular status, artifact type, and artifact combinations in Table II.
Data challenges: Sliding-window decomposition converts sparse anomalous orbits into many localized positive samples for training.A single anomalous orbit can yield tens of positive samples, with augmentation increasing spatial variation.

III. METHODOLOGY

The methodology makes artifact detection sensor-agnostic by analyzing spatial patches rather than relying on sensor-specific output details. Human-labeled examples and standardized artifact classes support automated quality control across instruments.

Patch-based framework: A patch is a sliding discrete sub-window that lets the framework operate independently of specific precipitation-product output details.This design addresses differences in passive-microwave radiometer operating conditions.
Artifact taxonomy: The image dictionary provides example artifact patches for SSMI or SSMIS orbits.Examples include lines, spots, parallel structures, coastline-following features, smooth shapes, and mosaic-like discontinuities.
Quality control: Existing automated quality controls often miss artifacts that are detected only through subsequent manual review.Human-labeled orbits therefore serve as ground truth for training and validation.

A. Data Visualization

Visualization standardizes how precipitation fields are inspected and how artifact morphology is represented as training input. Fixed map and color settings improve comparability, while overlapping windows capture localized structures.

Data Visualization: All geolocated precipitation pixels are plotted with Cartopy using a Plate Carrée projection.The same standardized colorbar and fixed value limits are applied across orbits and sensors.
Patch extraction: A 70×70-pixel sliding kernel extracts localized artifact features from each orbit.At 12.5 km pixel spacing, each window covers approximately 875 km by 875 km.
Patch extraction: The window size balances capturing artifact morphology against increasing the number of training samples.Overlapping extraction also reduces the chance of losing artifacts at window boundaries.
Negative sampling: Regular-orbit negatives are extracted on a non-overlapping 70 × 70 pixel grid and no-rain patches are undersampled.Only 20% of identified no-rain patches are retained to limit trivial empty-data bias.

C. Model Baseline Architecture Selection

The framework uses a pre-trained CNN architecture with patch-based, multi-channel inputs and incremental training procedures designed for scarce, evolving artifact data.

CNNs were preferred over Vision Transformers because their inductive bias helps reduce overfitting on smaller datasets.
MobileNetV3-Large was selected as the primary architecture to balance detection performance and computational overhead.ResNet-18 performed marginally better but required significantly longer training, while larger ResNet models performed worse.
Artifact patches were extracted with sliding windows and augmented through reflection, translation, and 90°, 180°, and 270° rotations.The extraction used 70 × 70 pixel windows with a 23-pixel stride to capture localized artifact features.
The input tensor combines globally scaled precipitation, locally normalized precipitation, and a binary landmask across three channels.Global scaling preserves absolute rainfall intensity, while local scaling emphasizes subtle artifact textures.
Training followed chronologically ordered orbit cycles to simulate the gradual discovery of new artifact categories during satellite operations.Validation data remained isolated and unaugmented, while cross-validation folds were adaptively limited by the rarest class.

H. Cross-Sensor Data and Transfer Learning Strategies

The study compares isolated target-sensor learning with cross-sensor training while evaluating artifact detection using metrics that emphasize positive-class sensitivity and calibration.

Three training paradigms were implemented to compare isolated sensor learning with models incorporating aggregate cross-sensor data.The Baseline target-only model uses MobileNetV3 pre-trained weights and trains exclusively on the target sensor.
Validation metrics were computed on hold-out orbits that were never used for training, hyperparameter tuning, or cross-validation.
F2 was selected because missing an artifact is treated as twice as important as falsely flagging a regular sample.The F-score combines precision and recall, with β=2 emphasizing recall.
Brier Score measures the mean squared difference between predicted artifact probabilities and binary outcomes, with lower values indicating better calibration.
PR-AUC provides a threshold-agnostic comparison with SPEEDe, while Balanced Accuracy addresses severe class imbalance across regular and artifact samples.Balanced Accuracy prevents performance from being inflated by unequal detection ability across artifact classes.

J. Assessment of Minimum Description Length

The minimum description length experiment measures how many labeled examples each artifact class requires for stable detection performance.

The study defines minimum description length as the training-sample threshold needed to achieve stable detection for an artifact morphology.
Independent models were trained separately for each artifact class in the SSMI and SSMIS datasets.
Training began with one positive sample and incrementally added samples until the class-specific repository was exhausted.Standard augmentation was applied to each expanded training pool.
F2 score and Balanced Accuracy were calculated iteratively on chronologically held-out validation data for the target artifact class.

IV. RESULTS

The results section fixes sensor-specific hyperparameters and reports the selected search-space values for SSMI and SSMIS experiments.

Hyperparameters were fixed across training cycles for each sensor after searching continuous learning-rate and weight-decay ranges and discrete optimizer and batch-size options.The selected values are reported in Tables V and VI.
Table V reports the SSMI hyperparameter search space and selected optimal values.
Table VI reports the SSMIS hyperparameter search space and selected optimal values.

B. Framework Performance on SSMI

Across SSMI and SSMIS, incremental performance depended strongly on when artifact classes entered training, while Joint data fusion reduced cold-start weakness. Orbit-level evaluation showed strong discrimination after sufficient target data, with calibrated probabilities improving over cycles.

SSMI Baseline F2 increased with available data, but improvement rates varied substantially across artifact classes.
F2 exceeded 0.9 for SSMI “Lines” after the first orbit containing 53 samples, with only marginal subsequent gains.
SSMI “Spots” began improving at Cycle 8 and exceeded F2 0.9 only at Cycle 14 after eight spot-only orbits.
In Cycle 1, Joint fusion achieved F2 0.33 for SSMI and F2 0.28 for SSMIS, while Baseline and Warm-Start detected no anomalies without target artifacts.
For SSMIS, Baseline achieved ROC-AUC and PR-AUC 1.000 by Cycle 2; SSMI values generally exceeded 0.9 after Cycle 2 and reached 1.000 by Cycle 13.
Brier Score decreased from 0.138 to 0.021 for SSMI and from 0.106 to 0.039 for SSMIS across cycles.

E. Spatial Distribution and Probabilistic Mapping of Artifacts

The framework combines class- and sensor-sensitive incremental learning with pixel-level probability maps for locating artifacts. Learning requirements vary with anomaly structure, density, and sensor-specific morphology.

The framework assigns artifact probabilities to each valid rain pixel, using transparency below p < 0.5 and darker colors for higher confidence.
The minimum description length is sensor- and morphology-dependent rather than universal, varying with textures and anomaly spatial density.
SSMI: SSMI “Lines” reached F2 0.919 and BA 0.972 after four samples, whereas SSMI “Spots” required approximately 14 samples to exceed F2 0.750.
SSMIS: SSMIS “Lines” required nearly 40 samples to surpass F2 0.400, illustrating the longer description length of sparser features.
SSMIS: SSMIS “Bands” reached F2 0.856 with two samples, while “Landmask” failed to exceed F2 0.300 after 27 samples.

G. Code and Data Repository

The paper provides its framework, code, trained models, and labeled data through public repositories. Its operational design emphasizes binary orbit integrity, calibrated probabilities, interpretable outputs, and refinement from uncertain or misclassified samples.

The complete framework, workflow, analysis code, and best models are available in the SAPAD GitHub repository.
The study developed and validated a supervised cross-sensor framework for identifying specific precipitation-retrieval artifacts in data-scarce settings.
High binary scores can coexist with conservative F2 scores because early models may confuse artifact types while still separating valid from anomalous orbits.
Brier Score improvement indicates increasingly representative artifact probabilities as more orbits are processed.
The framework provides artifact-type probabilities, unlike SPEEDe’s non-interpretable reconstruction-error score.
An auxiliary “IDK” category excluded ambiguous expert-labeled windows from training, prioritizing high-confidence data over conflicting signals.

B. Methodological Limitations

The framework is constrained by scarce, ambiguous labels and by assumptions about sensor geometry and training-data requirements. Its future extension to indeterminate samples and heterogeneous satellite platforms remains unresolved.

Labeling and data scarcity: A few mislabeled samples can disrupt convergence, while forcing ambiguous features into broad classes harms discriminative performance.The authors found that precise class definitions mattered more than total data volume.
Sensor geometry: Transfer-learning and joint-training paradigms rely on geometrically compatible sensors, and divergent scanning geometries remain untested.The conical-to-cross-track case is specifically identified as an untested boundary.
Heterogeneous platforms: A universal training-sample threshold for operational readiness remains impractical because visually similar artifacts require sensor-specific exposure volumes.The authors call for broader study of feature-extraction requirements across passive microwave sensors.
Indeterminate cases: Indeterminate samples at the boundary between nominal precipitation and confirmed artifacts remain a knowledge gap for current architectures.Human-in-the-loop validation is proposed as a future direction for resolving this IDK class.