ABSTRACT:

Operational long-range (30-day) streamflow forecasts from NOAA’s National Water Model (NWM) carry a combination of model structural error and forecast-forcing error that grows with lead time, limiting their practical value for water-resources decision-making. This paper presents a learned post-processing framework that improves NWM long-range forecast skill without retraining the underlying physics-based model. The framework adopts the Exponential Smoothing Transformer (ETSformer) as its backbone and augments it with three components: a Conditional Frequency Channel Attention (CFCA) module that incorporates static basin attributes such as dam-regulation status, Reversible Instance Normalization (RevIN) to address non-stationary distribution shift across basins, and a zero-cost Linear Scaling step that removes residual systematic bias. We train and evaluate the framework exclusively on the NWM’s archived operational long-range forecast product, retrieved from the public BigQuery archive for Sep,2018 to Nov, 025 across 20 USGS-gauged river reaches distributed across the contiguous United States. We benchmark the framework against four LSTM variants (LSTM, SAM-LSTM, MHAM-LSTM, BOA-LSTM) and two alternative Transformer architectures (Pyraformer, iTransformer). The framework reduces unbiased root-mean-square error by approximately 77% relative to the strongest LSTM baseline and maintains its skill across the full 30-day forecast window rather than degrading with lead time. Ablation experiments show that CFCA, RevIN, and Linear Scaling each contribute, and that their gains are complementary. A leave-one-basin-out generalization experiment indicates that the model transfers to unseen watersheds with only marginal loss in accuracy. The results demonstrate that a Transformer combined with explicit non-stationarity correction and static-attribute conditioning can deliver stable long-horizon bias correction for the NWM operational forecast product.

ABSTRACT:

Marine oil spill detection using Synthetic Aperture Radar (SAR) is crucial but challenged by dynamic marine conditions, diverse spill scales, and limitations in existing algorithms regarding model size and real-time performance. We introduce LSFE-YOLO, a lightweight marine oil spill detection model built upon an optimized YOLOv8s architecture. The SAR images we used were obtained from the Alaska Satellite Facility (ASF) platform. Publicly available oil spill data were obtained from the National Oceanic and Atmospheric Administration (NOAA), including information such as the geographic coordinates, date of occurrence, and image source of the oil spill. NOAA (www.ospo.noaa.gov/Products/ocean/) has published verified oil spill data from the Gulf of Mexico, the Pacific Ocean, the Atlantic Ocean, the Great Lakes, and international waters, utilizing Sentinel-1 satellites to gather information on various types of oil spills. The corresponding Sentinel-1 SAR images can be obtained from the ASF (https://search.asf.alaska.edu/) platform using information such as the geographical coordinates and occurrence dates of oil spill incidents. Sentinel-1 monitors the surface in all weather conditions and provides high-quality SAR images, making it an ideal data source for our research. Our approach integrates FasterNet into the backbone and adjusts the neck network width to reduce memory overhead and parameters, significantly boosting detection speed. We also propose a novel GroupNorm and Lightweight Shared Convolution (GN-LSC) Head detection module that features a shared convolutional structure. This structure not only decreases the number of parameters in the detection head but also minimizes redundant computations.