Satellite‐based precipitation retrieval is an essential and long‐standing scientific problem. With an increase of observational satellite data, the advances of data‐driven approaches such as machine learning (ML)/deep learning (DL) are favored to deal with large data sets and potentially improve the accuracy of precipitation estimates. In this study, we took advantage of new technologies by wrapping up a ML/DL‐based model pipeline (LinkNet segmentation + tree ensemble). This approach is applied to the Advanced Microwave Sounding Unit (AMSU) on National Oceanic and Atmospheric Administration 18 and 19 flight, and compared with the MultiRadar MultiSensor. Four simulations were configured to examine the performance gain by incorporating three components: (1) precipitation identification, (2) nonlocal features, and (3) precipitation classification. More importantly, we examined the interpretability of the “black box” model to get a better understanding of the underlying physical connections. First, the results by this model pipeline suggest the advantages of the ML model by reducing the systematic error and instantaneous error to a factor of two. Second, identifying precipitation pixels helps to reduce the systematic error by 130%, and predicting precipitation classification benefits improved correlations by 32%. Last, channels at higher frequencies (beyond 150 GHz) are favored to identify precipitation regions, and also channels at 89 and 150 GHz are ranked as the two most important features to precipitation retrieval. This study explores the potentials of AMSU precipitation estimates with ML algorithms and provides means of interpreting the models to facilitate the better prediction of precipitation.
1 August 2023 in Bulletin of the American Meteorological Society
The National Aeronautics and Space Administration (NASA) and National Oceanic and Atmospheric Administration (NOAA) have a long and successful history of weather radar research. The NOAA ground-based radars—WSR-88D network—provide nationwide precipitation observations and estimates with advanced polarimetric capability. As a counterpart, the NASA–JAXA spaceborne radar—the Global Precipitation Measurement Dual-Frequency Precipitation Radar (GPM DPR)—has global coverage and higher vertical resolution than ground-based radars. While significant advances from both NOAA’s WSR-88D network and NASA–JAXA’s spaceborne radar DPR have been made, no systematic comparisons between the WSR-88D network and the DPR have been done. This study for the first time generates nationwide comprehensive comparisons at 136 WSR-88D radar sites from 2014 to 2020. Systematic differences in reflectivity are found, with ground radar reflectivity on average 2.4 dB smaller than that of the DPR (DPR version 6). This research found the discrepancies between WSR-88D and DPR arise from different calibration standards, signal attenuation correction, and differences in the ground and spaceborne scattering volumes. The recently updated DPR version 7 product improves rain detection and attenuation corrections, effectively reducing the overall average WSR-88D and DPR reflectivity differences to 1.0 dB. The goal of this study is to examine the systematic differences of radar reflectivity between the NOAA WSR-88D network and the NASA–JAXA DPR and to draw attention to radar-application users in recognizing their differences. Further investigation into understanding and alleviating the systematic bias between the two platforms is needed.
19 June 2022 in Environmental Modelling & Software
This study presents a comprehensive review and comparative analysis of traditional machine learning (ML) and deep learning (DL) models for land cover classification in agricultural remote sensing. We evaluate the reported successes, trade-offs, and performance metrics of ML and DL models across diverse agricultural contexts. Building on this foundation, we apply both model types to the specific case of almond crop field identification in California’s Central Valley using Landsat data. DL models, including U-Net, MANet, and DeepLabv3+, achieve high accuracy rates of 97.3% to 97.5%, yet our findings demonstrate that conventional ML models—such as Decision Tree, K-Nearest Neighbor, and Random Forest—can reach comparable accuracies of 96.6% to 96.8%. Importantly, the ML models were developed using data from a single year, while DL models required extensive training data spanning 2008 to 2022. Our results highlight that traditional ML models offer robust classification performance with substantially lower computational demands, making them especially valuable in resource-constrained settings. This paper underscores the need for a balanced approach in model selection—one that weighs accuracy alongside efficiency. The findings contribute actionable insights for agricultural land cover mapping and inform ongoing model development in the geospatial sciences.