Core Concepts
The proposed MCSDNet model utilizes multi-scale spatiotemporal information to accurately detect Mesoscale Convective Systems (MCS) in remote sensing imagery.
Abstract
The paper presents a novel encoder-decoder neural network called MCSDNet for the detection of Mesoscale Convective Systems (MCS) in remote sensing imagery.
Key highlights:
MCSDNet targets multi-frame MCS detection, capturing the temporal evolution of MCS by exploring cross-frame variations in the image sequence.
It introduces a multi-scale spatiotemporal information module to extract multi-level semantic features from different encoder levels, making the model more suitable for extreme conditions with dense MCS distributions.
MCSDNet employs a Spatiotemporal Mix Unit (STMU) to capture both intra-frame features and inter-frame correlations, which is a scalable module that can be replaced by other spatiotemporal modules.
The authors propose a new spatiotemporal module called Dual Spatiotemporal Attention (DSTA) as the STMU, which can effectively capture both spatial and temporal dependencies.
The paper introduces the first publicly available dataset for multi-frame MCS detection, called MCSRSI, based on visible channel images from the FY-4A satellite.
Experiments on the MCSRSI dataset show that MCSDNet achieves state-of-the-art performance on MCS detection, particularly in extreme conditions with dense MCS distributions.
Stats
The proportion of MCS regions in the observation data collected in China in 2018 is less than 5% in over 60% of the satellite images.
Quotes
"Considering the dramatic and changeable characteristics of MCS, they are limited to achieve excellent performance."
"As far as we know, it is the first work to utilize multi-scale spatiotemporal information to detect MCS regions."
"MCSDNet considers both spatial and temporal information of MCS regions and gains better performance in MCS detection task."