合成孔径雷达以高空间分辨率进行大范围观测的能力使其成为10~30千米尺度大小的强降雨团的有效观测手段之一。针对Sentinel-1波模式获取的9种海面现象的SAR影像组成的数据集，本文使用融合特征的宽度学习系统（Broad Learning System，BLS）进行了海面强降雨团的智能检测研究。结果表明强降雨团的检测精度为98.51%，召回率为95.24%，该结果与ResNet50预训练模型的结果相当，但是同等计算条件下后者的模型训练时间却是前者的20倍。此外，与传统深度学习网络相比，BLS的结构是灵活的，即可以通过增加节点或者是增加数据集来优化、更新模型。对于BLS的节点增量学习功能，本文实验证实其可以在无需重训练整个模型的前提下更新模型。针对训练数据集增广导致的模型更新任务，本文综合利用增量学习方案和重训练方案的优点提出了模型的混合更新方案，新方案既能保证模型的高精度又能显著降低模型更新所需时间。

The ability of synthetic aperture radar (SAR) to probe with a wide swath and high spatial resolution makes it one of the effective observation approaches of rain cells with the scale of 10 ~ 30 km. Using the SAR images dataset composed of nine sea surface phenomena obtained by Sentinel-1 wave mode, this paper uses the fusion features-based broad learning system (BLS) to detect the rain cells. The results show that the detection accuracy is 98.51% and the recall rate is 95.24%, which is equivalent to the result of ResNet50 pre-trained model. However, the training time of ResNet50 is 20 times that of BLS under the same calculation conditions. In addition, compared with the deep learning network, the structure of BLS is flexible, that is, the model can be optimized and updated by adding nodes or adding input data. For the nodes incremental learning of BLS, the experiments show that it can update the model without retraining the whole model. For the model updating task caused by the expansion of training dataset, this paper proposed a hybrid model updating scheme according to the advantages of incremental learning scheme and retraining scheme. This new scheme can not only ensure the high accuracy of the model, but also significantly reduce the time cost for model updating.