**6. Discussion**

We also discuss the shadow detection performance of ShadowDeNet on another video SAR dataset to reveal its universal effectiveness and excellent migration ability. This data is provided by Zhou et al. [12] and China Aerospace Science and Industry Corporation (CASIC) 23 research institute. This data was obtained from the spotlight SAR that was equipped on an aircraft and flew along a circular trajectory at a height of 3000 m. The carrier frequency, resolution, and velocity of aircraft are Ka-band, 0.15 m, and 80 m/s. There are 369 images with 1000 × 1000 pixel size. Among them, 246 images are selected as the training set, and the remaining 123 images are selected as the test set.

Figure 20 shows the qualitative results of ShadowDeNet on the CASIC 23 research institute video SAR data, and the quantitative results are shown in Table 12. From Figure 20, many moving target shadows can be detected by ShadowDeNet successfully, showing the excellent migration ability of ShadowDeNet. From Table 12, ShadowDeNet offers comparable shadow detection performance on the CASIC 23 research institute video SAR

data with the SNL video SAR data, i.e., 66.01% *f* 1 vs. 66.28% *f* 1 and 52.39% *ap* vs. 51.87% *ap*. Therefore, ShadowDeNet is effective on video SAR data, showing its universal effectiveness.

**Figure 20.** Qualitative video SAR moving target shadow detection results of ShadowDeNet on the CASIC 23 research institute data. The ground truths are marked by green boxes. The numbers above boxes are the confidences. The IOU threshold is 0.50, the same as the PASCAL VOC criterion [79].

**Table 12.** Quantitative results of ShadowDeNet on the CASIC 23 research institute video SAR data.

