*5.4. Ablation Study on SGAAL*

Table 9 shows the quantitative detection results of ShadowDeNet with and without SGAAL. From Table 9, SDAL can improve the overall performance of ShadowDeNet by ~2.8% *f* 1 accuracy and ~3% *ap* accuracy, which shows its effectiveness. For one thing, its internal anchor location network can predict shadow-like regions adaptively to suppress false alarms. For another thing, its internal anchor shape network can predict better shapes to match moving target shadows adaptively to improve the detection rate. As a result, the false alarm rate is dropped by ~2.3%; meanwhile, the detection rate is increased by ~3.0%.


**Table 9.** Quantitative results of ShadowDeNet with and without SGAAL.

We perform another experiment to study the impact of the location filter threshold *εL*. The quantitative results are shown in Table 10. In Table 10, the value range of *εL* is suggested by Wang et al. [45]. From Table 10, when *εL* is set to 0.01, the accuracy reaches the best. Thus, the final *εL* is set to 0.01 in ShadowDeNet. Our experimental results are in line with [45] where *εL* is also set to 0.01. According to their findings, a small location filter threshold has already removed many false positives. However, a too-large location filter threshold is bound to remove many true positives, resulting in a lower detection rate. We find that such phenomenon is also in line with video SAR images.

**Table 10.** Quantitative results of ShadowDeNet with different location filter thresholds.

