Algorithm A1 Conservative loss and accuracy in once gradient descent
Input:		$\mathrm{Predicetion}: y_{pred}$
		$\mathrm{Conservatively} \mathrm{annotated} \mathrm{label}: labe{l}_{con}$
		$\mathrm{Treshold} \mathrm{for} \mathrm{sky}-\mathrm{annotation}: {\lambda }_s$
		$\mathrm{Treshold} \mathrm{for} \mathrm{ground}-\mathrm{annotation}: {\lambda }_g$
Output:		$\mathrm{Conservative} \mathrm{binary} \mathrm{cross}-\mathrm{entropy}: los{s}_{con}$
		$\mathrm{Conservative} \mathrm{binary} \mathrm{accyracy}: ac{c}_{con}$
1	Initialization:
2		Constant:$N_p$$, N_s$$, N_g$ = 0;
3		Array:$mas{k}_s$$, mas{k}_g$$, labe{l}_s$$, labe{l}_g$ = 0;
4	for$pixel$$\mathrm{in} labe{l}_{con}$ do
5
6
7
8
9
10
11
12
13
14
15	${\eta }_s=N_s/N_p$;
16	${\eta }_g=N_g/N_p$;
17	$labe{l}_s=mas{k}_s$;
18	$labe{l}_g=mas{k}_g$;
19	$y_{pred\_s}=y_{pred}\ast mas{k}_g$;
20	$y_{pred\_g}=(1-y_{pred})\ast mas{k}_g$;
21	$los{s}_{con\_s} \leftarrow y_{pred\_s}, labe{l}_s$;
22	$los{s}_{con\_g} \leftarrow y_{pred\_g}, labe{l}_g$;
23	$los{s}_{con} \leftarrow los{s}_{con\_s}+los{s}_{con\_g}$;
24	$ac{c}_s \leftarrow y_{pred\_s}, labe{l}_s$;
25	$ac{c}_{con\_s}=\left(ac{c}_s- {\eta }_s\right) / (1-{\eta }_s)$;
26	$ac{c}_g\leftarrow y_{pred\_g}, labe{l}_g$;
27	$ac{c}_{con\_g}=\left(ac{c}_g- {\eta }_g\right) / (1-{\eta }_g)$;
28	$ac{c}_{con}=\left(ac{c}_{con\_s}- ac{c}_{con\_g}\right) / 2$;