Complexity to Forecast Flood: Problem Definition and Spatiotemporal Attention LSTM Solution

Table 2

Comparison experiments on Lech dataset to show sensibility of different input hydrological features, where Wo- refers to operation of eliminating one specifical hydrological feature, ST-Water is short for hydrological feature of soil tension water, subscripts A and σ represent mean value and bias value respectively, R, M and D represents measurements of RMSE, MAPE, and DC, respectively. Note that bold texts indicate best performance among comparative methods. We define unit for M as %.

Methods	RT+3	RT+6	RT+9	RA	Rσ	MT+3	MT+6	MT+9	MA	Mσ	DT+3	DT+6	DT+9	DA	Dσ
Wo-precipitation	117.0	121.3	145.3	127.9	61.88	29.61	31.96	36.31	32.63	16.12	0.35	0.33	0.28	0.32	0.49
Wo-evaporation	65.82	74.08	85.16	75.02	9.00	16.49	18.53	22.39	19.14	2.63	0.80	0.76	0.64	0.73	0.08
Wo-ST-water	80.44	74.74	77.92	77.70	11.68	19.91	18.92	19.64	19.49	2.98	0.71	0.80	0.73	0.75	0.06
Wo-temperature	63.38	68.51	75.50	69.13	3.11	15.94	17.18	19.04	17.39	0.88	0.82	0.79	0.74	0.78	0.03
Wo-wind	63.81	68.12	75.91	69.28	3.26	15.98	17.06	18.32	17.12	0.61	0.82	0.78	0.76	0.79	0.02
STA-LSTM	60.94	65.56	71.56	66.02	—	15.24	16.39	17.89	16.51	—	0.84	0.81	0.77	0.81	—