Search

In hydrology, runoff predictions are challenging when the data is lacking (e.g., predictions in un-gauged basins (PUB) and predictions with limited data (PLD)). Here, PLD refers to the case that the data is not enough for training or fine-tuning a data-driven model well (e.g., a new-gauged basin). We also name PLD as PNB (predictions in new-gauged basins). The difference between PNB and PUB is that the new-gauged basins can provide some data (e.g., runoff observations) while the un-gauged basins cannot. The long short-term memory (LSTM)-based models have shown good performance in runoff predictions due to their advanced structures. However, those structures have low level of flexibility and two nonadjacent positions cannot communicate directly. For high level of flexibility and better performance, we propose a simple Transformer-based rainfall-runoff model named RRS-Former, and want to show its power and also the power of the previously proposed RR-Former in PNB compared with LSTM-based models. The main part of RRS-Former is attention modules in which two arbitrary positions can be connected directly. Four hydrological units including 241 basins in the Catchment Attributes and Meteorology for Large-Sample Studies (CAMELS) dataset are used to compare the performance of our Transformer-based models and that of LSTM-based models. Besides using the k-fold validation to test performance for PNB, we propose a new way named unit-fold validation, in which we train the models by using the basins in three hydrological units and then test the performance using basins in the rest one hydrological unit. The results based on both k-fold and unit-fold show that our RRS-Former and RR-Former have better performance and are more reliable compared with the LSTM-based models. • A data-driven rainfall-runoff model based on Transformer is proposed and named RRS-Former. • A Transformer-based model is more flexible than LSTM and has better transfer ability. • RRS-Former has less parameters and less training time compared with RR-Former. • A Transformer-based model is more reliable than LSTM in the case of data lacking. • Unit-fold has more advantages for testing the transfer ability compared with k-fold. [ABSTRACT FROM AUTHOR]

• A Transformer based rainfall-runoff model is proposed for the first time. • RR-Former can connect arbitrary positions directly and is more flexible than LSTM. • RR-Former achieves excellent performance on individual rainfall-runoff modeling. • RR-Former achieves excellent performance on regional rainfall-runoff modeling. • RR-Former suits for big datasets and provides multi-step-ahead runoff predictions. Recently, the long short-term memory (LSTM) based rainfall-runoff models have achieved good performance and thus have received many attentions. In this paper, we propose a novel rainfall-runoff model named RR-Former based on the Transformer, which is entirely composed of attention mechanisms. Compared with a LSTM-based model, the architecture of RR-Former can connect two arbitrary positions in a time series process directly by using attention modules. It can strengthen or weaken the connection of two arbitrary positions and thus is more flexible than a LSTM-based model. Therefore, the RR-Former has potential to achieve better performance. By employing the Catchment Attributes and Meteorology for Large-Sample Studies (CAMELS) dataset, we test the performance of RR-Former in two tasks: individual rainfall-runoff modeling and regional rainfall-runoff modeling. In the first task, our RR-Former outperforms two LSTM-based sequence-to-sequence models significantly for 7-day-ahead runoff predictions. For example, the median and the mean of Nash–Sutcliffe efficiency for the 673 basins provided by our RR-Former achieve 0.8265 and 0.7904 , respectively, while those provided by the benchmark model (the better one between two benchmark models) are 0.7448 and 0.6952 , respectively. In the second task, our RR-Former also shows its power and suits for a big dataset better. [ABSTRACT FROM AUTHOR]

• A Robust Training Sample Selection strategy is proposed to effectively solve the problem of hyperparameter dependence. • Mitigate the impact of centerness loss without decreasing by designing a dynamic optimization loss. • A mixed attention module is proposed to enhance the multi-scale representation ability of feature pyramids. [Display omitted] As a pioneering work of introducing the idea of full convolutional network into the field of object detection, the fully convolutional one-stage object detection network (FCOS) has the advantage of excellent performance with low memory overhead. However, there are certain problems with FCOS that merit more research: the centerness quality assessment loss does not decrease during the late training stage, and its adaptive training sample selection (ATSS) relies heavily on the hyperparameter. To solve the aforementioned problems, we propose a novel object detection network, named augmented fully convolutional one-stage object detection network (AugFCOS). First of all, we propose an improved dynamic optimization loss (DOL) to mitigate the impact of the original centerness loss not decreasing. Then, a Robust Training Sample Selection (RTSS) is proposed to get rid of the dependence of hyper-parameter in ATSS of FCOS. Finally, a novel mixed attention feature pyramid network (MAFPN) is presented to enhance the multi-scale representation ability of feature pyramid network (FPN) and further improve the ability of multi-scale detection. The experimental results on MS COCO demonstrate the effectiveness of our proposed AugFCOS, where AugFCOS achieves approximate 2.0% to 2.9% increase when compared with ATSS and FCOS. [ABSTRACT FROM AUTHOR]