Mapping of the wind erodible fraction of soil by bidirectional gated recurrent unit (BiGRU) and bidirectional recurrent neural network (BiRNN) deep learning models

The destructive consequences of wind erosion have been reported in many studies, but accurate assessment of wind erosion is still a challenge, especially on large scales. Our research introduces two deep learning (DL) al-gorithms consisting of bidirectional gated recurrent unit (BiGRU), and bidirectional recurrent neural network (BiRNN) for spatial mapping of wind-erodible fraction of the soil (EF). EF was measured in 508 soil samples using the Chepil method. 15 key factors controlling EF including: soil, topography, and meteorology parameters were mapped. The performance of the most efficient DL model was interpreted by Game theory. The uncertainty of the DL models was quantified by deep quantile regression (DQR). Results showed that both DL models were per-formed very well with the BiRNN performing slightly better than BiGRU. The aggregate mean weight diameter (MWD) was a key variable for the mapping of soil susceptibility to wind erosion. Based on the BiRNN model, most of the study region was moderately and highly susceptible to wind erosion regarding the EF value (between 32 and 98). This indicates the urgent need for soil conservation measures in the region. The DQR results showed that the observed values of EF fell within the EF values predicted by the model. Overall, the suggested meth-odology has proven to be helpful in mapping wind erosion susceptibility on a large scale.

Automated summary

Two deep learning algorithms, BiGRU and BiRNN, were introduced to accurately assess the spatial mapping of wind-erodible fraction of the soil. The results showed that the region was highly susceptible to wind erosion, highlighting the urgent need for soil conservation measures.