2.3 Selection of ephemeral gully identification model
The SegNet model is a convolutional neural network for pixel-level image segmentation proposed by Badrinarayanan et al. (2017) using end-to-end and pixel-to-pixel training to achieve image segmentation. Its core component is an encoder network and corresponding decoder network, followed by a pixel-level classifier, which outputs the probability map of the K channel, where K is the number of classification categories. Compared with other models, SegNet has fewer training parameters, smaller memory occupation, and shorter network training time (Deng et al., 2022; Jiang et al., 2020; Manickam et al., 2020). In 2021, six indexes (Accuracy, Precision, Recall, F1 value, ROC curve, and AUC) were used to compare the ephemeral gully recognition results and accuracy evaluation with U-Net, R2U-Net, and SegNet, with SegNet ranked first for ephemeral gully recognition in the hilly and gully region of the Loess Plateau, followed by R2U-Net and U-Net (Liuet al. , 2022). Ephemeral gully length and width between predicted and field measured values had RMSE values of 6.78 m (R2 = 0.9817) and 0.50 m (R2 = 0.8573), respectively, using the SegNet model, indicating its superior performance for ephemeral gully recognition and morphological feature extraction. Hence, this study used the SegNet model for ephemeral gully recognition and morphological feature extraction at the watershed scale. The process of identifying ephemeral gullies with the SegNet model was performed as follows: SegNet model carries out iterative cyclic training based on the data set until the loss value is minimum and obtains the corresponding optimal weight of the model. Based on the optimal weight and network structure of the SegNet model, ephemeral gullies were recognized using the sliding window and ignoring edge detection methods at the watershed scale (Fig. 3).