Fig. 1 Distribution of training data in the wildlife dataset from the Inner Mongolia Ulanba National Nature Reserve

Fig. 2 The architecture of Diff-SCC model. Conv, Convolution; SCConv2_x-SCConv5_x, Residual stages of SCConv; SRU, Spatial reconstruction unit; CRU, Channel reconstruction unit.

Fig. 3 Expansion of skewed dataset based on diffusion model. z0, Initial latent representation of the original image; zT, Latent representation after T steps of Gaussian noise addition; ${{{z}'}_{T1}}$, Intermediate latent representation in the reverse diffusion process;${{{z}'}_{0}}$, Latent representation of the image generated by the diffusion model.

Fig. 4 The workflow of SCConv convolution reconstruction. SRU, Spatial reconstruction unit; CRU, Channel reconstruction unit; Conv, Convolution unit; X, Input feature map; GN, Group normalization; XW 1, Feature map with high information content; XW 2, Feature map with low information content; XW 11, Self-reconstructed part of the high-information feature; XW 12, Cross features injected into the low-information regions; XW 22, Self-reconstructed part of the low-information feature; XW 21, Auxiliary supplement to the high-information regions; XW1, XW2, Enhanced feature maps after reconstruction; XW, Feature map after spatial convolutional reconstruction; αC, (1-α)C, Main and auxiliary channel splits; Xup, Feature map from the main channel group; Xlow, Feature map from the auxiliary channel group; GC, Group convolution; PC, Point convolution; Y1, Main features output by the upper branch; Y2, Auxiliary features output by the lower branch; Pooling, Pooling layer; β1, β2, Channel attention weights; Y, Final channel-reconstructed feature map.

Fig. 5 Comparison of predicted vs. actual classification for ResNet50 (a) and Diff-SCC (b) models on the ULB-12 Dataset. The color intensity in the confusion matrix reflects the number of samples at each position: Darker shades indicate a larger number of samples, whereas lighter shades indicate fewer samples. Therefore, a darker and more concentrated diagonal region suggests stronger recognition performance for the corresponding class, while more pronounced off-diagonal regions indicate more severe inter-class confusion.

Fig. 6 Threshold sensitivity curve analysis of D1, D2, and D3

Fig. 7 Effect of the mixing ratio of real and generated samples on low-frequency accuracy

Fig. 8 Comparison of different model methods