Conv 3 × 3 Conv 3 × 3 (a) Conv 3 × 3 Conv 3 × 3 Conv 3 × 3 (b) hi+1 FM FM FM (c) Conv 3 × 3 Conv 3 × 3 Conv 3 × 3 FM FM FM Depthwise Pointwise (d) FM FM FM Concat Add Add Add Add Add Sep 3 × 3 Identity Sep 3 × 3 Sep 5 × 5 Avg 3 × 3 Identity Avg 3 × 3 Avg 3 × 3 Sep 5 × 5 Sep 3 × 3 hi hi-1 (e) Atr 5 × 5 + Sep 3 × 3 Sep 3 × 3 HI-2 HI-1 Sep 3 × 3 Atr 3 × 3 + Sep 3 × 3 (f) + Atr 5 × 5 + Sep 5 × 5 Sep 5 × 5 + Sep 5 × 5 HI FIGURE 2. Illustration of the commonly used model compression techniques. (a)-(d) Some widely used handcrafted convolutional blocks: (a) bottleneck convolution [51], (b) grouped convolution [52], (c) inverted bottleneck convolution [53], and (d) depthwise separable convolution [54]. (e), (f) Some popular automatically learned cells: (e) NASNet [55] and (f) Auto-DeepLab [56]. (Continued) 98 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE DECEMBER 2023 Conv 1 × 1 Conv 1 × 1 Split to Groups Concatenation Conv 1 × 1 Depthwise Conv 3 × 3 Conv 1 × 1 Concat