The left panel demonstrates how a two-unit LSTM layer processes a sequence with three features at each time step. For better visual illustration, we represent the sum of two matrix multiplications, in Equations (C.1)–(C.4), as a single matrix multiplication between a concatenated weight matrix of shape (2, 5) and a stacked vector of x_t and h_t − 1 of shape (5, 1), as described in Equation (C.7). The output h_t has a length of 2. The right panel, adapted from Naz et al. (2024), illustrates the generalization to ConvLSTM2D, where matrix multiplication (similar to fully connected operations) is replaced by convolution operations (denoted by *), while handling both the short-term memory and input data, all of which are 2D at each time step. While the gating mechanism remains the same between an LSTM and a ConvLSTM cell, the two panels depict them differently to enhance clarity.