General architecture of the ConvEntion network. The image time series are first rearranged to embed the band information. Then each 3DCNN is fed with a sub-sequence of K inputs of the time series J(∊ ℝ^M×H×W×2 for M elements of images of size HxW) to create the new downsized sequence S (∊ ℝ^{N×H′×W′×D}). S is fed to the positional encoder in order to add the information about the position, which outputs F(∊ ℝ^{N×H′×W′×D}). Then F is passed to ConvBERT which has L layers. The 3D max-pooling is used to downsize the output of ConvBERT for the classifier.