I have a dataset of the form (samples, t, x, y, channels), where x, y are pixels and t the sequence length. The data is seasonal, so each t represents a season. t doesn’t extend far, as there’s only one temporal label that was build over many years. My hope is to use ConvLSTM2D to capture seasonal effects on my features and connect those seasonal changes to the same label. This model would then be used with features from the following year (and perhaps further) to predict the overall change of that year. As I understand it, it is a form of nowcasting, though what I want to predict are proportions (i.e. what classes make up an image).