Trjtdtk

Data: I have a spatio-temporal dataset which is approximately 5 years worth of crime data for New York City. This has been aggregated into a space-time grid so that the three dimensions of the matrix are longitude and latitude grid cells, and the time-frame. For this question let us say that there are 50x50 cells overlaid on the area and 100 time-frames for this question. So matrix dimensions are (50, 50, 100) and the cell value is the crime count.

Aim: I want to feed in data one time-frame at a time (as it becomes available) to predict the next timeframe.

Question: How should I shape the data for input as I am struggling on the understanding of how to make these forecasts and what data should be available/fed in?

Currently: Since the ConvLSTM2D Layer takes the following input (#Samples, Frame, Row, Col, Channel). I have reshaped as follows: (1, 100, 50, 50, 1) where I have a single video sample, which contains 100 frames, 50 rows and columns and a single channel. With x, y datasets as arr[:,:-1,:,:,:] and arr[:,1:,:,:,:] respectively.

Model I have created (might be wrong):

def createConvLSTMModel(dim0, dim1, dim2):

 #Create the model
 model = Sequential()

 ## Add layers to the model
 #Add the first convolutional LSTM unit (with input)
 model.add(ConvLSTM2D(filters=32,
 kernel_size=(3, 3),
 kernel_initializer='glorot_uniform',
 strides=(1,1),
 activation='relu',
 batch_input_shape=(None, dim1, dim2, 1),
 padding='same',
 #stateful=True,
 return_sequences=True))
 #Perform batch normalisation
 model.add(BatchNormalization())

 #Add the second convolutional LSTM unit
 model.add(ConvLSTM2D(filters=32,
 kernel_size=(3, 3),
 kernel_initializer='glorot_uniform',
 strides=(1,1),
 activation='relu',
 padding='same',
 #stateful=True,
 return_sequences=True))
 #Perform batch normalisation
 model.add(BatchNormalization())


 #Add the third convolutional LSTM unit
 model.add(ConvLSTM2D(filters=64,
 kernel_size=(3, 3),
 kernel_initializer='glorot_uniform',
 strides=(1,1),
 activation='relu',
 padding='same',
 #stateful=True,
 return_sequences=True))
 #Perform batch normalisation
 model.add(BatchNormalization())

 #Add the fourth convolutional LSTM unit
 model.add(ConvLSTM2D(filters=64,
 kernel_size=(3, 3),
 kernel_initializer='glorot_uniform',
 strides=(1,1),
 activation='relu',
 padding='same',
 #stateful=True,
 return_sequences=True))
 #Perform batch normalisation
 model.add(BatchNormalization())

 #Add a final 3D convolution Layer
 model.add(Conv3D(filters=1, kernel_size=(3, 3, 3),
 activation='sigmoid',
 padding='same', data_format='channels_last'))

 #Configure the model for training
 model.compile(loss='binary_crossentropy', optimizer='adadelta')

 #Return the model
 return(model)
```