Multi-modal crop yield prediction — Cameroon

Six deep learning architectures compared over three nested data configurations, for maize, cassava and cocoa in the Centre and Nord regions of Cameroon.

The interesting engineering here was not the models but the discipline around them. An early version of the pipeline carried five distinct defects that each inflated apparent performance: target scaling applied before the train/test split, scalers fitted on the full dataset, missing early stopping, and two related leakage paths through the validation fold. Correcting them lowered the headline numbers and made them mean something.

Architectures: CNN-LSTM, ConvLSTM, ConvLSTM-ViT, CNN-GRU, TCN, CNN-Transformer.