Indirect Weight Measurement Model for Mango Post-production Process Based on Gated Recurrent Unit and Bidirectional Long Short-Term Memory

Demand for faster postharvest processing has rocketed for the past decades with the main purpose of increasing throughput via machine-based productivity. Philippines, landing on a top spot as a worldwide mango producer, is expected to equip postharvest systems including packaging and storing subsystems with state-of-the-art technologies to address the demand for automated weight determination. To address this technical challenge, this study developed an RNN-based sequence layers approach composed of Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Bidirectional LSTM (BiLSTM) in indirectly measuring the weight of mango fruit based on four extracted morphological phenes, namely, length, width, and two-pixel count averages with relation to 15 fps video frame as the object-of-interest (mango) is moving along the conveyor belt. These phenes were set as input to the developed nine three-layer LSTM, GRU, and BiLSTM network variants which resulted in GRU-BiLSTM-LSTM (LWP2) model exhibiting an RMSE of 7.665, R
²
of 0.985, MAPE of 0.026, and inference time of 1.92 s – a model performance that sets apart LWP2 from other optimized RNN-based sequenced layers models. The four phenes were confirmed to have strong positive correlations with mango weight ranging from 0.765 to 0.959. Hence, this vision-based LWP2 model is recommended to be embedded in a mango postharvest processing system.