Process-Based Crop Modeling for High Applicability with Attention Mechanism and Multitask Decoders

Abstract

Crop models have been developed for wide research purposes and scales, but they have low compatibility due to the diversity of current modeling studies. Improving model adaptability can lead to model integration. Since deep neural networks have no conventional modeling parameters, diverse input and output combinations are possible depending on model training. Despite these advantages, no process-based crop model has been tested in full deep neural network complexes. The objective of this study was to develop a process-based deep learning model for hydroponic sweet peppers. Attention mechanism and multitask learning were selected to process distinct growth factors from the environment sequence. The algorithms were modified to be suitable for the regression task of growth simulation. Cultivations were conducted twice a year for 2 years in greenhouses. The developed crop model, DeepCrop, recorded the highest modeling efficiency (= 0.76) and the lowest normalized mean squared error (= 0.18) compared to accessible crop models in the evaluation with unseen data. The t-distributed stochastic neighbor embedding distribution and the attention weights supported that DeepCrop could be analyzed in terms of cognitive ability. With the high adaptability of DeepCrop, the developed model can replace the existing crop models as a versatile tool that would reveal entangled agricultural systems with analysis of complicated information.