Transfer Learning for Crop Yield Prediction Using Time Series Data: A Comparative Study of InceptionTime and ResNet Models

By Moses Okolo

Abstract
Predicting potato yield using limited real-world data is a challenge. Previous studies have shown that deep learning-based models perform below expectations when compared to simple linear regression with expert selected features. We propose using deeper, more complex model architectures on the timeseries features to improve overall model performance, namely the InceptionTime and ResNet architectures. We tested these models in two different domains, on a synthetic dataset (source domain) spanning weather conditions for the years 1990-2020 (991k data points), on a dataset of potato field trials carried out in the Netherlands spanning 1994-2003 (303 data points) and on a dataset from a commercial farm spanning 2015-2020 (77 data points) both of which represent the target domain. We compared our models to two baselines, a deep learning (I-D CNN) and machine learning (ROCKET) based baseline respectively, in the source domain, our hypothesis here being the Inception and ResNet models would outperform both baselines. Our hypothesis here was not completely confirmed because even though the best performing model in this domain was the ResNet, the deep learning baseline outperformed the InceptionTime model. We took both models (InceptionTime and ResNet) pretrained from the source domain and fine-tuned them on the target domain. We also compared them to the same baselines as in the source domain and a third model with the same architecture as the ResNet (best performing in the source domain) but trained with new randomly initialized weights. Our hypothesis here was that the pretrained models would outperform all other three models. Our hypothesis was confirmed in the commercial potato farm dataset with the InceptionTime model outperforming all followed closely by the pretrained ResNet model but not in the potato field trials. The InceptionTime model outperformed all other models again but the ResNet with new weights outperformed the pretrained ResNet model. Our conclusion was that these models lead to an improvement in performance overall even though their performances are not very predictable, with models performing well in one domain being outperformed by different models in the other domain, concretely the InceptionTime performing better than the ResNet in the target domain even though it underperformed in the source domain.

Keywords: crop yield prediction, potato yield, InceptionTime, ResNet, deep learning, machine learning