|dc.description.abstract||It is important to have an accurate estimate of the yields and prices of fresh produce (FP) for maintaining an effective Fresh Produce Supply Chain Management (FSCM). Since, the FP comprises of the perishable goods, it is cumbersome to manage and keep a track of logistics, which makes it important to have an estimate of the FP yield to have a better management of the supply and demand. In addition, having a reliable estimate of the FP prices helps the food company to bid the right price to the wholesalers. This prevents the food company from bidding unreasonable price and incurring any loss. Computational tools for forecasting yields and prices for fresh produce have been based on conventional machine learning approaches or time series modeling. These approaches can neither effectively capture the complex relationships between the inputs and the outputs to the models nor can they handle large datasets. To overcome such drawbacks, Deep Learning (DL) based approaches are proposed in this work for forecasting the yields and prices of FP. Soil and weather parameters of counties across California are used to forecast the yields and prices of FP like berries and apples.
Choosing the most effective input parameters for forecasting strawberry yields and prices is investigated. The set of parameters used for this investigation are soil parameters alone and soil parameters along with the weather parameters. For this forecasting, the ensemble of two DL models is used namely, Convolutional Neural Networks and Long Short Term Memory with Attention (Att-CNN-LSTM) and Convolutional LSTM with Attention (Att-ConvLSTM). It is found that using soil and weather parameters together gives better forecasting results than using soil or weather parameters alone. Also, various compound DL models like Att-CNN-LSTM, Att-ConvLSTM, Temporal Convolutional Network (TCN) and SeriesNet with Gated Recurrent Unit (SeriesNet-GRU) are tested for forecasting, to determine the best performing DL model. It is found that the ensemble of two compound DL models Att-CNN-LSTM and SeriesNet-GRU gives the best forecasting results with an improvement of around 7% in the value of Aggregated Measure (AGM) than the component compound DL models. It also outperforms the previous work done in literature with an improvement of around 14% in the value of AGM. The effect of using soil input parameters on yield forecasting is further studied. To study the effect of static soil parameters on forecasting performance, the compound DL model SeriesNet with GRU is used to forecast the annual apple yield using the static and dynamic soil parameters. The county level annual apple yield forecast, using both static and dynamic parameters together, proves to give promising results, it reduces the forecasting AGM by around 34% compared to the case of excluding the static parameter and only using the dynamic parameters set. It is also found that, on using an augmented training set to train the DL model improves the AGM value by around 12% on testing with the non-augmented test set.
To generalize the findings, transfer learning technique is utilized amongst the yield forecasting models of the similar crops. To overcome the computational complexity of retraining DL yield forecasting models for each type of FP, it is necessary to have a generalization of the models’ application to similar FP with minimal retraining. Two berries are considered in this work, California strawberries and raspberries which have similar yield, since the two follow similar time series on the basis of a number of parameters such as lag, seasonality and trend. The voting regressor ensemble of two compound DL models Att-CNN-LSTM and SeriesNet with GRU is used. First, the proposed DL model is trained using station-based soil data input mapped to the strawberry yield as output. The weights obtained from this learning are transferred to the raspberry yield forecasting ensemble model with minimal retraining. It is found that the transfer learning gives comparable results to training from scratch and reduces the processing time by half.||en