Abstract
Accurate forecasting of fresh produce yields and prices is critical for effective supply chain management but remains challenging due to the non-linear impact of environmental variables. This paper presents a hybrid voting ensemble framework that integrates local station-based soil data with remote sensing satellite imagery to predict strawberry market conditions five weeks in advance. By combining a sophisticated station-based model—utilizing SeriesNet and Attention-based CNN-LSTMs—with a remote sensing ensemble, the proposed approach achieves significant performance gains. Results demonstrate a 33% improvement in yield forecasting and a 21% improvement in price forecasting over existing state-of-the-art models, highlighting the efficacy of multi-source data fusion in agricultural analytics.
Introduction
Fresh produce Supply Chain Management (FSCM) is characterized by extreme price volatility driven by fluctuating yields. Traditional time-series forecasting often fails to account for the complex, non-linear relationships between soil health, weather patterns, and market outcomes. Furthermore, a significant "data gap" exists: while local weather stations provide high-fidelity soil data, their geographic coverage is limited. Conversely, satellite imagery offers global coverage but presents challenges in feature extraction for predictive modeling. This study addresses these challenges by developing a robust deep learning system that bridges ground-truth station data with remote sensing observations to stabilize fresh produce forecasting.
Methods
The researchers developed a multi-stage voting ensemble framework consisting of two primary branches:
- Station-based Model (ATT-CNN-LSTM-SeriesNet_Ens): This branch processes local soil temperature and moisture data. It employs a combination of SeriesNet, which utilizes dilated causal convolutions to capture multi-level features, and a CNN-LSTM architecture enhanced with an additive self-attention layer. The attention mechanism allows the model to prioritize historical time steps most relevant to future yields.
- Remote Sensing Model (SIM_CNN-LSTM_Ens): This branch utilizes satellite-derived surface temperature and soil moisture data. To handle the high dimensionality of raw imagery, data is converted into pixel frequency histograms, which are then processed by an ensemble of CNN-LSTM models.
The final output is a weighted average of these branches, leveraging the complementary strengths of local and global data sources.
Data and Preprocessing
The study focused on Santa Barbara County, California, using data from 2011 to 2019. Yield and price data were sourced from the California Strawberry Commission. Station-based soil parameters (temperature and moisture) were obtained from the National Oceanic and Atmospheric Administration (NOAA), while remote sensing data (MODIS surface temperature and NASA-USDA soil moisture) provided global context. Principal Component Analysis (PCA) was employed to reduce the station-based feature set from 280 to the 36 most influential parameters.
Results and Discussion
The model's performance was evaluated using an Aggregated Measure (AGM)—a composite metric combining MAE, RMSE, and $R^2$ (where lower is better).
- Yield Forecasting: The voting ensemble achieved an AGM of 6.0, representing a 33% improvement over previous state-of-the-art deep learning ensembles and a 56% improvement over a standard LSTM.
- Price Forecasting: The model achieved an AGM of 0.0555, a 21% improvement over existing literature and a 53% improvement over standard LSTMs.
Key insights include the discovery of a 20-week lag, where soil conditions from the preceding 140 days are the most critical indicators for forecasts five weeks out. Additionally, the self-attention layer was instrumental in filtering daily noise to identify seasonal trends. The study also noted specialization: the remote sensing model excelled at yield prediction, while the station-based model was superior for price forecasting.
Conclusion
The proposed hybrid ensemble significantly advances the state-of-the-art in agricultural forecasting. While the model demonstrates high accuracy in following general market trends, it still faces challenges in predicting "steep fluctuations" caused by extreme outlier events like heatwaves. Future research will explore transfer learning to adapt this framework to other perishable crops and investigate methods to better capture sudden market shocks.
Summary updated