Abstract
Accurate forecasting of fresh produce yields and prices is essential for stabilizing supply chains and ensuring food security. Traditional time-series and machine learning models often struggle with the non-linear complexities of agricultural data. This paper proposes a novel deep learning framework that ensembles localized station-based soil data with broad-scale satellite imagery. By utilizing a voting ensemble of specialized architectures—including Attention-based CNN-LSTMs and SeriesNet—the model achieves a 33% improvement in yield forecasting and a 21% improvement in price forecasting over existing state-of-the-art methods for a five-week lead time.
Introduction
Predicting the market dynamics of strawberries is notoriously difficult due to the interplay of weather volatility, soil conditions, and fluctuating market demand. While localized station data (e.g., soil sensors) provides high precision, it is often geographically sparse. Conversely, satellite imagery offers wide coverage but requires sophisticated processing to extract daily or weekly signals. This research bridges these gaps by developing a hybrid model that extracts complementary features from both data sources to provide reliable mid-term forecasts for Santa Barbara County, California.
Methodology
The proposed framework utilizes a Voting Ensemble architecture composed of two primary branches, each designed to handle a specific data modality.
The Station-Based Model (ATT-CNN-LSTM-SeriesNet_Ens)
This branch processes soil temperature and moisture data using two sophisticated sub-models:
- Att-CNN-LSTM: This architecture combines Convolutional Neural Networks (CNNs) for spatial feature extraction and Long Short-Term Memory (LSTM) networks for temporal patterns. A critical component is the Additive Attention layer, which allows the model to weigh historical time steps based on their relevance to the future forecast.
- SeriesNet with GRU: Designed specifically for time-series data, this model employs dilated causal convolutions and skip connections. This structure enables the model to aggregate multi-scale contextual information without losing resolution, which is then processed by Gated Recurrent Units (GRUs) to capture long-term dependencies.
The Remote Sensing Model (SIM_CNN-LSTM_Ens)
To process satellite images efficiently, the researchers converted raw imagery into pixel frequency histograms. This dimensionality reduction technique preserves essential environmental signals while ensuring the model is invariant to pixel shuffling. These histograms are then fed into an ensemble of CNN-LSTM models.
Final Fusion
The outputs from both branches are combined via a Voting Ensemble. This final stage averages the predictions, effectively smoothing out individual model errors and leveraging the fact that satellite data often excels at yield prediction while station data is more sensitive to price-influencing soil conditions.
Data and Preprocessing
The study utilized data from 2011 to 2019 in Santa Barbara County:
- Yield/Price: California Strawberry Commission.
- Soil Data: NOAA station-based moisture and temperature.
- Satellite Data: MODIS surface temperature and NASA-USDA soil moisture.
- Feature Engineering: Principal Component Analysis (PCA) was used to reduce 280 station features to the 36 most impactful variables. Land-cover masks were applied to satellite data to isolate croplands.
Key Results
Performance was measured using an Aggregated Measure (AGM), combining RMSE, MAE, and $R^2$ (lower is better).
- Yield Forecasting: The voting ensemble achieved an AGM of 6.0, representing a 33% improvement over previous deep learning ensembles (AGM 9.0).
- Price Forecasting: The model achieved an AGM of 0.0555, outperforming literature standards by 21%.
- Predictive Power: The model demonstrated a high coefficient of determination ($R^2 \approx 0.87$) for yield, indicating a strong ability to track actual market trends five weeks in advance.
Conclusion
The integration of localized soil sensors and remote sensing imagery through a deep learning ensemble significantly enhances the accuracy of strawberry market forecasts. While the model currently faces challenges in capturing extreme, sudden fluctuations, its superior performance over traditional methods marks a significant step forward. Future research will focus on using transfer learning to adapt this framework to other regions and produce types with minimal retraining.
Summary updated