[Paper Review] Spatial-temporal Multi-Task Learning for Within-field Cotton Yield Prediction

Long Nguyen, Jia Zhen+5 more|arXiv (Cornell University)|Nov 16, 2018

Smart Agriculture and AIAgricultural and Biological Sciences3 citations

TL;DR

This paper proposes a spatial-temporal multi-task learning framework that integrates soil, spectral, NDVI, and weather data across multiple years to predict within-field cotton yield at 30m resolution in West Texas (2001–2003). By sharing features across tasks and applying spatial regularization to exploit neighborhood correlations, the model achieves state-of-the-art performance, reducing RMSE to 83.7 and MAE to 63.6 in 2003, outperforming conventional methods including SVM and single-task models.

ABSTRACT

Understanding and accurately predicting within-field spatial variability of crop yield play a key role in site-specific management of crop inputs such as irrigation water and fertilizer for optimized crop production. However, such a task is challenged by the complex interaction between crop growth and environmental and managerial factors, such as climate, soil conditions, tillage, and irrigation. In this paper, we present a novel Spatial-temporal Multi-Task Learning algorithms for within-field crop yield prediction in west Texas from 2001 to 2003. This algorithm integrates multiple heterogeneous data sources to learn different features simultaneously, and to aggregate spatial-temporal features by introducing a weighted regularizer to the loss functions. Our comprehensive experimental results consistently outperform the results of other conventional methods, and suggest a promising approach, which improves the landscape of crop prediction research fields.

Motivation & Objective

• To address the challenge of accurate within-field cotton yield prediction under complex interactions between crop growth, weather, soil, and management factors.
• To improve prediction accuracy by leveraging heterogeneous data sources—soil, spectral, NDVI, and weather—across multiple years through multi-task learning.
• To incorporate spatial autocorrelation by introducing a spatial regularization term that minimizes yield differences between neighboring grid cells.
• To enable real-time, month-by-month yield forecasting during the growing season using only available data up to each prediction point.
• To evaluate the relative importance of different data sources (soil, spectral, NDVI) in yield prediction performance.

Proposed method

• The model employs a multi-task learning architecture where each year (2001–2003) is treated as a separate but related prediction task, sharing a common feature representation layer.
• Heterogeneous input features—including soil topography (elevation, slope, curvature), spectral bands (BAND1–BAND4), NDVI, and weather (temperature, rainfall)—are fed into a shared encoder network.
• A spatial regularization term is introduced to the loss function, penalizing large differences between a grid’s predicted yield and the weighted average of its neighbors, with neighborhood size tuned to 5.
• The model is trained end-to-end using a composite loss function combining task-specific regression losses and the spatial regularization term.
• Real-time predictions are evaluated monthly (May–September), using only data available up to that point, simulating online forecasting.
• Feature importance is assessed by ablation: removing one data source at a time and measuring performance degradation in MSE, RMSE, MAE, MAPE, and Max Error.

Experimental results

Research questions

• RQ1 Can a multi-task learning framework that shares representations across years improve within-field cotton yield prediction accuracy compared to single-task or conventional models?
• RQ2 How does incorporating spatial regularization based on neighboring grid relationships affect prediction error and spatial consistency?
• RQ3 What is the relative contribution of soil properties, pre-planting spectral data, and NDVI to the final prediction performance?
• RQ4 How does model performance evolve over the growing season when predictions are made incrementally with increasing data availability?
• RQ5 What is the optimal neighborhood size for spatial regularization in terms of minimizing prediction error?

Key findings

• The proposed spatial-temporal multi-task learning model achieved the lowest error metrics in 2003, with MSE = 7,013.5, RMSE = 83.7, MAE = 63.6, MAPE = 7.55, and Max Error = 254.4.
• The model consistently outperformed all baselines, including Support Vector Regression and standard multi-task learning, across all three years (2001–2003).
• Performance improved significantly from May to September, with the largest gains occurring in August and September, indicating strong data accumulation benefits.
• Spatial regularization reduced error metrics most effectively when neighborhood size was set to 5, after which performance stabilized.
• NDVI was the most influential data source, as removing it caused the largest performance drop; spectral data before planting had the least impact.
• The ablation study confirmed that removing spectral data before planting yielded the best performance among all ablation settings, indicating its limited predictive value compared to NDVI and soil data.