Learning-Based Tracking of Crop Biophysical Variables and Key Dates Estimation From Fusion of SAR and Optical Data

Abstract

Monitoring crop development is of crucial importance to ensure sustainable management practices while promoting efficient land use. The ability of satellite remote sensing data to cover large areas offers a robust tool to aid this task. In this article, we propose a filtering framework, which uses Gaussian-process-based dynamic and observation models, an unscented Kalman filter, and the fusion of multitemporal SENTINEL-1 and SENTINEL-2 data to monitor crop biophysical variables. This method complements state-of-the-art filtering frameworks given its ability to learn models and uncertainties from data and to exploit the imagery temporal dimension. This enables the method to be transferable to other crop types, biophysical variables, and locations. We test the methodology to track asparagus below-ground carbohydrates and the season crop age and to forecast crop key dates. The amount of carbohydrates stored below ground in the plant's root system is highly associated with the yield of asparagus and the ability to establish a healthy canopy. Validation with ground truth showed that the use of more than one SENTINEL-1 orbit and SENTINEL-2 data provided the best tracking performances and a reliable way for handling missing data from a sensor. Under this configuration, the method achieves a mean absolute error (MAE) of 1.802 Brix degrees (surrogate for carbohydrates). Similarly, it can retrieve crop age and forecast harvest date, with the MAE of six days. Remotely tracking below-ground carbohydrates may contribute toward reducing the destructive sampling required for its measurement in the field.

Keywords

• Agriculture
• asparagus
• crop development
• data fusion
• forecasting
• gap filling