Optimizing Gaussian Process Regression for Image Time Series Gap-Filling and Crop Monitoring

Area of Science:

• Remote Sensing and Geospatial Analysis
• Agricultural Science and Crop Monitoring
• Artificial Intelligence and Machine Learning Applications

Background:

• Satellite image time series analysis is crucial for monitoring crop phenology, including start and end of season estimations.
• Gaussian Process Regression (GPR) is a powerful Bayesian algorithm for time series gap-filling, offering uncertainty estimates.
• Standard per-pixel GPR processing is computationally intensive, hindering large-scale application.

Purpose of the Study:

• To develop a computationally efficient GPR approach for satellite image time series gap-filling.
• To assess the accuracy and performance of the proposed method compared to standard per-pixel GPR.
• To evaluate the impact of the optimized GPR on crop phenology indicator retrieval.

Main Methods:

• Proposed a novel method involving cropland-based precalculations of GPR hyperparameters (θ) to replace per-pixel optimization.
• Utilized Sentinel-2 Leaf Area Index (LAI) time series data from an agricultural region in Spain for validation.
• Compared image reconstruction accuracy (RMSE) and processing time against standard per-pixel GPR and a scene-wide hyperparameter approach.

Main Results:

• The proposed method achieved comparable accuracy to per-pixel GPR (12% RMSE degradation) while accelerating processing time by approximately 90 times.
• Using scene-wide hyperparameters yielded similar accuracies and computational performance for crop areas.
• Calculated crop phenology indicators (start and end of season) showed analogous temporal patterns across all tested methods, with differences within five days.

Conclusions:

• The proposed GPR hyperparameter precalculation significantly enhances computational efficiency for satellite time series analysis without substantial accuracy loss.
• This optimized approach is suitable for large-scale crop monitoring applications, providing reliable phenological insights.
• The developed gap-filling and phenology retrieval techniques are integrated into the freely available DATimeS GUI toolbox.