Crop development monitoring from Synthetic Aperture Radar (SAR) imagery

Abstract

Satellite remote sensing plays a vital role in providing large-scale and timely data to stakeholders of the agricultural supply chain. This allows for informed decision-making that promotes sustainable and cost-effective crop management practices. In particular, data derived from satellite-based Synthetic Aperture Radar (SAR) systems, provide opportunities for continuous crop monitoring, taking advantage of its ability to acquire images during day or night and under almost all weather conditions. Moreover, an abundance of SAR data can be anticipated in the next 5 years with the launch of several international SAR missions. However, research on crop development monitoring with data from SAR satellites has not been as widely studied as with data derived from passive multi-spectral satellites and contributions can be made to the current state-of-the-art techniques.

This thesis aims at improving the current knowledge on the use of satellite-based SAR imagery for crop development monitoring. This is approached by developing novel methodologies and detailed interpretations of multitemporal SAR and Polarimetric SAR (PolSAR) responses to crop growth in three different test sites.

Chapter two presents a detailed analysis of the Sentinel-1 SAR satellite response to asparagus crop development in Peru, investigating the capabilities of the sensor to capture seasonality effects as well as providing an interpretation of the temporal backscatter signature. This is complemented with a case study where a multiple-output random forest regression algorithm is used to successfully retrieve crop growth stage from Sentinel-1 data and temperature measurements.

Following the limitations identified with this approach, a methodology that builds upon ideas of Bayesian Filtering Frameworks (BFFs) for crop monitoring is proposed in chapter three. It incorporates Gaussian processes to model crop dynamics as well as to model the remote sensing response to the crop state. Using this approach, it is possible to derive daily predictions with the associated uncertainties, to combine in near-real-time data from active and passive satellites as well as to estimate past and future crop key events that are of strategic importance for different stakeholders.

The final section of this thesis looks at the new developments of the SAR technology considering that future open access missions will provide Quad Polarimetric SAR data. An algorithm based on multitemporal PolSAR change detection is introduced in chapter four. It defines a Change Matrix to encode an interpretable representation of the crop dynamics as captured by the evolution of the scattering mechanisms over time. We use rice fields in Spain and multiple cereal crops in Canada to test the use of the algorithm for crop monitoring. A supervised learning-based crop type classification methodology is then proposed with the same method by using the encoded scattering mechanisms as input for a neural-network-based classifier, achieving comparable performances to state-of-the-art classifiers.

The results obtained in this thesis represent novel additions to the literature that contribute to our understanding and successful use of SAR imagery for agricultural monitoring. For the first time, a detailed analysis of asparagus crops is presented. It is a key crop for agricultural exports of Peru, the largest exporter of asparagus in the world. Secondly, two key contributions to the state of the art BFFs for crop monitoring are presented: a) A better exploitation of the SAR temporal dimension and an application with freely available data and b) given that it is a learning-based approach, it overcomes current limitations of transferability among crop types and regions. Finally, the PolSAR change detection approach presented in the last thesis chapter, provides a novel and easy-to-interpret tool for both crop monitoring and crop type mapping applications.