[Paper Review] Deep Recurrent Neural Networks for mapping winter vegetation quality coverage via multi-temporal SAR Sentinel-1

Dinh Ho Tong Minh, Dino Ienco+5 more|arXiv (Cornell University)|Aug 11, 2017

Remote Sensing in Agriculture7 references16 citations

TL;DR

This study proposes a deep recurrent neural network (RNN) approach using multi-temporal Sentinel-1 SAR data to map winter vegetation quality coverage in La Rochelle, France. By leveraging Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) networks to model temporal dependencies in VV and VH polarization time series, the method significantly outperforms traditional classifiers, achieving 99.05% F-measure with GRU and 98.83% with LSTM, demonstrating superior discrimination of complex temporal patterns in vegetation classes including low and very low quality cover.

ABSTRACT

Mapping winter vegetation quality coverage is a challenge problem of remote sensing. This is due to the cloud coverage in winter period, leading to use radar rather than optical images. The objective of this paper is to provide a better understanding of the capabilities of radar Sentinel-1 and deep learning concerning about mapping winter vegetation quality coverage. The analysis presented in this paper is carried out on multi-temporal Sentinel-1 data over the site of La Rochelle, France, during the campaign in December 2016. This dataset were processed in order to produce an intensity radar data stack from October 2016 to February 2017. Two deep Recurrent Neural Network (RNN) based classifier methods were employed. We found that the results of RNNs clearly outperformed the classical machine learning approaches (Support Vector Machine and Random Forest). This study confirms that the time series radar Sentinel-1 and RNNs could be exploited for winter vegetation quality cover mapping.

Motivation & Objective

• To address the challenge of mapping winter vegetation quality coverage in regions with high cloud cover, which limits the use of optical sensors.
• To evaluate the effectiveness of deep learning, specifically Recurrent Neural Networks (RNNs), in exploiting temporal dependencies in multi-temporal SAR data for land cover classification.
• To compare RNN-based models (LSTM and GRU) against classical machine learning methods (SVM and Random Forest) for classifying five winter vegetation quality classes.
• To assess the impact of temporal correlation in Sentinel-1 SAR time series on classification accuracy, particularly for classes with overlapping or ambiguous temporal profiles.

Proposed method

• Multi-temporal Sentinel-1 SAR data (13 acquisitions from October 2016 to February 2017) in VV and VH dual-polarization were used, with a 12-day revisit cycle.
• Preprocessing included coregistration, radiometric calibration, and five-look intensity estimation to reduce speckle noise while preserving spatial resolution.
• Temporal filtering was applied to enhance radiometric resolution without degrading spatial detail, improving signal-to-noise ratio in time series data.
• Two deep learning architectures—LSTM and GRU—were trained end-to-end on the time series intensity data to model temporal dynamics and classify vegetation quality into five classes.
• A 5-fold cross-validation strategy was employed to evaluate model generalization, with performance measured using F-measure, accuracy, and Kappa statistics.
• Confusion matrices and per-class F-measure analysis were used to assess misclassification patterns, particularly between low, very low, and bare soil classes.

Experimental results

Research questions

• RQ1 Can deep recurrent neural networks effectively model temporal dependencies in multi-temporal Sentinel-1 SAR data for winter vegetation quality classification?
• RQ2 How do RNN-based models (LSTM and GRU) compare in performance to classical machine learning methods (SVM and Random Forest) on this classification task?
• RQ3 Why do traditional classifiers struggle with distinguishing between low and very low vegetation quality classes, and can RNNs overcome this limitation?
• RQ4 To what extent do temporal correlations in SAR time series improve classification accuracy, especially for classes with overlapping or ambiguous temporal profiles?

Key findings

• The GRU-based RNN achieved the highest F-measure of 99.05%, followed by LSTM with 98.83%, significantly outperforming both SVM (91.22%) and Random Forest (91.77%).
• The RNN models demonstrated consistent performance across all five vegetation quality classes, including low and very low, which were frequently misclassified by traditional methods.
• Both SVM and Random Forest exhibited the highest accuracy on the High (class 4) and lowest on the Low (class 2) class, indicating poor generalization for ambiguous temporal profiles.
• Misclassification between Low (2) and Bare soil (5) classes was common across all models, but RNNs reduced this error compared to classical classifiers.
• The confusion matrices revealed that RNNs better discriminated between Low and Very Low classes, which are temporally similar, due to their ability to model sequential dependencies.
• The study confirms that RNNs are well-suited for SAR time series classification because they explicitly leverage temporal correlations, which standard classifiers ignore.