Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Prediction of Sentinel-2 multi-band imagery with attention BiLSTM for continuous earth surface monitoring

Weiying Zhao, Natalia Efremova

TL;DR

The paper tackles the problem of maintaining continuous remote sensing records by predicting missing Sentinel-2 data under cloud cover. It introduces an attention-enhanced BiLSTM with a time-difference input channel to perform sequence-to-one forecasts across multiple bands, including NDVI and other vegetation indices. Empirical results show the model superiorly predicts NDVI, several vegetation indices, and all Sentinel-2 bands compared with LSTM-based and ConvLSTM baselines, demonstrating robustness under cloud contamination. This approach can substantially improve data continuity for crop monitoring and environmental surveillance, enabling reliable assessments on user-defined dates even in adverse weather conditions.

Abstract

Continuous monitoring of crops and forecasting crop conditions through time series analysis is crucial for effective agricultural management. This study proposes a framework based on an attention Bidirectional Long Short-Term Memory (BiLSTM) network for predicting multiband images. Our model can forecast target images on user-defined dates, including future dates and periods characterized by persistent cloud cover. By focusing on short sequences within a sequence-to-one forecasting framework, the model leverages advanced attention mechanisms to enhance prediction accuracy. Our experimental results demonstrate the model's superior performance in predicting NDVI, multiple vegetation indices, and all Sentinel-2 bands, highlighting its potential for improving remote sensing data continuity and reliability.

Prediction of Sentinel-2 multi-band imagery with attention BiLSTM for continuous earth surface monitoring

TL;DR

The paper tackles the problem of maintaining continuous remote sensing records by predicting missing Sentinel-2 data under cloud cover. It introduces an attention-enhanced BiLSTM with a time-difference input channel to perform sequence-to-one forecasts across multiple bands, including NDVI and other vegetation indices. Empirical results show the model superiorly predicts NDVI, several vegetation indices, and all Sentinel-2 bands compared with LSTM-based and ConvLSTM baselines, demonstrating robustness under cloud contamination. This approach can substantially improve data continuity for crop monitoring and environmental surveillance, enabling reliable assessments on user-defined dates even in adverse weather conditions.

Abstract

Continuous monitoring of crops and forecasting crop conditions through time series analysis is crucial for effective agricultural management. This study proposes a framework based on an attention Bidirectional Long Short-Term Memory (BiLSTM) network for predicting multiband images. Our model can forecast target images on user-defined dates, including future dates and periods characterized by persistent cloud cover. By focusing on short sequences within a sequence-to-one forecasting framework, the model leverages advanced attention mechanisms to enhance prediction accuracy. Our experimental results demonstrate the model's superior performance in predicting NDVI, multiple vegetation indices, and all Sentinel-2 bands, highlighting its potential for improving remote sensing data continuity and reliability.
Paper Structure (8 sections, 5 figures, 1 table)

This paper contains 8 sections, 5 figures, 1 table.

Table of Contents

  1. Introduction
  2. Method
  3. Experimental results and discussion
  4. Dataset
  5. NDVI forecasting
  6. Multi vegetation indices forecasting
  7. Sentinel-2 bands forecasting
  8. Conclusion

Figures (5)

  • Figure 1: Flowchart of the vanilla attention BiLSTM. The processing flow in this figure takes 5 time series bands as an example. The additional times series, which is highlighted in red, is the time difference time series.
  • Figure 2: Time difference distribution. The time difference is between the target image acquisition time and the previous 5 image acquisition time.
  • Figure 3: Scatter plot of predicted and real NDVI values. The predicted image is unseen during the model preparation.
  • Figure 4: Multibands time series mean and standard deviation across different bands. (up) Original time series. (bottom) Time series vegetation indices. Most of the cloud-affecting Sentinel-2 images have been removed based on the NDSI band.
  • Figure 5: Scatter plot of predicted and real vegetation indices.