A deep learning approach for forecasting non-stationary big remote sensing time series
- PDF / 4,827,348 Bytes
- 11 Pages / 595.224 x 790.955 pts Page_size
- 103 Downloads / 203 Views
2ND CAJG 2019
A deep learning approach for forecasting non-stationary big remote sensing time series Manel Rhif1
· Ali Ben Abbes1 · Beatriz Martinez2 · Imed Riadh Farah1
Received: 13 May 2020 / Accepted: 15 October 2020 © Saudi Society for Geosciences 2020
Abstract Remote sensing (RS) data are undergoing an explosive growth. In fact, RS data are regarded as RS big data which generates several challenges such as data storage, analysis, applications, and methodologies. In this paper, a suitable method to forecast the Normalized Difference Vegetation Index (NDVI) time series (TS) from RS big data is introduced. In fact, we propose a non-stationary NDVI TS forecasting model by combining big data system, wavelet transform (WT), long short-term memory (LSTM) neural network. In the first step, the MapReduce algorithm was investigated for RS data storage and NDVI TS extraction. Then, the WT was used to decompose the TS into different components. Finally, LSTM was used for NDVI TS forecasting. Additionally, we have compared the forecasting results using only LSTM, recurrent neural network (RNN), and WT-RNN. Our results show that the proposed methodology using WT-LSTM model provides us an efficient method for forecasting NDVI TS in terms of root mean square error (RMSE) and Pearson correlation coefficient (R). Finally, we have evaluated the performance of the big data model. Keywords Remote sensing · Vegetation · Non-stationary time series · Big data · Deep learning · Wavelet transform
Introduction Nowadays, land cover faces several change due to the pollution, environmental change, etc. Recently, a large and open set of remote sensing (RS) data is available to track changes over time and forecast the future. Specifically, the Normalized Difference Vegetation Index (NDVI) (Tucker 1979) time series (TS) proves its effectiveness for vegetation analysis and forecasting in several works (Reddy and Prasad 2018; Andrea et al. 2019).
This paper was selected from the 2nd Conference of the Arabian Journal of Geosciences (CAJG), Tunisia 2019 Responsible Editor: Biswajeet Pradhan Manel Rhif
[email protected] 1
Laboratoire RIADI, Ecole Nationale des Sciences de l’Informatique, Mannouba, Tunisia
2
Departament de F´ısica de la Terra i Termodin`amica, Universitat de Valencia, Val`encia, Spain
The large availability of RS data confers the RS data as a proper big data (BD) resource. Therefore, RS-BD (remote sensing big data) data can be characterized by three BD features (3V): volume, variety, and velocity (Chi et al. 2016). However, a big volume of RS data is available. For example, from NASA archives, the Earth Science Data and Information System (ESDIS) holds 7.5 PB of data with nearly 7000 unique datasets and 1.5 million users in 2013 (Ramapriyan et al. 2013). Additionally, a variety of data are available from different sensors (Lidar, radar, etc.), multi-temporal and multi-resolution data. Also, different TS can be extracted using different indices depending on the application such as NDVI and Enhanced Vegetation Index
Data Loading...
