IIIT Hyderabad Publications |
|||||||||
|
Prediction of vegetation dynamics using NDVI time series data and LSTMAuthor: Devireddy Sushma Reddy Date: 2021-05-29 Report no: IIIT/TH/2021/53 Advisor:Ramachandra Prasad Pillutla AbstractThe structure of vegetation changes rapidly due to various reasons influenced by both anthropogenic and natural factors, potentially resulting in degradation and deforestation. Understanding and analyzing the changes in vegetation cover is very important in several aspects including climatic changes, water budget, ecological balance and especially to undertake necessary conservation measures. Spatiotemporal mapping and monitoring of vegetation using remote sensing satellite data help in assessing the health of vegetation and imply better management practices to safeguard them. Further, the development of vegetation indices using remote sensing satellite data provides better insights about the growth patterns, seasonal changes and health conditions of the vegetation. Normalized Difference Vegetation Index (NDVI) is one of the most widely acknowledged indices for vegetation related studies. The generalized annual NDVI profile for vegetation increases with the plant growth and reaches a peak or plateau. Later the profile falls off eventually with plant death or leaf senescence / seasonal changes. Thus, the NDVI series provides a means to describe plant phenology. To interpret changes in vegetation along with the seasonal and inter-annual changes, time series of NDVI data is used in the current study. The concept of neural network has gained much significance in the analysis of vegetation dynamics using remote sensing satellite data. After exploring various approaches available for prediction of vegetation dynamics in the literature and their drawbacks this study has proposed the application of Long-short term memory (LSTM) network, a variant of Recurrent Neural Network using MODIS NDVI time series data sets to assess vegetation dynamics. In the current study various experiments were carried out to demonstrate the potentiality of LSTM network in predicting vegetation dynamics. The first experiment includes a dataset of 861 7-day composite MODIS NDVI images from January 2000 to June 2016 for making the time series. The data is segregated into three sets as the training set (70%), validation set (20%), and testing set (10%). To check the reliability of the experiment two different regions were selected after extensive research in the current investigation. These include different terrains in the Great Nicobar Island, one region along the coast where vegetation has severe ecological damage due to the 2004 Indian Ocean tsunami and the other, an interior region that remained imperturbable during the tsunami. LSTM network is trained with the NDVI values for both the regions separately to predict the future vegetation dynamics. To measure the accuracy of the LSTM network, root mean square error is calculated. The resulting plots for both the experiments indicate that the LSTM neural network follows the series in addition to coinciding with the required time series. Also, an unanticipated change in the trend of the NDVI series was well adapted by the network and was able to predict future NDVI values with good accuracy maintaining RMSE less than 0.03 without providing any supplementary data. To further investigate the robustness of the proposed prediction technique to variations in data composite periods, noise and the study area two additional experiments are executed. The results obtained by using 16-day composite MOD13Q1 datasets containing 455 images of Great Nicobar Island suggested that the composite period has no effect on the performance of LSTM network when there is fairly enough data to train the network. To analyze the influence of noise in the time series on LSTM network, three noise reduction techniques (Fast Fourier Transform, Mean Value Iteration and Whittaker smoother) which have different noise reduction levels are selected. The results have indicated that FFT based approach had the least error rate as it is best at smoothing the time series of all the three methods and MVI had high error rates. Although FFT had least error rate, the original time series was not preserved well by the approach and the series was over smoothed. MVI performance was relative to the noise levels in the original series as this approach is based on the mean of adjacent values. Whittaker smoother on the other hand was able to smooth the series reasonably well and preserve true values. The RMSE values were observed to be around 0.02 for this approach in general. It is further understood that LSTM is independent of the properties of the study area by conducting the above experiments in Uttarakhand region that has different climatic, vegetation and topological conditions from. Further, to comprehend if a complex model like LSTM is required for vegetation studies and whether LSTM performs better than the vegetation predictions techniques in practice, all the above experiments are computed with two popularly used prediction techniques Artificial Neural Networks (ANN) and Support Vector Regression (SVR). Results showed that LSTM outperforms other models in comparison. This deduction is further validated by evaluating the performance of the three techniques in predicting vegetation with longer time intervals. By adopting the prescribed method in this thesis, the anticipation of vegetation changes can be done accurately much ahead of time and take proactive measures accordingly to safeguard and improve the vegetation in any area. Full thesis: pdf Centre for Spatial Informatics |
||||||||
Copyright © 2009 - IIIT Hyderabad. All Rights Reserved. |