Loading...

Spatial-temporal Variation of Urmia Lake Basin Using Artificial Intelligence Algorithms

Novin, Soroush | 2021

2279 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 54746 (09)
  4. University: Sharif University of Technology
  5. Department: Civil Engineering
  6. Advisor(s): Torkian, Ayoub
  7. Abstract:
  8. Water shortages resulting from macro-environmental climate changes as well as local inefficient agricultural practices and dam constructions activities have resulted in the gradual reduction of water level in Urmia Lake, located in the northwest of Iran. As such, restoration efforts were initiated to prevent further adverse impacts exacerbating the conditions and creating secondary problems such as regional salt dust generation and dispersion, resulting in health issues for the greater area population in the neighboring vicinities. The utilization of advanced forecast modeling based on deep learning algorithms can assist the authorities to manage better multi-dimensional issues affecting the agricultural industry as well as the health conditions of the general population. The dataset consisted of monthly water levels and local environmental conditions from 2000 to 2016. Vegetation variation data were extracted from MODIS satellite images. Two different recurrent neural networks (RNN) configurations of Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) were used to receive input data from a six-by-five matrix trained and tested on a 70-30 basis. Three layers were utilized using the dropout layer and linear activation function. Input variables consisted of water level, land surface temperature, precipitation, runoff, and NDVI index, and the monthly average water level was used as the output. Root mean square error (RMSE), R-squared (R2), and mean absolute error (MAE) were used as performance criteria of the networks. Results indicated improved performance monthly and yearly water level predictions with an increasing number of input variables for both configurations. Optimized versions for GRU and LSTM networks produced 5.1%, 5.6%, 94.3% and 10.6%, 13.7%, 66.8% for MAE, RMSE and R2, respectively for 12-month future predictions. Better performance of GRU memory cells seems to be related to the less complex architecture. GRUs have fewer parameters and thus train a bit faster; generalization is also more consistent with short sequences used in datasets such as this study. Based on the results of this research, it seems special attention should be directed towards improved agricultural practices as well switching to farming products with fewer water requirements. More water should also be released from dams to the lake, and the construction of new dams and wells in the basin should be curtailed
  9. Keywords:
  10. Urumieh Lake ; Recurrent Neural Networks ; Time Series Prediction ; Normalized Difference Vegetation Index ; Urmia Lake Level ; Google Earth Engine ; Water Level Fluctuations

 Digital Object List

 Bookmark

...see more