پيش‌بيني خشك‌سالي با بهره‌گيري از مدل تركيبي ماشين بردار پشتيبان موجكي و شاخص SPI (مطالعه موردي: حوضه درياچه اروميه-ايران)

Drought Forecasting Using Wavelet - Support Vector Machine and Standardized Precipitation Index (Case Study: Urmia Lake-Iran)

زﻣﯿﻨﻪ و ﻫﺪف: ﺧﺸﮏ ﺳﺎﻟﯽ ﺗﻬﺪﯾﺪي ﺟﺪي ﺑﺮاي اﻧﺴﺎن و ﻣﺤﯿﻂ زﯾﺴﺖ ﺑﻮده ازاﯾﻦ رو ﯾﺎﻓﺘﻦ ﺷﺎﺧﺼﯽ ﺟﻬﺖ ﭘﯿﺶﺑﯿﻨﯽ اﯾﻦ ﭘﺪﯾﺪه از اﻫﻤﯿﺖ ﺑﻪﺳﺰاﯾﯽ ﺑﺮﺧﻮردار اﺳﺖ. ﺷﺎﺧﺺ ﺑﺎرش اﺳﺘﺎﻧﺪاردﺷﺪه )SPI( ﯾﮏ ﺷﺎﺧﺺ ﺟﺎﻣﻊ ﺟﻬﺖ ﻃﺒﻘﻪ ﺑﻨﺪي ﺷﺪت ﺧﺸﮏ ﺳﺎﻟﯽ ﺑﻪ ﺣﺴﺎب ﻣﯽآﯾﺪ. ﻣﺪلﻫﺎي ﻫﻮش ﻣﺼﻨﻮﻋﯽ ﮐﻼﺳﯿﮏ از ﻣﺘﺪاولﺗﺮﯾﻦ ﻣﺪلﻫﺎﯾﯽ ﻫﺴﺘﻨﺪ ﮐﻪ ﺟﻬﺖ ﭘﯿﺶﺑﯿﻨﯽ ﺷﺎﺧﺺ SPI ﻣﻮرداﺳﺘﻔﺎده ﻗﺮارﮔﺮﻓﺘﻪ اﻧﺪ. ازآن-ﺟﺎﯾﯽ ﮐﻪ اﯾﻦ ﻣﺪلﻫﺎ ﺑﺮ ﭘﺎﯾﻪي وﯾﮋﮔﯽ ﺧﻮدﻫﻢﺑﺴﺘﮕﯽ اﺳﺘﻮار ﻫﺴﺘﻨﺪ، ﺑﻨﺎﺑﺮاﯾﻦ ﺗﻮاﻧﺎﯾﯽ رﺻﺪ ﻧﻤﻮدن ﺳﺮيﻫﺎي زﻣﺎﻧﯽ درازﻣﺪت و ﻓﺼﻠﯽ را دارا ﻧﻤﯽﺑﺎﺷﻨﺪ. در اﯾﻦ ﭘﮋوﻫﺶ ﺑﺮاي ﭘﯿﺶﺑﯿﻨﯽ ﺧﺸﮏ ﺳﺎﻟﯽ از ﻣﺪل ﺗﺮﮐﯿﺒﯽ ﻣﺎﺷﯿﻦ ﺑﺮدار ﭘﺸﺘﯿﺒﺎن ﻣﻮﺟﮑﯽ و ﺷﺎﺧﺺ SPI اﺳﺘﻔﺎده ﺷﺪه اﺳﺖ. روش ﺑﺮرﺳﯽ: ﺑﺮاي اﯾﻦ ﻣﻨﻈﻮر ﺳﺮي زﻣﺎﻧﯽ ﺷﺎﺧﺺ SPI ﻣﺮﺑﻮط ﺑﻪ ﺣﻮﺿﻪ اروﻣﯿﻪ ﺗﻮﺳﻂ آﻧﺎﻟﯿﺰ ﻣﻮﺟﮏ ﺑﻪ ﭼﻨﺪﯾﻦ زﯾﺮ ﺳﺮي ﺑﺎ ﻣﻘﯿﺎسﻫﺎي زﻣﺎﻧﯽ ﻣﺨﺘﻠﻒ ﺗﺒﺪﯾﻞ ﺷﺪه و اﯾﻦ زﯾﺮ ﺳﺮي ﻫﺎي زﻣﺎﻧﯽ ﺑﻪ ﻋﻨﻮان ورودي ﻣﺪل ﻣﺎﺷﯿﻦ ﺑﺮدار ﭘﺸﺘﯿﺒﺎن ﺑﺮاي ﭘﯿﺶﺑﯿﻨﯽ ﺧﺸﮏ ﺳﺎﻟﯽ در ﻧﻈﺮ ﮔﺮﻓﺘﻪ ﻣﯽﺷﻮﻧﺪ. ﯾﺎﻓﺘﻪ ﻫﺎ: ﻧﺘﺎﯾﺞ ﺣﺎﺻﻞ از ﺻﺤﺖ ﺳﻨﺠﯽ ﻣﺪلﻫﺎ ﺑﯿﺎنﮔﺮ آن اﺳﺖ ﮐﻪ ﺑﯿﺶﺗﺮﯾﻦ ﻣﻘﺪار ﺿﺮﯾﺐ ﺗﺒﯿﯿﻦ و ﮐﻢﺗﺮﯾﻦ ﻣﻘﺪار ﺟﺬر ﻣﯿﺎﻧﮕﯿﻦ ﻣﺮﺑﻊ ﺧﻄﺎ ﺑﺮاي ﻣﺪل ﻣﻨﻔﺮد ﻣﺎﺷﯿﻦ ﺑﺮدار ﭘﺸﺘﯿﺒﺎن ﺑﻪ ﺗﺮﺗﯿﺐ 0/865 و 0/237 و ﺑﺮاي ﻣﺪل ﺗﺮﮐﯿﺒﯽ ﻣﺎﺷﯿﻦ ﺑﺮدار ﭘﺸﺘﯿﺒﺎن ﻣﻮﺟﮑﯽ ﺑﻪ ﺗﺮﺗﯿﺐ 0/954 و 0/056 ﻣﯽﺑﺎﺷﺪ. ﺑﺤﺚ و ﻧﺘﯿﺠﻪ ﮔﯿﺮي: ﺑﻨﺎﺑﺮاﯾﻦ ﻣﺪل ﺗﺮﮐﯿﺒﯽ ﻣﺎﺷﯿﻦ ﺑﺮدار ﭘﺸﺘﯿﺒﺎن ﻣﻮﺟﮑﯽ در ﻣﻘﺎﯾﺴﻪ ﺑﺎ ﻣﺪل ﻣﻨﻔﺮد ﻣﺎﺷﯿﻦ ﺑﺮدار ﭘﺸﺘﯿﺒﺎن ﺗﻮاﻧﺎﯾﯽ ﺑﻪ-ﺳﺰاﯾﯽ ﺟﻬﺖ ﭘﯿﺶﺑﯿﻨﯽ ﺳﺮي زﻣﺎﻧﯽ ﺷﺎﺧﺺ SPI و ﻧﯿﺰ رﺻﺪ ﻧﻤﻮدن ﻧﻘﺎط ﺑﯿﺸﯿﻨﻪ اﯾﻦ ﺳﺮي زﻣﺎﻧﯽ ﺑﻪ ﺳﺒﺐ در ﻧﻈﺮ ﮔﺮﻓﺘﻦ ﺗﻐﯿﯿﺮات ﻓﺼﻠﯽ دارا ﻣﯽﺑﺎﺷﺪ. از ﺳﻮﯾﯽ ﻧﺸﺎن داده ﺷﺪ ﮐﻪ اﯾﻦ ﻣﺪل ﺗﺮﮐﯿﺒﯽ در ﻣﻘﺎﯾﺴﻪ ﺑﺎ ﺳﺎﯾﺮ ﻣﺪلﻫﺎي ﺧﻮدﻫﻢﺑﺴﺘﻪ ﮐﻼﺳﯿﮏ ﻫﻢﭼﻮن ﺷﺒﮑﻪ ﻋﺼـﺒﯽ ﻣﺼﻨﻮﻋﯽ از دﻗﺖ و ﮐﺎراﯾﯽ ﺑﺎﻻﺗﺮي ﺑﺮﺧﻮردار اﺳﺖ.

Background and Objectives: Drought is regarded as a serious threat for people and environment. As a result, finding some indices to forecast the drought is an important issue that needs to be addressed urgently. The appropriate and flexible index for drought classification is the Standardized Precipitation Index (SPI). Artificial intelligence models were commonly used to forecast SPI time series. These models are based on auto regressive property. So, they are not able to monitor the seasonal and long-term patterns in time series. In this study, the Wavelet-Support Vector Machine (WSVM) approach was used for the drought forecasting through employing SPI. Method: In this way, the SPI time series of Urmia Lake watershed was decomposed to multiple frequent time series by wavelet transform; then, these time series were imposed as input data to the Support Vector Machine (SVM) model to forecast the drought. Findings: The results showed that, the maximum value of R2 and minimum value of RMSE indexes for SVM model are 0.865 and 0.237 and for WSVM model are 0.954 and 0.056 respectively in verification step. Discussion and Conclusion: So, the propounded hybrid model has superior ability in forecasting SPI time series comparing with the single SVM model and also it can accurately assess the extreme data in SPI time series by considering the seasonality effects. Finally, it was concluded that, the proposed hybrid model is relatively more appropriate than classical autoregressive models such as ANN.