برآورد تبخیر تعرق مرجع روزانه با استفاده از روش جنگل تصادفی بهینه‌سازی شده با الگوریتم ژنتیک

تبخیر تعرق یک متغیر مهم در فعل و انفعالات بین خاک، پوشش گیاهی، جو، انرژی سطح زمین و آب است. از طرفی، اندازهگیری آن از طریق روشهای مستقیم، هزینه و زمان زیادی میطلبد. هدف از پژوهش حاضر، بررسی توانایی روش جنگل تصادفی (rf) در دو حالت منفرد و بهینه شده با الگوریتم ژنتیک (rf ga) می‌باشد. بدین منظور، داده‌های روزانه برخی از متغیرهای هواشناسی اثرگذار بر پدیده تبخیر تعرق در دوره آماری 20 ساله (1400 1380) در سه ایستگاه تبریز، سراب و مراغه واقع در استان آذربایجان شرقی جمعآوری شد. سپس، شش سناریو ترکیبی از متغیرهای هواشناسی برای واسنجی و صحت‌سنجی مدل‌های مذکور مد نظر قرار گرفتند. علاوه براین، عملکرد سه گروه از روشهای تجربی برآورد کننده تبخیر تعرق مرجع نیز مورد بررسی قرار گرفت. در نهایت، با استفاده از معیارهای آماری کارایی روش‌ها مورد ارزیابی قرار گرفت. نتایج نشان داد که به‌منظور تخمین et0 با استفاده از متغیرهای هواشناسی کمتر سناریو 4 با شاخص پراکندگی 0.131 در ایستگاه تبریز، 0.171 در ایستگاه سراب و 0.134 در ایستگاه مراغه دقت بالایی دارد. همچنین سناریو 2 در ایستگاه‌های تبریز، سراب و مراغه به ترتیب با شاخص پراکندگی0.184، 0.220 و0.172 با دقت قابل قبولی می‌تواند مورد استفاده قرار گیرد. در حالت مقایسه جنگل تصادفی عملکرد بهتری نسبت به روشهای تجربی در ایستگاه‌های مورد مطالعه نشان داد. در نهایت، استفاده از روش جنگل تصادفی بهمنظور برآورد دقیقی از میزان تبخیر تعرق مرجع در استان آذربایجان شرقی پیشنهاد گردید.

estimation of daily reference evapotranspiration using random forest optimized by genetic algorithm

background and objectives: reference evapotranspiration (et0) is an important parameter in the interactions among soil، vegetation، atmosphere، surface energy and water. direct measurement of evapotranspiration values is costly and time consuming. on the other hand، modeling this complex process in which many variables interact with each other is not feasible without considering multiple assumptions. in this regard، the fao penman monteith method is used in a wide range of climatic and environmental conditions. one of the weaknesses of fao penman monteith method is its dependence on various meteorological parameters. therefore، it is necessary to implement methods with lower meteorological variables that can estimate et0 with suitable accuracy. thus، in the present study، an attempt was made to estimate et0 with acceptable accuracy using machine learning models. methodology: in the present study، daily meteorological parameters in the time period of 2000 2020 including maximum and minimum air temperature (tmax، tmin)، mean temperature (t)، wind speed (u2)، average relative humidity (rh)، maximum and minimum relative humidity (rhmax، rhmin) and sunshine hours (n) were obtained on a daily basis in three stations of east azerbaijan province (tabriz، sarab، and maragheh). moreover، six scenarios were defined as input combinations. then، using random forest (rf) method in two cases: single random forest and using the genetic algorithm (ga) to optimize its effective parameters with considering the fao penman monteith model as a basis، the machine learning models were calibrated and validated for estimating et0 values at studied stations. furthermore، the performance of empirical equations in three groups based on temperature (hargreaves، blaney criddle and romanenko)، radiation (irmak) and mass transfer (meyer) were also investigated. it should be noted that 75% of the data were considered for calibration and 25% for the validation of machine learning methods. finally، using the statistical criteria of correlation coefficient (cc)، scattered index (si) and willmott's index of agreement (wi)، a suitable machine learning method was introduced to estimate the reference evapotranspiration. also، the most suitable combination of meteorological parameters for et0 estimation was suggested. findings: the obtained results showed that in all studied stations، scenario 6 has the best performance، either in the case of single random forest (rf) or in the case of random forest optimized by genetic algorithm (ga rf). meteorological parameters of this scenario include minimum and maximum air temperature، minimum and maximum relative humidity، sunshine hours and wind speed. by optimizing the rf 6 parameters with the genetic algorithm at tabriz station، the statistical criteria were improved (cc from 0.990 to 0.991، si from 0.103 to 0.098). at sarab station، the cc was increased from 0.980 to 0.982، the si was decreased from 0.140 to 0.132 and the wi was increased from 0.989 to 0.990. at maragheh station، cc was increased from 0.990 to 0.991، si was decreased from 0.103 to 0.098 and wi remained unchanged at 0.995. in general، the decreasing trend of the scattered index for rf method from scenarios 1 to 6 can be understood by increasing the input parameters of the random forest method. among the three groups of empirical methods based on air temperature، radiation and mass transfer for estimating et0، the best performance was seen for the blaney criddle method based on air temperature. in all studied stations، the ga rf model showed better performance than the empirical methods. also، ga rf 5 with similar meteorological parameters with blaney criddle method provided accurate et0 estimations.conclusion: determining the amount of daily evapotranspiration and consequently accurate estimation of water requirement of plants provide the basis for proper designing of irrigation systems by reducing installation costs and providing a suitable program for the use of water resources in the agriculture sector. so، in the present study، meteorological data from tabriz، sarab and maragheh stations were used to evaluate the ability of machine learning methods including rf and ga rf to estimate the values of reference evapotranspiration. the results showed the high accuracies of rf 6 and ga rf 6 for all studied stations and belany criddel among the empirical models. in a more detailed look، the genetic algorithm had positive effects on increasing the model accuracies by reducing scattered index of ga rf scenarios 1، 4، 5 and 6 in tabriz and maragheh stations as well as scenarios 1، 5 and 6 at sarab station. finally، it can be concluded that both rf and ga rf models provided the most accurate estimates of daily reference evapotranspiration in the east azerbaijan province.