پهنه‌بندی ریسک آتش‌سوزی مناطق جنگلی با استفاده از روش تلفیقی شبکه عصبی مصنوعی و سیستم اطلاعات مکانی (مطالعه موردی: منطقه حفاظت شده شیمبار، استان خوزستان)

بهره‌گیری همزمان از سیستم اطلاعات جغرافیایی (gis) و روش‌های مبتنی بر هوش مصنوعی، همواره نتایج خوبی را در تحقیقات حوضه منابع طبیعی به دنبال داشته است. این تحقیق در همین قالب و به منظور اولویت‌بندی عوامل تاثیر‌گذار بر گسترش حریق و شناسایی مناطق پرخطر در جنگل‌های منطقه حفاظت شده شیمبار، بر اساس آتش‌سوزی های سال های 1390 تا 1397 انجام شد که در این خصوص، شاخص هایی برای روش شبکه عصبی مصنوعی در نظر گرفته شد. در پیاده‌سازی روش شبکه عصبی مصنوعی با شاخص های موثر بر آتش‌سوزی جنگل، به تهیه نقشه پهنه‌بندی خطر آتش‌سوزی با پنج کلاس خطر خیلی کم، خطر کم، خطر متوسط، خطر زیاد، خطر خیلی زیاد با صحت کلی 0/83 و خطای rmse برابر با 0/75 پرداخته شد. نتایج تحقیق نشان داد که بیست درصد مساحت منطقه در طبقه متوسط پتانسیل وقوع آتش‌سوزی، یازده درصد در طبقه زیاد و ده درصد در طبقه خیلی زیاد قرار دارد. همچنین مهم‌ترین متغیر‌های موثر بر وقوع آتش‌سوزی شامل فاصله از رودخانه، تیپ اراضی، ارتفاع و حداقل دما است. نتیجه پژوهش این است که با توجه به شاخص های در نظرگرفته شده، مدل‌های تلفیقی شبکه عصبی مصنوعی (ann) و سیستم اطلاعات مکانی، در تهیه نقشه پهنه‌بندی خطر آتش‌سوزی کارایی بالایی دارد و پیشنهاد می‌ شود از این مدل‌ها برای پیشگیری، کنترل و مدیریت آتش‌سوزی در سایر نقاط کشور هم در مقیاس وسیع استفاده شود.  

forest risk fire zoning using an integrated method of artificial neural network and spatial information system (murray study: shimbar protected area)

1- introduction   the diverse ecosystems in iran with their own unique climate and wildlife have witnessed uncontrolled fires annually to the extent that in terms of forest fires, iran ranks fourth among mena countries (naghipoor borj, 2018). in 2017, a total of 252 incidents of wildfire occurred in iran, with damage to 3,006 hectares; while in 2018, 187 wildfires occurred damaging 2,385 hectares (sabzali et al., 2019). the zagros forests cover an expanse of 6 million km2 in the west of iran (approximately %4 of iran’s total land mass) (sadeghi et al., 2017) of which 33,920.2 km2 are located in the khuzestan province (alli mahmoodi sarab et al., 2013). the shimbar mountains are chiefly composed of limestone formations, and only a small area is composed of alluvial deposits. the average annual rainfall in the area is roughly 815 mm, and the average annual temperature is 20-26°c. the average evaporation rate for the area is 2,523 mm. (sharifi et al., 2020). the vegetation cover of the shimbar natural reserve is composed of two types of vegetation: the marshland vegetation cover which is chiefly artemisia as the main vegetation cover, and the mountain vegetation cover which is a type of iranian oak.due to the high security of the wildlife reserve, a variety of mammals thrive in this region such as the iranian squirrel, martens, wolves, wild bear, and the mongoose. birds such as quail, partridge, bee-eaters and woodpeckers are also native to the area (dinarvand et al., 2018).  shimbar region was decreed by national legislature to be among the four areas under the jurisdiction of the national environmental protection agency (nepa). in the early 1940s, the first attempts to apply a logic-based model to simulate fire hazard risks was carried out by warren mcculloch and walter pitts, and this logic model is the basis of all present day artificial neural networks (laurent fast, 2016). in the present study, the underlying reason for the selecting of an artificial neural network was its capability in the creating a relationship between the input and output data for non-linear complex phenomena, its extensive application in the prediction of fire hazards, and its ability to create a model out of the relationship between the number of fires and the factors impelling such fires (ouache et al., 2021 & islami et al., 2020 & polinova et al., 2019 & jaafari goldarq et al., 2013).2- methodologyinitially, the data related to forest fires that had occurred in the period spanning 2011 to 2018 were collected from the andika regional environmental protection agency, and in the next stage, the ground reality maps of these points were prepared. the environmental protection agency had recorded 79 fires in the wildlife preserve; therefore, the data was used for the training and testing of the model used in this study. in order to prepare a map identifying potential fire hazard zones, the factors affecting the forest fires in the region were identified as chiefly being physiological features such as slope, aspect, and elevation. all topographic maps with a scale of 1:25,000 were obtained from the andika regional environmental protection agency, and the national cartographic center. features of the vegetation cover such as soil types, land classification types, vegetation cover, and land use were obtained from landsat 8 images and the complementary data were obtained from the natural resource’s organization of andika. climate features such as relative humidity, wind speed, minimum temperature, maximum temperature, wind direction, amount of rainfall, and average temperature were accessed from the archives of the regional meteorological office. with due regards to the fact that the data values obtained were at various points, and in order to generate data for the whole region, the interpolation functions in arcgis were utilized. anthropological features considered were far from residential areas, road accessibility, and distance from the river. the data for road accessibility were obtained from google earth layer maps and the data for residential areas and distance to rivers were provided by the andika regional environmental protection agency and natural resources offices. the information layers for roads, rivers and residential areas used a vector format; therefore, by using euclidean distance analysis, raster-geomatics with the capability of spatial segregation for the required zones were developed in a way that the value allocated to each cell indicated the distance from the nearest road, the river or residential area. once the features for each of the variables were identified, a spatial map was prepared in a gis environment. 3- results the data used in this study encompass historical data of forest fires occurring from 2011 to 2018. by analyzing the maps created for various parameters such as the forest fire dispersion map, physiological features map, vegetation cover map, meteorological map, anthropological specification map, the results showed that the vegetation cover; the distance to the available bodies of water, and the type of lands are the main factors to be considered. the validation of the model was assessed by utilizing rmse- roc-auc criteria in order to verify the accuracy of the obtained results for determining the extent of potential forest fires with actual events. it was observed that the method proposed in this research has an accuracy of 0.83 in predicting fluctuations along the trend which is relatively high. the data were then transferred to the arc gis software and the zoning map for determining the potential fire hazard areas based on the existing historical data was created and divided into five categories representing very low risk, low risk, moderate risk, high risk, and extremely