پيش بيني كاربري اراضي با استفاده از GIS و شبكه هاي عصبي در مرحله كارون جنوبي،تهران

Predicting urban land uses using GIS and neural networks in southern Karoon, Tehran

اين مقاله روشي نوين براي پيشبيني كاربري اراضي در مرحله كارون جنوبي در منطقه 10 تهران با استفاده از GIS و بر اساس شبكه هاي عصبي مصنوعي ارايه مي دهد.در اين راستا از سه نوع الگوريتم پس انتشار :-آموزش دسته اي كاهش شيب تدريجي با نرخ يادگيري انطباق پذير(Batch gradient descent with momentum and adaptive learning rate)- پس انتشار ارتجاعي ( Resilient back propagation) - Levenberg-Marquardt براي آموزش شبكه هاي عصبي چند لايه استفاده مي شود و با تغيير لايه هاي و نرون ها سعي در يافتن كاراترين شبكه در راستاي اهداف اين تحقيق مي گردد.براي آموزش شبكه،اطلاعات مربوط به هر قطعه زمين شناخت مساحت،مجاورت يا عدم مجاورت با خيابان هاي اصلي، فاصله تا كاربري هاي مختلف موجود در منطقه و كاربري آن در محيط GIS استخراج شدند.درنهايت روش آموزش دسته اي كاهش شيب تدريجي با نرخ يادگيري انطباق پذير براي شبكه اي با يك لايه مخفي شامل 10 نرون، به 60% از آزمايش ها پاسخ صحيح داده و بالاترين كارايي را كسب كرد. در ادامه از اين شبكه براي پيشبيني كاربري آينده اراضي باير در محدوده مطالعاتي استفاده شد.

This paper presents a new approach to predict the land use of parcels in southern Karoon parish, district 10, Tehran, Iran, using GIS and based on Artificial Neural Networks (ANNs). In this context, three types of back-propagation algorithms: 1) Batch gradient descent with momentum and adaptive learning rate, 2) Resilient backpropagation and 3) Levenberg-Marquardt were used for training of multi-layers ANNs with different input data. This paper tries to find the most efficient network for the purpose of this paper with changing the number of layers and neurons, training algorithm and number of input data. The information regarding to each parcel includes area, neighbourhood to main streets and distances to different land uses. This information together with the parcel’s land use (considered as the output of the designed ANN), were extracted using GIS and then were used to train the network. 100 parcels with various land uses were selected randomly. Half of them were used for training and others for evaluation. The results showed that the batch gradient descent with momentum and adaptive learning rate algorithm for a network with one hidden layer including 10 neurons estimates 60% of the tests correctly and obtained the most efficiency. The input data for this network were area, neighbourhood to main streets and distance to commercial areas. The network performance was reduced by increasing the input data for parcels that indicates the noise effect of those data on the network performance. In addition, this network was used to predict the future land uses for existing sterile parcels in the study area. The efficiency of this network could be enhanced by adding