بهبود شناسايي تغييرات زمين هاي زير كشت با تاكيد بر استفاده از تصوير هم مرجع شده چندزمانه و روش تركيب رنگي

An Approach for Improving Change Detection in Agricultural Lands Using Georeferenced Multi-Temporal Image and Color Fusion Method

در این تحقیق تعیین روال مناسب جهت شناسایی تغییرات زمینهای زیر كشت و برآورد درصد آن تغییرات در یك منطقه نیمه شهری با تكیه بر الگوریتم های شبكه عصبی و SVM بصورت نظارت شده مورد بررسی قرار گرفته است. برای این منظور از تصاویر چندزمانه سنجنده های Landsat استفاده شده است. در واقع در این روش برخلاف روشهای رایج، برای شناسایی تغییرات از مقایسه تصاویر استفاده نشده است بلكه هدف كشف تغییرات با استفاده توام از روش تركیب رنگی و الگوریتم های طبقه­بندی می­باشد. ایده اصلی به این طریق است كه با استفاده از دو تصویر، یك تصویر هم مرجع شده چندزمانه تهیه و سپس با استفاده توام از روش تركیب رنگی جمعی(تركیب كردن لایه های رنگی تصاویر) و روشهای تشخیص الگو و اعمال آنها بر روی تك تصویر هم مرجع شده بدست آمده، نواحی مناطق تغییر یافته و تغییر نیافته ایجاد و درنهایت با انتخاب نمونه های آموزشی فقط از یك تصویر، الگوریتم بر روی تصویر تلفیقی اعمال و نقشه تغییرات نهایی بدست آمده است. نتایج حاصل شده حاكی از آن است كه این روش از جهت كاهش نمونه های آموزشی، افزایش دقت (تقریبا 3 درصد) و سرعت عمل شناسایی تغییرات بر روشهای معمول كه از مقایسه دو تصویر چندزمانه با هم استفاده می­كنند برتری دارد.

During the time, landcover and associated landuse patterns are changing very fast and the human factors play a major role in such drastic changes. Scientists have formerly attempted to identify the landuse altering processes and related environmental impacts. In the recent studies, evaluating the agricultural outputs and arable lands is regarded so important that their organization and management can be mentioned as a very critical factor for all countries. Nowadays, satellite images could be accurately processed, as an advanced technique in remote sensing, to determine the environment changes in a particular object of study between two or more time periods.Therefore, favorable results are achieved in recognition pattern-based remote sensing methods in order to achieve these global aims. Although there are various methods in photography and remote sensing for dealing with revealing changes, none of them can be considered as an optimum one completely. In the present paper an appropriate Supervised approach was proposed in odrer to identify changes in semi-urban areas based on both neural network algorithms and SVM (Suport Vector Matchene). To achieve this purpose, Landsat7 multi-temporal images are applied. In principle, this method unlike conventional ones, is not introduced to identify changes, but our method can be addressed as determining changes without comparing multi-temporal single-source images with each other and principally relying on color fusion ( fusing colors in different bands and creating a different color ) in the resulted single image which contains all the layers of two multi-temporal images. The main basic idea is to produce a multi-temporal single-source image using two images and then using color fusion and pattern recognition methods on the georeferenced single-sourced image, afterward, was produced a map for changed and unchanged regions, finally, was applied algorithm on image to provide the final change map. Simplicity and increases performance can be proposed as the advantages of this method. In fact, mixed collective color (color fusion) method with pattern recognition methods or classification methods and using them for rsulted reference single image the basis of this method in order to identify the changed and unchanged zones. Finally, our main idea was based on that after selecting training data from one single ( common data in both images ), use training data in unmodified and stable zones and remove the data which located in changed zone. In practice, after revealing the modified zones showing an overlap with training set data, the existing data in mentioned zones were removed. Finally, applying training data and conventional classification methods such as SVM and neural networks classes were identified and introduced and final map of changes developed. Our achieved results suggest that this approach is far better than traditional methods and significantly reduces training samples and increases accuracy (2.5 – 3 percent), pace and spectral information for performed classification.