توسعه يك روش دانش‌مبناي مكاني براي كشف تغييرات در اراضي زراعي

Developing a Spatial Knowledge-Based Approach to Detect Changes of Cultivation Fields

تشخیص وقوع یا عدم وقوع تغییر در پدیده های زمینی كه باید در حجم عظیمی از داده های مكانی چند زمانه صورت پذیرد، نوعی استدلال زمانی-مكانی است. به طور متداول آشكارسازی تغییرات با استفاده از تصاویر ماهواره ای و با مقایسه مقادیر پیكسلی مربوط به دو تصویر (با زمان های متفاوت)، انجام می گیرد. نتایج این فرآیند علاوه بر وابستگی به نوع روش استفاده‌شده، به نحوه تفسیر نتایج نیز بستگی دارد. در بیشتر روش های كشف تغییرات با استفاده از تصاویر ماهواره ای، تشخیص نوع تغییر، نیازمند نمونه برداری های میدانی و صرف هزینه های بالا می باشد. در این تحقیق روشی دانش مبنا برای كشف تغییرات در اراضی زراعی توسعه داده می شود تا بتواند با دقت بالا، تغییرات اراضی زراعی و نوع آن تغییرات را شناسایی نماید. مجموعه ای از شروط مكانی (هندسی)، زمانی و طیفی، تشكیل‌دهنده پایگاه دانش مكانی موردنیاز در این روش هستند. این شروط با آنالیز سابقه كشت محصولات زراعی، سری زمانی تصاویر ماهواره ای و نقشه مزارع، استخراج‌ می گردد. پیاده سازی روش در اراضی زیر كشت گندم در دشت مغان نشان گر دقت حدود 98 درصدی این روش در كشف تغییر و دقت 94 درصدی در كشف نوع تغییر می باشد؛ كه در مقایسه با نتایج حاصل از پیاده سازی روش های تفاضل شاخص گیاهی نرمال شده و مقایسه تصاویر طبقه بندی شده (Post-Classification Comparison) در منطقه مطالعاتی كه به ترتیب دارای دقت 70 و 81 درصدی هستند، دارای دقت بالایی است. همچنین استفاده از این روش باعث كاهش نیاز به برداشت داده زمینی (به‌منظور مشخص كردن نوع تغییر) و عدم نیاز به تفسیر نتایج گردید.

The dynamics of agriculture together with their in-depth influences on human and environmental conditions made the agricultural managers to look for new methods of information gathering. Such information must be accurate, effective and accessible on demand to lead the managers for better decisions. Agricultural changes are usually monitored in voluminous spatial-temporal databases.  Spatial-temporal analysis of agricultural fields can confirm the changes, and its type to provide useful statistics. One way to collect such statistics is through spatiotemporal reasoning. The process of identifying differences in the state of a phenomenon by observing it at different times is called change detection. In remote sensing, change detection is done by using satellite images that are taken in different time epochs. The images are compared based on the corresponding pixel values. The results of this process are highly depended on the used methods as well as the interpretation strategies. Several techniques of change detection in remote sensing have been developed. Most of these methods are based on ground sampling. Control points are also used to evaluate the final results. Because of dynamic nature of corps, ground samples are only valid for one cultivation season and this process (ground sampling) should be repeated in every cultivation season. This, itself, increases the cost and the time required for change detection. In most of these methods interpretation of results is a complex approach that needs lots of experiences in remote sensing. Therefore, selecting appropriate method of change detection is important. Evaluation of studies about knowledge-based systems and their applications in Geomatics showed that, proper use of knowledge can increase the accuracy of existing methods. Therefore, in this research a knowledge-based change detection method was developed to detect changes of farms and to identify the type of changes. The proposed method has two main stages: 1) creating a spatial knowledge base and 2) inferencing stage. The knowledge base of this method includes three sets of spatial (geometric), temporal, and spectral information. These rules are achieved by analyzing of rotation history, time series of satellite imageries, farms maps and other facts which can increase the accuracy of change detection. Mugan plain in northwest of Iran was selected as the study area to test the proposed methodology. Rotation history of wheat farms and time series of Landsat 8 imageries were used to execute the test. Different sources affect multitemporal satellite-image datasets such as atmospheric effects, the sensor’s stability and responsiveness. Because of the importance of homogeneity of multitemporal satellite-image datasets, especially in vegetation change detection by remote sensing data, a relative radiometric normalization method was used. To achieve the temporal stability in series of images, this step is taken. In this process the radiometric properties of an image time series is adjusted to match that of a single reference image. Implementation of the method in wheat farm, proved to 86 percent accuracy in change detection and 80 percent of accuracy in type of changes. In this method, the type of changes is recognized through spatial knowledge-based, and no needs were found for using rendition. By removing mixed pixels, the proposed method resulted in an increase in accuracy of change detection and in the identification of the type of changes up to 95% and 90% respectively. Therefore, it is concluded that the results of this method are more accurate than that of Normalized Vegetation Index (NDVI) differencing and Post-Classification. NDVI resulted in 70% accuracy while Post-Classification Comparison has 81% percent accuracy. In addition, the proposed method reduced field work for data collecting.