Mısır Ve Pamuk Ekili Alanların Çok Zamanlı Uydu Görüntüleri Ve Obje Tabanlı Sınıflandırma Yöntemi İle Tespiti

Abstract

Uzay teknolojilerinin gelişmesi, uzaktan algılama ve uydu sistemlerinin de gelişimini beraberinde getirmiştir. Her geçen gün daha gelişmiş uydular yörüngeye yerleşmekte, bu durum neticesinde farklı mekânsal, zamansal ve spektral çözünürlükteki yeni uydulardan daha hassas, doğru ve güvenilir bilgiler üretilebilmektedir. Gelişen teknoloji ile uzaktan algılama sistemlerinin kullanım alanları artmış, elde edilen görüntüler tarım uygulamaları yaygın olmak üzere çevre, jeoloji, meteoroloji gibi pek çok disipline bilgi kaynağı olmuştur. Tarım uygulamaları göz önünde bulundurulduğunda, geniş alanlardaki tarımsal faaliyetlerin yersel yöntemlerle izlenmesi ve belirlenmesi oldukça zordur. Bu noktada, uydu görüntüleri ve uzaktan algılama teknikleri kullanılarak tarım arazilerinin mekânsal dağılımlarının belirlenmesi, tarla sınırlarının üretilmesi ve arazideki ürün çeşitleri hakkında bilgi elde edilmesi mümkündür. Ürün çeşitlerinin mekansal dağılımını tespit etmek ve ardından ürünlerin büyüme dönemlerinin takibini yapabilmek amacıyla uydu görüntülerinin doğru bir şekilde analizi büyük önem arzetmektedir. Uydu görüntülerinden ürün tiplerinin ve mekansal dağılımlarının belirlenmesi sırasında kullanılan temel analiz yöntemi sınıflandırma işlemidir. Tarımsal uygulamalarda ürünlerin gelişim dönemlerinin hassas bir şekilde belirlenemediği durumlarda çok zamanlı görüntü seti ile yapılan uygulamalar daha iyi sonuçlar vermektedir. Fakat bu noktada veri setleri farklı uydulardan elde edilmiş ise mekansal çözünürlük ilişkilerinin doğru kurulması gereklidir. Zira sınıflandırma yöntemi seçimi ve analiz tasarımı özellikle bu parametre ile doğrudan bağlantılı olacaktır. Bu çalışmada Şanlıurfa iline ait 3 ilçede (Harran, Ceylanpınar ve Viranşehir) farklı tarihlerde algılanmış Landsat8 ve SPOT6 görüntülerinin çeşitli kombinasyonları kullanılarak pamuk ve mısır tarlalarının tespitine yönelik sınıflandırma analizleri gerçekleştirilmiştir. Orta ve kaba çözünürlükteki uydu görüntülerinin sınıflandırılmasında kullanılan piksel tabanlı yöntemlerin yüksek çözünürlüklü görüntülerin sınıflandırılmasında yetersiz kalması, nesne tabanlı sınıflandırma yöntemine yönelimin artmasını sağlamıştır. Bu çalışmada mısır ve pamuk tarlalarının tespiti için nesne tabanlı sınıflandırma yöntemi kullanılmış, aynı zamanda nesne tabanlı sınıflandırma yönteminden yüksek tematik doğruluk elde edebilmek ve tarım alanlarını parsel seviyesinde tanımlayabilmek için ihtiyaç duyulan parametrelerin seçilmesi için birçok deneme yapılmıştır. Ayrıca yüksek mekânsal çözünürlüğe sahip ve yüksek maliyetli SPOT6 görüntülerinin yerine alternatif olarak Landsat8 görüntüleri ile denemeler yapılarak yüksek tematik doğrulukta ve parsel seviyesinde sonuçlar elde edilmesi amaçlanmıştır.

Remote sensing and satellite technologies have developed significantly in the recent years. New Very High Resolution (VHR) satellites such as SPOT-6, SPOT-7, Pleiades 1A, Pleiades 1B etc. were launched to their orbits and these new systems have extensive acquisition capacity of collecting several millions km2 area per day. Extracting precise and accurate spatial information using these data sets by applying different methodologies is among the most common subject that explored by scientists and experts. With the increasing capability and usage of remote sensing systems, satellite images with different qualities have been used widely by several disciplines like environment, geology, meteorology etc, but especially in agriculture. When dealing with large agricultural areas, determination of the agricultural activities by ground based methods can be time and source consuming and sometimes become nearly impossible. On the other hand, spatial distribution, boundaries of agricultural areas and variety of the crops can be determined by using satellite images and remote sensing techniques. The accuracy of image analysis is extremely important to identify different crop types and monitor the growing stages of these crops periodically. The main analysis method for cultivated area determination is image classification. In agricultural analysis with satellite imagery, multi-temporal image analysis are mostly preferred and provides more accurate results if the phonologic development stages are not precisely determined with in situ data. If the multi-temporal data set is acquired from different satellites, spatial resolution relationships should be carefully established in order to acquire better and more accurate results, since the classification algorithm selection and analysis design are directly affected from this relationship. In multi-temporal analysis, dealing with whole image bands for multiple images is time consuming and may reduce the analysis performance. At this point, spectral vegetation indices, with their sensitivity to biophysical properties of crops, are commonly used for identifying crop types and vegetation index calculation of multi-temporal satellite data helps to emphasize time dimension as well. Normalized difference vegetation index (NDVI) is one of the most widely used indices in crop type identification, which focuses on red and near infrared portion of electromagnetic spectrum. In the NDVI images, dense vegetation is expected to have higher NDVI values therefore they appear as bright on the image and the areas with little or no vegetation are expected to have very low NDVI values therefore they appear as dark on the image. In this study, parcel based identification of cotton and corn fields was conducted using different combinations of multi-temporal SPOT6 and Landsat8 images in three selected districts (Harran, Ceylanpinar and Viransehir) within Sanliurfa which is one of the largest provinces of Turkey. Sanliurfa is located in the southeastern part of Turkey and has great agricultural production potential. Harran Plain, which is partially located in Harran district, is one of the most fertile, biggest and irrigated agricultural areas of Turkey having a considerable impact on agricultural economy. Between 1984 and 2011, irrigated arable lands in Harran Plain had been increased % 59.77 because of GAP (Guneydogu Anadolu Projesi - Southeastern Anatolia Project), which is a project that aims to contribute economic development and social stability by increasing productivity and employment possibilities in rural areas of southeastern part of Turkey. TIGEM (Tarım İşletmeleri Genel Müdürlüğü - General Directorate of Agricultural Enterprises), which is located in Ceylanpinar district, contains % 4.5 of total irrigated arable lands of GAP. TIGEM area specifically focuses on producing all kinds of goods and services to meet the needs of agriculture and agro-based industries. For this reason, that area has especially large fields which contains a wide variety of products. Also because of different types of irrigation systems, the area has regular geometric shaped fields. Viransehir is one of the biggest districts of Sanlıurfa, but the agricultural production rate is lower than other districts, so it contains smaller lands and different product types. Accurate identification of crop types using remote sensing data depends on many variables such as extent of the study area, variety of crop types, phenology of these crops and size of agricultural parcels. Suitable combination of spatial, spectral and temporal resolution as well as a proper balance of quality and price is required in order to increase the accuracy and feasibility. There are many different types of sensors with different resolutions that have been used for environmental studies. In this research, high spatial resolution SPOT-6 and medium spatial resolution Landsat-8 satellite imager were used to identify cotton and corn fields. SPOT-6 satellite sensor, which was built by AIRBUS Defence and Space, was launched on September 9, 2012. It is the 6th member of SPOT family and the twin of SPOT7 which was launched in 2014. It has 1.5 m spatial resolution in panchromatic and 6m in multispectral mode with blue, green, red and NIR bands. It provides images with 60x60 km frame size. Landsat8, which is an American Earth observation satellite, was launched on February 11, 2013. It is the 8th member of Landsat family. It has 15 m spatial resolution in panchromatic and 30m in multispectral mode with 175x185 km frame size. Although selection of the sensor is crucial for agricultural mapping with satellite image, selection of the data acquisition dates is very important and it strongly depends on the crop types since different crops have different growth cycles. Data acquisition dates of SPOT-6 and Landsat8 satellites over the region were determined by considering phenological cycles of corn and cotton; in addition possible dates in which both of these crops had high spectral difference was pointed out using field data. Since cotton is generally planted in the beginning of summer and grow slowly, corn is planted mid summer and grows very fast, discrimination between these two plants would be very distinctive with the usage of multitemporal data sets. As a result, SPOT-6 images acquired on 27 June 2013, 23 July 2013 and 13 September 2013 and Landsat8 images acquired on 27 June 2013, 29 July 2013 and 15 September 2013 were used in this study. The specific aim of this study is the identification of the optimal number of acquisition dates and satellite images from two different sensors to receive high classification accuracy in parcel level. In this research, seven different applications were conducted in order to analyze the best combination of dates and different sensors. For the first application, NDVI images for three different data acquisition dates for SPOT6 were produced, then object-based classification approach was applied to multi-temporal NDVI image to accurately, precisely identify corn, and cotton fields in parcel level. For the next three applications, three different single date SPOT 6 images were classified by using object based classification approach for identification of same two products. For the last three applications, NDVI images for three different data acquisition dates for Landsat8 were produced, then object-based classification approach was applied to multi-temporal NDVI images by using single date SPOT6 images in the segmentation stage and multi-temporal Landsat8 NDVI image for classification stage. Accuracy assessments for all seven applications were conducted using confusion matrix and kappa statistics to analyze the accuracy of the proposed approaches. Our results showed that by only using single date imagery it is not possible to distinguish these two different crop types; however, usage of multi-temporal NDVI image assisted to monitor for phenological changes and resulted in very good delineation of corn and cotton parcels. In addition, using Landsat8 images for spectral information and SPOT6 images for parcel level identification gives almost as accurate results as using multi-temporal SPOT6 NDVI image for classification. Moreover, our analyses emphasized that pixel-based classification approach could not produce successful result for very high-resolution satellite images therefore it is out most important to apply object-based classification approach for these data sets.