Developing New Index to Object-Based Shadow Detection from High Resolution Satellite Images over Urban Area

Height variation in different urban objects e.g. buildings and trees coincide to occurrence of shadows in aerial and satellite images. Areas casted by shadow, appear darker than neighbouring areas in the image and it makes an unwanted contrast to the other brighter areas. This phenomenon attenuate different expectations from remote sensing data. In particular shadow areas ruins the result of automatic image matching algorithm and in land cover classification cause the misclassified pixels.
Detection of overshadow areas is the primary step to deal with this problem. Different strategies have been used to detect shadow in remote sensing images. To name a few we can consider classification based methods, region-growing methods and different spectral indices. In classification based method, some ground truth from shadow areas are collected and supervised machine learning algorithms are used to classify shadow and non-shadow pixels. Region-growing algorithms use the high contrast between shadow and bright areas. Spectral indices are made by simple arithmetic equations between spectral bands.
There is some deficiencies in the result of previous methods and strategies. In machine learning methods, existence of ground truth information is essential and somehow affect the results. Using region growing and spectral indices usually leads to addition of roads and vegetation to shadow areas. The result of all this methods are presented in pixel level, labelled shadow pixels. Wrongly detected shadow pixels appear as noises in classification map.
In high resolution aerial and satellite imagery single pixels are not meaningful independently. This is the outcome of decreasing the ground sampling size of sensors versus natural objects on the earth. The solution to deal with this problem is to integrate similar neighbouring pixels which belong to the same ground object. Object-based image analysis (OBIA) is developed based on this idea, considers image object, created by segmenting the image, as processing unit. The power and possibilities of image objects are less discussed and considered in detecting shadow areas.
In this paper we propose a new object-based framework for shadow detection which simultaneously benefits from OBIA, machine learning and spectral indices. Our proposed framework consists of four main steps. First step is the pre-processing of data. In this step spectral bands are pan-sharpened to enhance the spatial accuracy and the panchromatic band is segmented by eCognition Software. In the second step new spectral indices are proposed to overcome the weakness of existing indices in mixing roads and vegetation to shadow areas. To automate the process of detecting shadows from index values the Otsu thresholding algorithm is employed.
Third step is object-based shadow detection. To detect shadow areas in object level, majority analysis of shadow pixels in each image object is considered. To solve the ambiguity between vegetated and shadow objects an extra condition is checked to confirm that an object belongs to shadow class. This condition uses the mean NDVI value of pixels in each image object. Finally in the fourth step evaluation of produced map is obtained using completeness, correctness and F-measure. In this step the result of shadow detection using spectral indices, proposed index, machine learning and proposed method are compared and analysed.
GeoEye-1 satellite data comprised 4 spectral bands over Qom city in Iran is used in the experiment. 800 shadow objects are selected manually to evaluate the result. Correctness, completeness and F-measure obtained from confusion matrix of shadow map are calculated to compare the results. The result of shadow detection by spectral indices and SVM and random forest classifiers have been compared to the result of proposed method. Result of our experiments demonstrates the superiority of proposed object-based over the pixel-based method respect to correctness and F-measure for different classifiers. The proposed method succeed to detect shadow area with 93% correctness and 92% Completeness It is also evident that object-based method have well behaviour on the edge of shadow areas and perfectly detect shadows.