Shadow Detection and Removal in High Spatial Resolution Imagery

Abstract

<p>Remotely sensed imagery of the Earth’s surface is acquired as multi-scale image data over coarse and fine scales due to variable sensors mounted on spaceborne, airborne or Unmanned Aerial Vehicles (UAV). Measuring and monitoring of the Earth surface requires repeatable and accurate measurements, so the quality and consistency of remotely sensed data must be maximised. For sensors that acquire images in the electrooptical wavelength range (0.4–0.7 μm), the effects of shadow and illumination require compensation to maximise the accuracy and quality of data. These effects contaminate image scenes and are a result of sun-object-sensor geometry, surface morphology and the Earth’s atmosphere. The research in this thesis is method-focussed and consists of four studies to examine shadow and illumination effects offering techniques and directions to compensate for these effects in images of high spatial resolution.</p> <p>The first study applied an <i>n</i>-dimensional colour space approach to separate shadowed and illuminated pixels from naturally dark or bright objects. An adjustment factor was derived that delineates shadow from directly illuminated areas and separates naturally dark objects from shadow. The factor was applied to the image and accuracy assessment showed a 2.7% classification improvement on the shadow-compensated image. The method required no<i>a priori</i> information and the minimal umbra recovery highlighted the requirement to quantify diffuse skylight in compensation techniques.</p> <p>The second study presented an alternative technique for shadow detection and abundance for high spatial resolution imagery acquired under clear sky conditions from airborne or spaceborne sensors. The method quantified the proportion of diffuse skylight in each image pixel, termed Scattering Index (SI), thus providing a per pixel measure of shadow extent and abundance. Comparative evaluation was performed against two other methods on high-resolution Worldview-3 (1.2 m) and ADS40 (50 cm) images captured over a common scene. Evaluation showed the method improved the accuracy of classifying shadow pixels and, unlike the other methods, it was invariant to scene and sensor characteristics. The method negated the need for complex sun-object-sensor corrections, was simple to apply, and was invariant to the exponential increase in scene complexity associated with higher-resolution imagery.</p> <p>The third study was a field-based examination of shadow behaviour at different spectral wavelengths to quantify shadow empirically and accurately. A “FieldSpec® Pro FR” Spectroradiometer and a Canon 450D digital SLR camera were used to measure signatures of cast shadow. The field-based experiment used an occulter to cast shadow onto a Spectralon white plate to produce incrementally adjusted shadow depths. Results showed that shadow depth was darker and more ‘blue’ at the proximal areas and conversely that image brightness values increased towards distal areas. Since image brightness is a result of sun-object-sensor geometry, the conclusion was that a normalised spectral signature is invariant to geometry and can be used to quantify shadow depth.</p> <p>The last study used all previous results to guide a closer examination of the physics principles behind shadow and illumination effects that resulted in a more concise definition of shadow. The characteristics of illumination, reflectance and Bidirectional Distribution Functions (BRDF) were examined and resulted in a recommendation that ‘atsurface’ reflectance be used as a standard radiometric unit for shadow compensation in remotely sensed imagery. A review of current physics-based approaches for compensation techniques helped define an alternative approach for shadow detection and removal in high-resolution imagery. </p> <p>These studies demonstrated that an alternative approach to compensation of shadow and illumination effects in high-resolution imagery is required. Current approaches require supporting data such as Digital Surface Models (DSM) or BRDF references and these are rarely available or adequate for the exponential increase of detail in high resolution imagery. The findings provide an alternative approach that uses independent physics-based references to overcome existing limitations and can be applied to any electro-optical sensor. Additionally, this research provides a direction for research into shadow compensation techniques that can overcome the challenges associated with the exponential detail that is inherent in high-resolution imagery.</p>