EFFICIENT
PROCESSING

To explain what we mean by light touch, we need to start with the technology used in what we refer to as a multistep approach or the mainstream approach. To make it simple we focus on the typical steps for optical data. In this approach, sensor data is processed in many steps. Each step is specialized for a specific problem. In Figure 1 the steps are visualized as processing steps with storage buckets. The steps are named raw; system corrected; ortho; mosaic; and delivery. This is just an example of how it can be done. It is usually even more complex.

Fig 1: The multistep or mainstream approach to processing optical data

It usually starts with the decompression and decrypting stage where the data is made available for further processing. In a step following this, the pixels are made useful by compensating anomalies in the sensor, for example equalizing pixel values based on calibration. Here we often have band registration using complex techniques to make sure bands are aligned. The result is an image where the pixels can be viewed in a meaningful way. You can see features and recognize surroundings without problems with coloration and stripes. This is usually referred to as a system corrected image, an image where all system effects have been compensated for. This is a heavy step as it requires computation resampling of all bands to be aligned.

Before you can use the image in a geographic context, geo-registration must take place. The process takes all kinds of metadata from the sensor and craft, for example, GPS position and attitude angles. A physical model is defined based on the parameters to describe the x/y coordinates on the ground for each pixel. The accuracy of the model and hence the position for each pixel on the ground is dependent on the accuracy of the metadata. In real life, this is where we start to have problems. The metadata parameters are often not accurate enough leading to problems when mapping the pixels to the right place on the ground on a map. The solution is to measure ground truth and recalculate the metadata parameters, that is recalculating the GPS position of the craft and the attitude parameters (pointing angles). This process is today in most cases automatic using photogrammetric analytical models and the process is referred to as geo-registration including adjustment.

To be able to put the image on a map, it needs to be orthorectified. An orthorectified image has all pixels of equal size on the ground and the north is upwards. The orthorectified image can be put under a map of equal size and scale and will fit. This is also a heavy step as it requires computation resampling of all pixels. The output of an orthorectification can be used in a GIS system or other types of systems specialized for analysis.

The next step is where image data gets useful in a GIS context. Images are put together forming large mosaics. A mosaic can cover large areas such as Europe. When making mosaics it is important to handle the remaining radiometric effects making parts of images light and others dark. When combining images, the image edges can be visible due to these radiometric effects. To conclude on the mainstream approach, we can see that each steps gives new pixels in new images to store. Each time pixels are processed a rather costly process is run introducing delays.

The mainstream approach has serious effects on the ability to provide near-real-time data. Each step introduces a delay as data needs to be read and written on a disc. Another problem is that the process is usually dependent on larger segments of data to be possible to process. Despite the need for processing power and disc performance, a typical multistep approach can generate data within 10-15 minutes. This is good enough for most cases but the problem comes when further processing is needed. For example, a geometric enhancement process cannot start until the full scene or image is completed, which adds to the total time.

A more efficient way from a storage and resource perspective is to combine all processing steps into a combined physical/mathematical model. The model includes parameters for each step and makes it possible to map the raw or system corrected image to ortho or mosaic using a light touch in a single step. The model can be applied to the complete image dataset or a selected part. Figure 2 shows the single-step approach used for on-the-fly processing or on-demand processing.

The single-step approach is faster as it avoids repeated reading and writing of image data. It is also faster from a computational perspective as pixel computations are done in a single step. The model parameters can also be versioned making it possible to achieve versions of the image data without the need to store yet another product. The versions are stored as parameters and can be handled as ordinary database data, for example in a database.