GEOMETRIC QUALITY ASSESSMENT OF BUNDLE BLOCK ADJUSTED MULTI-SENSOR SATELLITE IMAGERIES

The integration of multi-sensor earth observation data belonging to same area has become one of the most important input for resource mapping and management. Geometric error and fidelity between adjacent scenes affects large-area digital mosaic if the images/ scenes are processed independently. A block triangulation approach “Bundle Block Adjustment (BBA)” system has been developed at ADRIN for combined processing of multi-sensor, multi-resolution satellite imagery to achieve better geometric continuity. In this paper we present the evaluation results of BBA software along with performance assessment and operational use of products thus generated. The application evaluation deals with functional aspects of block-adjustment of satellite imagery consisting of data from multiple sources, i.e. AWiFs, LISS-3, LISS-4 and Cartosat-1 in various combinations as single block. It has provision for automatic generation of GCPs and tie-points using image metafile/ Rational Polynomial Coefficient’s (RPC’s) and ortho/ merged/ mosaicked products generation. The study is carried out with datasets covering different terrain types (ranging from high mountainous area, moderately undulating terrain, coastal plain, agriculture fields, urban area and water-body) across Indian subcontinent with varying block sizes and spatial reference systems. Geometric accuracy assessment is carried out to figure out error propagation at scene based ortho/ merged products as well as block level. The experimental results confirm that pixel tagging, geometric fidelity and feature continuity across adjacent scenes as well as for multiple sensors reduced to a great extent, due to the high redundancy. The results demonstrate that it is one of the most affective geometric corrections for generating large area digital mosaic over High mountainous terrain using high resolution good swath satellite imagery, like Cartosat-1, with minimum human intervention.


INTRODUCTION
The integration of multi-sensor earth observation data pertaining to a particular area is one of the most crucial inputs for resource mapping, management, planning and natural hazard mitigation, which require both multispectral coarse resolution and a number of high resolution data in a timely manner.Hence, geometric processing of heterogeneous sensors need to be handled in a large single block triangulation with minimum human intervention.IRS and Cartosat series provides a stack of multi-resolution multispectral and high resolution pan imagery (i.e.AWiFs, LISS-3, LISS-4 and Cartosat-1) but mapping quality gets affected by the heterogeneity of data sources in structure, semantics and geometry of sensors.Cartosat-1 alongtrack stereoscopic imagery (fore +26 and -5 aft) with 2.5 m spatial resolution (B/H ratio -0.62) add an advantage to multiscale data synchronisation.Most common practice is inputs to the photogrammetric processing system are Rad-orthokit or Geo-orthokit (mono/stereo) along with Rational Polynomial Coefficients (RPCs) and or Level-1A image along with Ancillary data information file (ADIF).The RPCs for each image is generated using orbit attitude information of single image scene/ image strip.Geometric error and fidelity between adjacent scenes are affected in large-area digital mosaic if the images/ scenes are processed independently.A number of scenes/ strips from different sensors need to be processed as a single block with various types of reference data to achieve geometric consistency and feature continuity for state/ district/ city/ watershed level mapping.Moreover to deal with the high data volume of available satellite systems, a highly automated processing is required for cost effective timely production of large number of data products.A digital system (Bundle block adjustment and production system -BBA) with block triangulation approach is developed at ADRIN, providing full photogrammetric production line from satellite triangulations to ortho mosaics, for combined processing of multi-sensor, multiresolution data products with substantially reduced human intervention in a short time.
This software system is substantiating IMGEOS requirement (BBA user Manual, 2014).In this paper we present the evaluation results of BBA software system before operationalization along with performance assesment and product quality.

BBA APPROACH
The BBA software system supports multi-sensor approach, the ability to process geometrically the imageries collected from different sensors in a single adjustment.The principle of bundle block adjustment is based on the co-linearity equation, a method to calculate the orientation parameters from ground control points and their positions in the image, precisely with RFM model and Rigorous Sensor Model (Nagasubramanian et. al, 2007 andRadhadevi et. al, 2010).Ground control points (GCPs) and Tie points are handled simultaneously in one single adjustment process which assures high geometric precision and fidelity.Automatic generation of control points, tie points with adjacent scene overlap areas are generated using techniques of image correlation, mutual information, edge matching etc and it is refined by blunder detection methods.The advantages of block adjustment system are -Reduce the number of ground control points Obtain a better relative accuracy between the images/ image-strips and precise mosaic over large areas Accurate co-registration/ pixel tagging between multisensor/ bands products Typical satellite imagery blocks, collected across multiple seasons, over several years and under different atmospheric conditions (fog, cloud cover), contributing in 'difference in appearance' resulting in matching failures.Additionally, the image quality of varying resolution further complicates image matching resulting in problems when performing automatic tie point matching on blocks with hundreds of images.

Data Description -
Several images covering various geographic conditions from different sensors have been chosen over the six mentioned test sites (  The data processing in one block is planned to assess Cartosat-1 Geometric data processing robustness for cases which causes common error (i.e.across single row, state level area processing, area over land and ocean, area with thin cloud patch etc. as part of block processing).For multi-sensor processing in a single block, sensors can be combined whose spatial resolutions is up to 4 times.For example-Cartosat and LISS-4 can be processed in a single block but AWiFs and Cartosat can't be as the resolution differences between these two sensors are much more than 4 times.

Methodology -
These test data sets belong to a period of last 5 years at different seasons, with varying sun-angle and azimuth, Cloud cover (0% to 6%) and existing with ADIF information or in Orthokit.The products comparison is performed using the in-house developed specific quality analysis tool and other COTS software.The software generates Ortho, bundled and merged products depending upon input and output defined.(Fig- 3).Users have to specify the ROI and sensors for generating BBA products.A number of Strip-Ortho images have been generated in semiautomatic way, using a large volume of data, to enable operational production of ortho products with low internal distortion.Initial experiments and analysis carried out promise generation of products with good geometric quality and feature matching

RESULTS AND DISCUSSION
The software is able to process a large no of imagery in a single block.Scenes with more than 60% cloud or water fail for auto-GCP identification and 3 seed points per scene need to be provided by the user for further processing.

Multi-sensor Block adjustments
Large area multi-sensor products generated over test area 2 (desert) and test area 5 (high mountain) with multiple scenes.The dotted seam lines are indicating the scene extent.

Overall observations
Following are the salient features of this exercise-Good feature continuity with minimal internal distortion is observed in plain area as well as in high and moderately hilly terrains.
Multi-sensor block triangulation can assure better geometric accuracy and feature continuity across a fairly large number of scenes of high resolution as well as pixel tagging in images from multiple sensors.
Results indicate that products from input with ADIF information or with Ortho-kit are equal in terms of geometry and radiometry.Ortho correction with control points from ETM-pan data can give a planimetric accuracy of ± 30m for entire block without any specific error bias.
Error plots confirm that reference datasets of better resolution and accuracy (like Cartosat-1 or Quickbird ortho instead of ETM or LDCM) are more suitable for control points, which is providing better accuracy (within 10 m).It is also observed that better spatial resolution of input image assured products with better accuracies.Co-registration between different bands of LISS-3 and AWiFs are also taken care and showing better consistency through Block adjustment.AOI ortho products, Merged product (Pan + multispectral) and mosaic of pan or multispectral products can be generated within the software environment.
The tool is able to process Cartosat imagery of entire Maharastra (approx.600 no) in a single block.and generate 50 no ortho products per hour with minimal human intervention.

CONCLUSIONS
The software is suitable to make products from medium resolution to very high resolution satellite data and be able to cater the need of future missions.The spatial consistency and positional accuracy of single/ multi-sensor Ortho-corrected products has been significantly improved through the use of Bundle Block Adjustment.Further investigations should be done regarding the statistical evaluation of the product with surveyed controls and understanding the correlation of error distribution in block level as well as in scene level with absolute accuracy.

Fig - 1 .
Fig -1.Process flow at BBA software system Input data preparation component takes care of image selection based on sensors details/metadata, Reference database preparation/ selection (Fig-1).Matching and block processing component happens without human interventions to the maximum extent.If automatic matching fails, it triggers the user to identify seed points /GCP in the failed images and the process will continue.

Fig- 3 .
Fig-3.Single block processed for Cartosat-1 imagery of a mountainous area with error-plot and ortho-productsThe data assessment process will highlight the following aspects: Geometric quality assessment o Absolute geo-location accuracy of the products, estimated against known/ input reference points.

Fig- 5 .
. Feature continuity and differences in ortho outputs

Fig- 8a .
Fig-8a.Example of geometric accuracy with overlaid vector data (road in green line) over multi-sensor outputs scenes

Table 2 :
Test area details 4.3 Test Dataset details A Number of scenes in each block references used and output datum projection details are mentioned in Table-2.

Table 3 :
Details about reference and input datasets.