This task is performed by the Resampler on a “per-pass” base (half an orbit), such that subsequent SHARP1-images of the same orbit are treated simultaneously. It involves 4 steps:

STEP1:

The chip matching results are read from the L2-file(s) MATCHES.TXT, and for each GCP two “delta's” are computed: dC=C_it - C_ie and dR=R_it - R_ie. As before the subscripts t/e indicate the true/estimated values, and the latter are derived by inputting the hot spot longitude and latitude (λ_t, β_t) in the L2 GEO_GRID: (C_ie, R_ie)=CR_i[λβ_t]. The delta's are then submitted to three different filters in order to remove possible outliers (distance from uni-variate mode, bi-variate Euclidean distance from general mean, and uni-variate Gauss filter).

STEP2:

This step involves an update of the original ephemeris and a first actualisation of the L2-grids. It is only executed in the normal case where the Pre-Processor indeed has found the appro­priate TBUS-ephemeris for the concerned image – and if enough GCP are available after the previous filtering (default: at least 3). SpacePC assumes that the errors in the initial grids are main­ly due to a bias in two orbital parameters: the scene centre time t_SC and the longitude of the ascending node Ω. The AVHRR internal clock (and hence t_SC ) indeed can show errors of up to 1 second, which corresponds with an orbital displacement of 6 lines (6.6 km on the ground). Moreover, adjustment of t_SC and Ω indirectly corrects for effects not covered by the model, espe­cial­ly the attitude (the fact that the sensor doesn't look perfectly downward to nadir at σ=0°).

For each GCP, a new set of estimates of t_SC and Ω can be computed, by assuming all the other orbital and sensor parameters are correct, and by traversing all the equations of the geometrical modelling (§ 2.5.3) in the opposite way. Starting from the true GCP-values (λ_t, β_t, C_it, R_it), the procedure briefly goes as follows – the numbers of the concerned equations are given between brackets.

At first, δ_P(inverse of 2-10), σ(2-5) and r_P(2-7) can directly be recomputed. Then follows a group of 5 variables: κ(2-6, r_S still unknown), γ(inverse of 2-8), ν_S= γ - ω(2-3), E_S(inverse of 2-2) and r_S(2-2), which have to be solved by an iterative search until stabilisation and taking (a+b)/2 as initial value of r_S. From E_S, the pixel registration time t can be derived (inverse of 2-2) and then the updated value of the scene centre time t_SC (inverse of 2-3). On the other hand, the new value of Ω can be computed through the following sequence: δ_S (2-4), Δλ (2-9), λ'_S =λ_P ± Δλ - (t-t_A) ω_E (2-9 and 2-10 combined), χ (2-4) and finally Ω (inverse of 2-4).

This procedure is repeated for each GCP and at the end, the obtained values of t_SC and Ω are averaged. These mean values represent the “updated ephemeris”. By means of this new ephemeris, the entire geometrical model (§ 2.5.3) is run again in order to compute new versions of the L2-grids (IMG_GRID and GEO_GRID).

STEP3:

The just created L2-grids are still considered imperfect. The residual errors which are mainly due to the weakness of the model, are now removed by means of a statistical regression on the GCP's. This step does not require the presence of an initial ephemeris, so contrary to the previous grid update, it can always be applied. The only conditions are that enough GCP's survived the filtering step (default: at least 4) and that they are well spread over the entire scene (for SHARP1: set of scenes). This clustering degree is quantified by means of a “City Block” distribution index (see figure 2-14) which should exceed a minimum threshold (default: 3).

Figure 2-15 For the chip matching, the L2-scene is divided in 16 equal-sized “quadrants” with fixed numbers. Quadrant 7 shows a chip window moving around in its search window. Suppose only the 4 shaded quadrants contain one or more GCP’s (chip matches). The computation of the “City Block” distribution index is shown in the table on the right. The values are the numbers of horizontal or vertical steps needed to go from one quadrant to the other. The distribution index is the sum of all these steps (32).

If all conditions are fulfilled, for each GCP the estimated latitude and longitude (λ_e,β_e) are computed, now by inputting the true (C_it, Rit) in the IMG_GRID: (λ_e,β_e)= λβ[C_it,Rit]. The relationship between the true and estimated values is then established by means of the following two polynomial regressions:

The above equation represents a 3^th order polynomial. In practice, the number of included terms (respectively 3, 6 and 10 for polynomial orders 1, 2 and 3) is adapted to the available number of GCP's. Finally, the IMG_GRID λβ[CR_i] – which so far retrieved biased values (λ_e , β_e) – is updated by application of the found polynomials (λ_t=f₁(λ_e,β_e) and β_t=f₂(λ_e,β_e)) on all the values in the grid. As before, the GEO_GRID is also recreated by inversion of the new IMG_GRID.

STEP4:

This last step performs the output of the entire procedure. The updated grids are stored in the L2-directory. The “updated ephemeris” is written to the L2 PAR-file and also to the AUXILIARY/REGISTRATION-directory. The latter also receives the found polynomial coefficients. Finally, the matched L2-pixels (at positions C_it, R_it) are labelled by resetting bit 16 of band 5 from 0 to 1. In this way, the retained GCP's can later be traced back on the L2-images.