CHRIS-PROBA PERFORMANCE EVALUATION: SIGNAL-TO-NOISE RATIO, INSTRUMENT EFFICIENCY AND DATA QUALITY FROM ACQUISITIONS OVER SAN ROSSORE (ITALY) TEST SITE

Transcription

1 CHRIS-PROBA PERFORMANCE EVALUATION: SIGNAL-TO-NOISE RATIO, INSTRUMENT EFFICIENCY AND DATA QUALITY FROM ACQUISITIONS OVER SAN ROSSORE (ITALY) TEST SITE Alessandro Barducci (1), Donatella Guzzi (1), Paolo Marcoionni (1), Ivan Pippi (1) (1) National Research Council Applied Physics Institute Nello Carrara via Panciatichi 64, 5017 FIRENZE, ITALY - I.Pippi@ifac.cnr.it ABSTRACT A methodology for evaluating CHRIS performance is presented. Starting from hyperspectral images acquired over San Rossore (Italy) test site, different issues concerning instrument response are investigated. Particular attention is devoted to the assessment of the noise (due to fully stochastic process) affecting the images and to the estimation of signal-to-noise ratio by means of an innovative algorithm developed by the authors. This procedure analyses the bit-planes extracted from any monochromatic image in the hyperspectral data cube, then it assesses the randomness of every bit-plane, and finally computes the signal-to-noise ratio for each spectral channel. Differently from more traditional signalto-noise estimators, which need to select an homogeneous area in the concerned image to isolate noise contribution onl our algorithm is insensitive to scene texture. This circumstance allows the developed method to completely process the image of any observed ground area. The paper discusses other possible sources of systematic disturbance (like stripe-noise, smear effect, and so forth) in order to assess the overall data quality. Finall comparisons between estimated and expected instrument efficiency are presented and discussed. 1. INTRODUCTION Nowadays a great effort is being devoted to the development of hyperspectral imagers with increased spectral, spatial and radiometric resolution for Earth remote sensing. However, these technical advantages may give rise to a lot of problems like as those related to data calibration that have great significance for the remote sensing of Earth. For example, the extended dynamic range of these sensors can be fully exploited only if the instrument response is correctly calibrated in order to avoid the effects of noise patterns and other disturbances that seriously dim the image quality. Moreover, the images acquired by hyperspectral device are generally affected by random noise, and non-periodic disturbances that are characterized by high degree of spatial and spectral coherence like the well known stripe-noise [1][][3]. Proc. of the 3rd ESA CHRIS/Proba Workshop, 1 3 March, ESRIN, Frascati, Ital (ESA SP-593, June 005) The amplitude of this disturbances pattern superimposed to the observed target may be quite large as much as 0-30% of the unperturbed signal [4][5][6]. To reject this kind of systematic disturbance a very often used technique is to perform a measurement of flatfield, i.e. illuminating the instrument entrance slit with a homogeneous and uniform radiance source. However, this measurement becomes a less practical strategy as long as the digitised accuracy and the detector s dynamic range increase. This is due to the difficulty in finding laboratory sources whose output can be controlled with high spatial and spectral accuracy (1 part in 4095 for a 1 bit ADC). For example, satellite sensors can be influenced by aging and some types of detector require continuous re-calibration to accommodate the gain drift among the built-in read-out channels [7][8]. With particular reference to imaging spectrometer operating in push-broom mode like CHRIS sensor, the problems of removing systematic disturbance and estimating the image signal-to-noise ratio (SNR) are investigated in this paper. After introducing some details about CHRIS spectrometer in Section, outcomes from the procedure of rejection of stripe-noise are shown in Section 3. The description of an innovative algorithm aimed to estimate image SNR is given in Section 4 together with first results. In Section 5 conclusion and open problems are drawn with the discussion of future activities to be carried out in order to better assess CHRIS data quality.. IMAGING SPECTROMETER: THE CHRIS SENSOR CHRIS is a push-broom imaging spectrometer that was designed by SIRA Electro-Optics Ltd. (U.K.) to collect data for land investigation and aerosol measurement. CHRIS is the main instrument payload on ESA small satellite platform PROBA-1 launched on October, 001. The mission is being used as a demonstrator in order to evaluate the performance of compact design technology. The knowledge derived from CHRIS-PROBA will guide the design of hyperspectral imaging systems for future missions. The main scientific CHRIS-PROBA goal is the measurement of Earth surface directional reflectance in the visible and near-infrared spectral bands using the platform pointing capability [9].

2 To this aim CHRIS acquires a set of five images of the same scene during the same sun-synchronous polar orbit. Each image set has an associated fly-by position on the ground (roughly the image center) that corresponds to the Minimum Zenith Angle (MZA), defined as the off-nadir inclination of the sensor viewing direction in the plane perpendicular to the satellite orbit. The line of sight inclination in the along track direction (held in the orbit plane) is indicated by the Fly-by Zenith Angle (FZA). The geometrical composition of the aforementioned angles gives the true zenith angle of the sensor viewing direction. CHRIS acquires images of the same scene at FZA of +55 o, +36 o, 0 o, -36 o, and -55 o during the same orbit with different spatial and spectral configurations (MODE). In fact, CHRIS collects 18, 37 or 63 spectral bands if operated at spatial resolution of 18 or 36 m with a spectral coverage ranging from 405 nm until 1050 nm. The main characteristics of CHRIS spectrometer are listed in Table 1. Table 1. Main characteristics of CHRIS-PROBA. Instrument: Field of view: Ground swath: Altitude: Orbit inclination: Descending node: Across track pixel size: Along track pixel size: Number of images: Spectral range: Spectral resolution: Number of spectral bands Sensor type: Digitalization Signal-to-noise ratio push-broom imaging spectrometer 1.3 o 13.5 km Apogee: 688 km, perigee: 556 km 97.8 o 1:10 local time 18m or 36m Finest resolution is 18m 5 acquisitions of the same area at +55, +36, 0, -36, -55 view angles during the same orbit 410 nm to 1050 nm From nm to nm and binning possibility From 18 bands at a spatial resolution of 18 m, to 63 at 36m EV CCD5-0 (115 x 780 pixels, 5 µm x 5 µm pixel size, full frame transfer, thinned and back-illuminated) 1 bits max. target albedo 0., λ 800 nm, gain Since June 00 up to now many hyperspectral images were collected over San Rossore forestry test site (latitude: o N, longitude: o E, altitude 5 meters a.s.l.) by CHRIS utilizing different MODEs and for various angular geometries. Table lists the complete set of images acquired so far. Table. Complete list of CHRIS images so far acquired over the San Rossore forestry test site. DATE and TIME (GMT) MZA MODE 1 June 19, 00 10:7-6 o 4 April 16, :41-19 o 4 3 June 16, :3 0 o 4 4 July 1, : o 4 5 July 5, :34 - o 3 6 August 5, :3 + 4 o 30 7 September 18, :37-8 o 3 8 October 1, :4-16 o 3 9 October 19, : o 3 10 January 15, :34-6 o 3 11 February 9, :43-15 o 3 1 March 7, : o 3 13 May 7, : o May 8, : September 7, : September 8, : September 16, : October 7, : January, : March 10, : Let us note that the CCD detector operates in a frame transfer mode, hence the sensor may be affected by an additional disturbance usually called smearing effect. 3. RANDOM NOISE AND SYSTEMATIC DISTURBANCE 3.1 Physical background Remotely sensed images are generally affected by different kinds of noise that can be distinguished in two main classes: pattern and temporal noise. To the first class belong periodic and coherent noise while random and isolated noise give rise to temporal disturbances [10]. Periodic noise shows a periodic spatial pattern whose phase and amplitude may or may not be constant. Coherent noise has a partially deterministic nature and, due to the circumstance that it can generate different patterns for each sensor spectral channel, this noise is also spatially and spectrally coherent. The random noise is generated by a fully stochastic process and it represents the last limitation to the image SNR. On the contrary isolated noise is characterised by large amplitude perturbation of the signal (spikes) that is measured in a limited number of samples. The temporal noise and in particular the random noise should be carefully considered for theoretical interpretation of data. In the case of an imaging spectrometer the main sources which contribute to random disturbances are well known and can be distinguished as in the following: - Photonic noise: it results from the inherent statistical variation in the photon flux impinging on a sensitive element of the CCD. Photoelectrons generated within the semiconductor device constitute the signal whose amplitude will randomly fluctuate at each CCD pixel.

3 The photon number received in a given time interval obeys to Poisson s statistics, and, therefore, the corresponding noise is proportional to the square root of the average signal amplitude. The term shot noise is applied to any noise component reflecting a similar statistical variation, or uncertaint in measuring the number of photons collected during a given time interval. - Dark noise: it arises from statistical variation in the number of electrons thermally generated inside the silicon structure of the photosensitive area. For frame transfer devices there is an additional contribution originated by the memory area of the detector. Other random noise components are due to noise in readout electronics, which often holds a d.c. offset produced by bias. This phenomenon is independent of photoninduced signal, but it highly depends on device temperature. The flux of thermal electrons at a given time referred to as dark current. Cooling the CCD significantly reduces the dark current. Another phenomenon that produces a similar effect is electronhole pair generation due to high-energy radiation impinging on the detector area. Both corpuscular (mainly b-rays) and E.M. radiation may contribute to this phenomenon. - Read noise: it is inherent to the process of converting CCD charge carriers into a voltage signal, and the subsequent process devoted to analogue-to-digital conversion. The major contribution to read noise usually originates from the on-chip amplifier chain, which originates frequency dependent noise. Inside this family we can distinguish reset noise and flicker noise. - Reset noise (ktc noise): Prior to the measurement of each pixel charge packet the CCD capacitor is reset to a reference level. Noise is generated at the sense node by an uncertainty in the reference voltage level due to thermal variations in the channel resistance of the reset transistor; - Flicker noise: it is referred to as 1/f noise because of its approximate inverse dependence on frequency; - Charge transfer errors: it is due to imperfection in the process of transferring the charge packets between adjacent detector locations. In particular it results from the trapping of charge within a pixel either during integration time or as a charge packet passes through the pixel during the transfer time. This phenomenon gives rise to a random error whose amplitude may depend on the signal amplitude, array size, and read-out electronics. - Round-off error: it is due to the finite ADC quantization accuracy. Moreover, due to the operation mode of a frame transfer CCD another important source of systematic disturbance has to be considered: the smearing effect. This phenomenon is caused by light falling on the CCD during the charge transfer phase. As long as the read-out phase duration is not negligible the charge packets significantly grow, originating a signal-dependent charge error. All these terms are summarized in the following Table 3. Table 3. Mathematical expression of possible sources of random noise < σ > (root mean square) in a remotely Shot noise: Dark noise: Read noise: sense d image. < σ > < σ < σ > photon > dark read N e ~ ς ( T ) τ 0.1 N τ int int well N Round-off error: < σ > ( bit round off N well 1) Smearing effect: where: λ f δs ( λi ) k( τ shift, τint ) s( dλ λi N well : Full well capacity Number of electrons generated after N e absorption of photons τ int Integration time ς (T ) Function depending on sensor temperature T N bit Accuracy of digitalisation τ Time for transferring carrier shift from a pixel-to-pixel s ( Spectral response of sensor k ( τ shift, τ int ) Function depending on transfer time and integration time λ i, λ f Wavelength extension of a spectral channel 3. Coherent noise in push-broom imaging spectrometers The typical layout of an imaging device operating in the push-broom mode includes a CCD matrix array to resolve the spectral content of each target line imaged by the spectrometer. Thus a certain CCD line of elements corresponds to a fixed wavelength of the spectrometer s free spectral range and each monochromatic image is sequentially acquired by the same row of CCD element located at the corresponding wavelength. Due to this circumstance the non-uniform distribution of sensitivity among the elements of the matrix detector superimposes a systematic pattern of noise organized by vertical lines upon the target image. We refer to this

4 disturbance as a spatial and spectrally coherent noise pattern, which would affect each pixel spectrum as well as the intensity distribution (texture) of each monochromatic image. From a theoretical point of view the signal g output by the detector is expressed as in Eq.1: [ i( * H ( ] S( η( ) g ( η( )) g ( + 0 λ (1) where x and y are the coordinates of a generic image pixel, S is the detector sensitivit g 0 the detector dark current, i the impinging radiance and H the Instrument Chromatic Response (ICR). In this way the convolution i*h represents the spectral radiance which illuminates the detector area placed in the focal plane of the spectrometer. Let us note that we have assumed the ICR to be a constant for all points of the spectrometer entrance slit. In addition the offset term g 0 contains contributions from random noise and bias sources. The term S ( η( ) represents the response of the pixel at image column x that is dragged by the platform along the y (or t) dimension of the image to acquire the whole image at the wavelength (spectral channel) λ. Due to this circumstance S ( η( ) gives rise to a systematic pattern of disturbance that is constant along the image columns and that could change varying the selected spectral band. Let us note that the S ( η( ) fluctuations may be so large (~ ±30%) to require accurate sensor calibration. 3.3 Rejection of coherent noise: results from CHRIS images In the last year [1] we have developed an original algorithm to remove coherent noise patterns from remotely sensed images without the necessity of any flatfield calibration measurement. The basic idea is to separate contributions due to scene texture from those originated by the noise pattern. The algorithm utilises a subset of the Hough transform defined by all the integrals computed over image columns. In this way a profile p corresponding to the average image horizontal line is obtained as indicated in Eq.: depend on the selected spectral channel (i.e. wavelength) and x coordinate. We note that image texture contribution to the mean line p is reduced by spatial averaging but it is not necessarily eliminated. Assuming that disturbance contributions rapidly change with position x differently from the smoothed texture contributions, it should exist a cut-off spatial frequency f c which separates the two contributions. To find and extract the higher-spatial-frequency components of profile p, originated from noise contributions alone, a two-stage filtering procedure has been designed. Firstly the profile p is convolved (smoothed) with a Gaussian weighting function w whose width σ is connected to the cut-off frequency f ~ 1 c σ : s ( w( x ξ ) p( ξ, dξ (3) The obtained smoothed profile s contains contributions that are due to the scene s texture, thus, the p-to-s ratio represents the shape of noise pattern alone as indicated in Eq.4: p( c ( (4) s( Finally the corrected image is obtained as the ratio between the original signal g and the correction profile c: i ( (5) c( Let us note that the correcting profile c oscillates around the value 1, hence the newly defined image i has on average the same intensity distribution of the original signal g. We have applied our algorithm to CHRIS images gathered over San Rossore since the radiance images as far distributed were always affected by notable striping. Fig. 1 shows an example of these radiance images of San Rossore corrupted by coherent noise. p ( λ ) dy () Due to the embedded averaging this profile is poorly influenced by any additive random noise, and it may

5 Fig. 1. At-sensor radiance image acquired by CHRIS over San Rossore on September 18, 003 with MZA -8 o and FZA 0 o and displayed in the 1 st band at 44 nm. (a) (a) (b) nm nm nm nm nm 63.0 nm nm nm nm pixel nm nm nm nm nm 63.0 nm nm nm nm pixel (b) Fig. 3. Scatterplots for the correcting (a) and (b) smoothed profiles for the first and last spectral bands. Application of our image-restoration procedure (destriping) to the shown CHRIS image is reported in Fig.4. Fig.. Calculated profiles for the first nine bands of CHRIS hyperspectral image of Fig.1. Correcting (a) and (b) smoothed profiles were vertically shifted for clarity. In Fig. the smoothed and corrected profiles for image of Fig.1 are shown. As expected the smoothed profiles slowly change with wavelength according to the assumption that they mainly contain contributions from texture. In Fig.3 the cross correlation between the correcting and smoothed profiles are detailed between different spectral channels. These scatterplots show that the correcting profiles are mutually uncorrelated while a strong correlation is found for the smoothed profiles. Fig. 4. at-sensor radiance image acquired by CHRIS over San Rossore on September 18, 003 with MZA -8 o and FZA 0 o and displayed in the 1 st band at 44 nm. As can be seen this algorithm originates impressive results, a quality level that is achieved for all spectral channels. Finally the algorithm s ability to maintain unchanged the image brightness and the pixel spectra has been verified within ± 1% by using different sets of CHRIS images. In Fig.5 the same horizontal profiles extracted from image of Fig. and Fig.4 are compared. Let us note how

6 texture variations are maintained while the strong striping on the left side of the original data is suppressed. represented with N bit and let be the extracted k bit-plane: { } y, 0,1 (5) At-sensor radiance (µw m- sr-1 nm-1) pixel L1B BSQ Fig. 5. Horizontal profiles extracted from Sira image (as shown in Fig. ) (BSQ on right vertical axis) and from IFAC corrected image (as shown in Fig.3) (L1B on left vertical axis). 4. IMAGE SIGNAL-TO-NOISE RATIO ESTIMATE 4.1 Bit-plane method Differently from coherent disturbances, random is due to a fully stochastic process. In remotely sensed images this kind of noise appears as a salt-pepper pattern which affects the scene qualit and can t be removed without any spatial resolution losses. The knowledge of random noise amplitude is important to estimate the corresponding SNR. To this aim an estimation method has been with the essential property to be insensitive to scene texture. The algorithm is based on the assumption that noise statistics is spatially stationary and not auto-correlated (almost ergodic noise). The central idea of our algorithm is to extract from every image the corresponding bit-plane sequence starting from the less significant bit, and to establish if the bit-plane obeys a random distribution. To this aim the difference between any pixel of the bit-plane and each of its neighbour is computed, and the distribution of these differences is calculated. This distribution is analysed in order to assess if it has been generated by a random bit-plane or by a textured intensity distribution. The procedure is repeated for the entire bitplane sequence and the SNR is computed. Let us suppose a N rows by N columns image y) to be r c At-sensor radiance (µw m- sr-1 nm-1) Information about the randomness of the bit-plane distribution can be obtained by analysing the distribution of differences i ( between any pixel and each of its neighbour as in the following Eq.6: 1( ( 3( 4( x + 1, y + 1, x + 1, y + 1, x 1, y + 1, We only consider differences computed along four spatial directions being the remaining ones already included in difference computation performed around neighbouring pixels. It easily found that the statistics of the allowed values for the differences i ( are given in the following table: Table 6. Possible values for the differences i ( as computed in Eq.6. g g A0 B1 1 C1 D0 Let p be the probability to find a bit in the upper state 1 (or success event) and q 1 p the probability to find the bit in the lower state 0 (or failure event). Now we compute the probabilities, for the single P i 0 P i 1 difference i ( as indicated in the Eq.7: P i 0 P i 1 P + P A P + P B D C q + (1 q) q(1 q) This expression means that, until the concerned bit-plane is dominated by random noise, the probabilities of the two only possible values for the difference i ( are equal to 1. In other wordings the variable i ( obeys to the binomial distribution with 1 mean. This holds true even if the sensor ADC (Analogical-to-Digital Converter) would introduce a polarization effect in the output signal as demonstrated in the following. Let us suppose that the probability q of the (6) (7)

7 failure event is perturbed by a small amount ε << 1/ so that: 1 q + ε (8) It is easy to verify that: 1 P i 1+ q q + o( ε ) 0 (9) Our algorithm computes the variable δ and compares it with the expected standard deviation. The bit-plane is labelled as noisy if δ is less than the threshold nσ ξ being n an user-defined multiplier that affects the confidence of algorithm predictions. Let us note that accepting as noisy the values δ 3σ δ we assess noisy bit-planes with roughly only 1% probability of missed hits (noisy bit-planes classified as textured ones). Moreover, calculation of false alarm probability is really complex and its prediction requires strong assumptions concerning textured bit-planes difference distribution i (. and in a similar way: 4. SNR estimates 1 P i q(1 q) + o( ε ) 1 (10) which proves the probabilities and P are 1/ P i 0 i 1 in the less of a second order power of ε. When the concerned bit-plane is dominated by scene texture the average value of the process i ( will tend to fall below 1/, hence the difference δ 1 < i ( x, > i between the expected 1 and the actual < ( x, > mean values should i x, i represent a reliable randomness index. In order to use this index for bit-plane randomness assessment it is necessary to account for the expected variability of it when considering an unknown pure-noise bit-plane. To this purpose the local theorem (Moivre-Laplace formulation) can help. Let n 0 be the number of 0 (failure) occurrences in the concerned bit-plane and δ the new random variable defined as: 4qNrNc n0 δ (11) 4NrNc where 4N r N c represent the total number of the computed differences. According to Moivre-Laplace theorem as long as the number of measurements is high and the difference i ( is a fully stochastic field, this new variable δ obeys to a normal distribution with vanishing mean and a standard deviation σ δ given by: q( 1 q) σ δ (1) 4N r N c Fig.7 shows an example of this processing as applied to the image reported in Fig.4. This image is obtained after composing together all the bit-planes recognised as noisy in the original image. As can be seen the extracted bitplanes truly seem to represent a fully random distribution, i.e. its outcome has the expected salt-pepper aspect. Fig. 7. Result of extraction of noisy bit-plane of the CHRIS image acquired in the 4 th spectral channel. The processing reveals an average of 18 bits of data per pixel (information plus noise) in any spectral band. Only 7 out the 18 found bit-planes are recognized as significant (i.e., bit-planes which carry physical information), the remaining ones being dominated by noise (about 11 bit of noise per pixel). Let us note that the effective number of bits per pixel revealed by the algorithm is higher than the accuracy digitalisation of the CHRIS 1-bit sensor because calibrated images have been processed instead of raw data. Therefore radiometric levels higher than 4096 are expected. However, if we extract the bit-planes sequence from each CHRIS spectral channel some troubles arise. An example of this processing for the 4 th spectral band is shown in Fig.8.

8 (bit-plane 01) (bit-plane 0) (bit-plane 11) (bit-plane 1) (bit-plane 03) (bit-plane 04) (bit-plane 13) (bit-plane 14) (bit-plane 05) (bit-plane 06) (bit-plane 15) (bit-plane 16) (bit-plane 07) (bit-plane 08) (bit-plane 17) (bit-plane 18) Fig. 8. Bit-planes sequence for the 4 th spectral channel from the least significant bit up to the most significant bit carrying information. (bit-plane 09) (bit-plane 10) The obtained outcome in the 4 th spectral band has a different looking: As an instance the 10 th bit-plane of the analysed sequence obeys a fully random distribution over land and inland waters but is affected by image texture over sea. In other wordings while over land the concerned bit-plane has the expected salt-pepper aspect, over sea some scene textures already emerge from the image.

9 The algorithm reveals a noise distribution among bitplanes, which is systematically shifted with respect to that one measured over land. The problem does not appear related to the particular radiometry or spectral signature of water, since all inland waters have exactly the same noise distribution as soils with different cover types. It is worth noting that the sensor itself should not originate any signal-statistics variation between different image regions, a difference that would have been added in the subsequent processing step. In order to obtain a reliable estimate of the SNR we have selected an image stripe, as depicted in Fig.9, that excludes sea areas SNR P eak_snr wavelength (nm) (a) Fig. 9. (a) CHRIS image acquired in the 4 th spectral channel over San Rossore containing only soil and vegetation areas. (b) Result of extraction of noisy bitplane of the same CHRIS image. (b) Fig. 11. CHRIS SNR plot versus wavelength computed from image of Fig.9. In Fig.1 the sequence of bit-planes extracted from Fig.9(a) is plotted. As can be seen the part of the image signal dominated by texture is cleaned by the meaningless bit-planes which do not add information to the data. Let us note that the cleaned image does not have a higher SNR than the original one, since the first encountered bit-plane which is not recognized as noisy still contains important noise contribution. The obtained SNR is reported in the Table 10 and its spectral profile is plotted in Fig.11. Table 10. CHRIS SNR computed from image shown in Fig.9 at each spectral channel. (bit-plane01) (bit-plane0) (bit-lane03) Band Peak_SNR SNR

10 (bit-plane04) (bit-plane05) (bit-plane06) (bit-plane16) (bit-plane17) (bit-plane18) Fig. 1. Sequence of bit-planes extracted from the image shown in Fig.9(a). (bit-plane07) (bit-plane08) (bit-plane09) (bit-plane10) (bit-plane11) (bit-plane1) As can be seen the first 11 bit-planes of the 18 ones per pixel revealed by the algorithm are recognised as noisy while the remaining ones (about 7 bit per pixel) carry information containing scene texture. As expected the noise distribution is confined to the less significant bits while the information occupy the upper states of the word length. 5. CONCLUSIONS In this paper hyperspectral images acquired by CHRIS over San Rossore (Italy) site have been processed in order to assess information and data quality. Particular attention has been devoted to estimate noise and SNR by utilising an innovative algorithm developed by the authors. From this analysis, some troubles have arisen concerning the noise distribution among bit-planes as extracted from any monochromatic image. The algorithm revealed a noise distribution among bit-planes, which was systematically shifted with respect to that one measured over land. Finall an example of SNR retrieved by the algorithm has been presented and its behaviour with changing wavelength has been investigated. 6. AKNOWLEDGMENTS (bit-plane13) (bit-plane14) (bit-plane15) We would like to thank all the scientists from different institutions that are involved in remote sensing activities carried out in San Rossore test site. Moreover, we would like to thank people from SIRA and RSAC in particular M. Cutter and P. Fletcher for planning the acquisitions and providing CHRIS data, and for useful discussions. 7. REFERENCES 1. Barducci, A., and Pippi, I. Analysis and rejection of systematic disturbances in hyperspectral remotely sensed images of the Earth. Applied Optics, Vol. 40, pp , 001.

11 . Watson K., Processing remote sensing images using the -D FFT-Noise reduction and other applications. Geophysics, Vol. 58, n. 6, pp , Barducci, A., and Pippi, I., Environmental monitoring of the Venice lagoon using MIVIS data. Proceedings of the International Geoscience and Remote Sensing Symposium IGARSS 97, T. I. Stein, ed. (Institute of Electrical and Electronics Engineering, Piscatawa J.N.), Vol. II, pp , Barducci, A., and Pippi, I., The airborne VIRS for monitoring of the environment. Sensors, Systems and Next-Generation Satellites, H. Fujisada, ed., Proceedings SPIE, Vol. 31, pp , Aiazzi, B., Alparone, L., Barducci, A., Baronti, S., Pippi, I., Estimating Noise and Information of Multispectral Imagery. Optical Engineering, Vol. 41, n. 3, pp , Aiazzi, B., Alparone, L., Barducci, A., Baronti, S., Pippi, I., Information-Theoretic Assessment of Sampled Hyperspectral Imagers. IEEE Transactions on Geoscience and Remote Sensing, Vol. 39, n.7, pp , Curran, P., J., and Dungan, J., L., Estimation of signalto-noise: a new procedure applied to AVIRIS data. IEEE Transactions on Geoscience and Remote Sensing, Vol. 7, pp , Nelson, M., D., Johnson, J. F., and Lomhein, T., S., General noise processes in hybrid infrared focal planes arrays. Optical Engineering, Vol. 30, pp , Cutter, M., A., Johns, L., S., Lobb, D., R., Williams, T., L., Settle, J., J., Flight Experience of the Compact High Resolution Imaging Spectrometer (CHRIS). Proceeding of SPIE Conference 5159 Imaging Spectrometry IX, August 003, San Diego, California, USA, Nieke, J., Solbrig, M., and Neumann, A., Noise contributions for imaging spectrometers. Applied Optics, Vol. 38, pp , 1999.