Hyperspectral remote sensing refers to a remote spectral detection of light, reflected or scattered from a target. Each pixel of a hyperspectral imager can contain hundreds of spectral channels, as opposed to the traditional three-color RGB cameras. Hyperspectral imaging has been widely applied in applications such as medical imaging and diagnostics [1
], food safety inspection [2
] and agriculture studies [3
]. Hyperspectral cameras are usually dependent on ambient lighting. This limits the accuracy of the spectral signal since any variation in the illumination spectrum translates into a misinterpretation of the target response.
Active hyperspectral sensing refers to a method where the investigated target is artificially illuminated by a broadband light source. A supercontinuum (SC) light source is most commonly used, however optical parametric oscillators (OPO) [4
], as well as frequency combs [5
], have also been used. SC is generated via spectral broadening of a short high-power monochromatic laser pulse in an optically non-linear material, usually an optical fiber [6
]. The resulting light can have an optical bandwidth spanning over multiple octaves [7
] but can still be directed and focused over long distances [8
] as a regular laser. The potential applications for the technology include mineral survey, the automotive industry and agriculture.
The simplest case of an active hyperspectral sensor (AHS) or imager is a passive hyperspectral imager combined with an external light source to provide more light at the measured wavelengths [12
]. However, as with passive illumination, the spectral radiant intensity of the external light source can drift. To extract the accurate spectral response of the target, a calibration target should be used at the target scene. A more attractive way is to calibrate the response with a calibration target only once and actively measure the spectrum of the outgoing beam for reference. Any drifts in the received signal caused by external illumination can be compensated by normalizing the received signal with the reference signal. This enables much more accurate measurements of the target spectral response compared to passive illumination, without the need for frequent calibrations of the spectral response. The active referencing is important when using SC illumination. The SC output is highly dependent on any variation in the fiber geometry and laser pulse properties. This usually results in large pulse-to-pulse variations and a drift of the output optical power and spectrum.
The active monitoring of the outgoing beam can be realized with fast detection where the outgoing and received pulse are recorded with a single element. This approach has high requirements for the detection speed and cannot be used with integrating detectors. Nevertheless, fast detection will allow for combining the ranging measurement with the spectral measurement [13
]. An alternative approach is to use two spectrometers, one monitoring the transmitted spectrum and another for the received spectrum. This approach requires two separate components to disperse the light, which results in a higher cost and complexity of the instrument. Thirdly, if the SC is already spectrally filtered prior to transmission, separate single-color point detectors can be used to detect both transmitted and received light. This approach poses a practical problem of implementing spectral scanning methods, which do not result in wavelength dependent illumination patterns over long distances. This would be the case if prisms or diffraction gratings were used for dispersion.
To date, relatively few prototypes of hyperspectral sensing using SC sources have been published and no commercial instruments exists. More work remains in both instrumental design and application studies in order to find the full commercial potential of the technology. In this work, we present an AHS prototype, based on a voltage-tunable near-infrared MEMS Fabry-Pérot interferometer (FPI) [17
]. The FPI allows for a compact, cost-efficient and fast on-axis wavelength selection between 1300 nm and 1650 nm. We used the FPI to select the illumination band of the in-house built SC source. This approach enables the use of low-cost point detectors for both outgoing and received light. It also allows for more light to be transmitted in eye-safety critical applications compared to illumination of the whole bandwidth simultaneously. To demonstrate the approach, a compact battery-powered hyperspectral sensor was built, characterized and tested in the field.
This article is organized as follows. In Section 2.1
we present the system design of the prototype. Section 2.2
and Section 2.3
present the SC generation and the FPI characterization, respectively. In Section 3.1
we present profiling of the outgoing beam after the FPI and transmission optics. Based on the results, we extrapolate the spectral illumination pattern over long distances. In Section 3.2
we show the results from Allan-Werle deviation analysis of the received pulse energy and demonstrate how long period, calibration-free operation can be achieved with the active referencing. In Section 3.3
we present the results from the field measurements. The prototype is mounted on a test car and we present measurement results provided by the sensor, probing the road condition ahead of the moving vehicle. Finally, in Section 4
we conclude the work.
An AHS sensor based on MEMS FPI was built and demonstrated in both laboratory and in field. The FPI enables simple on-axis spectral filtering of the SC. For eye-safety limited applications, the FPI-based sensor allows for greater transmitted spectral power density due to the spectral scanning of the transmitted light compared to transmitting the whole SC. Furthermore, a simple active reference measurement using a single-point detection was shown to reduce the long- and short-term drifts considerably.
The SC source was characterized for precise directionality. The large observed M2 values of the SC could be reduced by using single-mode fiber for SC generation. However, it seemed that the lower M2 of the longer wavelengths helped to uniform the illumination, correcting the difference in divergence of different wavelengths due to diffraction. Also, the use of multimode fiber has practical advantages on coupling the high-power pump light in to the fiber. The high peak-power pump pulses can damage the fiber end if the light is focused to the cladding instead of the fiber core. The larger core of the multimode fiber allows for more relaxed alignment.
The small low-cost MEMS FPI optical bandpass filter was used for wavelength selection. The FPI filter does not refract different wavelengths and thus allows for light filtering before transmission of the SC. The FPI did not show any signs of degradation or change in tuning properties due to high-intensity laser light during this work. One of the limitations of using the FPI is the reduced bandwidth. The FPI used in this work covered merely 25% of the whole SC spectrum. Future development of the FPI filter might alleviate this limitation. Another practical limitation is the small aperture of the FPI that can govern the transmitters optical design. Finally, with FPI used only at the transmission in the presented prototype, the main detector is exposed to background radiation from the whole 350 nm FPI bandwidth. For this reason, it might be advantageous to synchronize a second FPI to filter ambient light.
A best precision of 0.1% and long-term stability of 0.5% per spectral channel was demonstrated by using the Allan-Werle analysis. This performance, however, is limited to stable atmospheric conditions and received signal-strength. While this is not an issue in indoor applications, correction of the atmospheric attenuation is critical in outdoor measurements due to changes in relative humidity. Fog and dust particles also induce wavelength dependent scattering, affecting the spectral measurement.
With different detector gain levels and signal strengths, the overall 1σ standard deviation increased to 2.5% for all wavelengths, a factor of 5 greater than with the stable signal. It is possible, that small differences in the spectral uniformity affected the spectra measured with the light obstruction. More work is required in order to asses the reasons behind the non-ideal behaviour.
The robustness of the instrument was put to test by mounting on a test vehicle. The sensor acquired spectra of the road 10 m in front of the vehicle. Several road conditions and objects were visually distinguished from the spectra. Future work should include building and integrating a model for automated recognition of the road types and common objects. In the case of autonomous driving or smart cars, the long-distance performance needs to be optimized. The sensor used in this work has been tested for only relatively short distances. With the current prototype, stationary solid targets, with reflectance larger than 0.1, have been measured with good SNR up to 50 m in lab conditions. To increase the distance, avalanche photodiodes could be employed in the design.
With the addition of a scanner, the AHS could be fused as a part of a sensory platform in self driving vehicles, providing target classification for objects, which are not recognized by conventional methods, such as lidar or camera. Future work will also investigate the potential applications for the sensor in other areas, such as mineral survey.