1. Introduction

Remote Sensing (RS) of air quality has progressed immensely due to technological advances in other fields (e.g. military research) and the growing need for reliable real-time information about air quality. Hyperspectral sensing in the thermal infrared (IR) range has been successfully used for such applications, and both active (with internal IR source) and passive RS systems have shown great potential. Passive RS in the IR range relies on the ambient radiation as IR “source”, as opposed to active RS in which an artificial IR source is used. While active remote sensing is a well-established method for monitoring atmospheric gases [1–3], the use of passive systems is much less common. The few studies with passive RS have mainly focused on gases, such as ammonia emissions [4], smokestack emissions [5], or methanol and ethanol emissions [6]. Active RS is considered to be easier to use, mainly since the post-processing required for analyzing such signals is much less complex than the one required to analyze signals acquired with a passive system. Despite this, passive RS has advantages, such as the lack of a reflector or source and faster deployment time that make it ideal for military or environmental applications. One of the main challenges with passive RS is the inherently low Signal to Noise Ratio (SNR) [7]. Monitoring in passive mode relies, among others, on the temperature difference between the cloud, the background and the atmosphere. Cloud detection, let alone quantification, usually requires a detailed and complex radiative transfer model.

Remote sensing of particulate matter using passive IR instruments, which is the focus of the present study, is much more difficult due to radiation scattering which is virtually non-existent in gases. Although a few studies, such as the detection of biological aerosols [8,9], Tri (2-bis)-ethyl hexyl phosphate [10], and dust particles [11], have shown great potential in sensing particulate matter, detection and quantification of such aerosols remain major challenges.

Conceptually, if a pollutant cloud is present in the line of sight (LOS) of a passive IR RS device, the signal measured by the detector is determined by four factors [12]: (1) the atmosphere between the RS sensor and the pollutant cloud, (2) the pollutant's cloud, (3) the atmosphere between the cloud and the background, and (4) the background. More detailed models that take into account phenomena such as multiple scattering [11] or two dimensional radiative transfer [13] have also been developed. These models require much additional information (e.g. the re-scattering amount of the emitted radiation from the cloud to the LOS and the upwelling and downwelling scattering toward the LOS) in order to isolate the cloud's signal for detection.

One of the factors which determine whether an aerosol cloud can be detected by passive RS is the type of particles which form the cloud, and in particular their optical properties. In principle, a unique refractive index leads to a unique IR signature and the stronger the signature and the spectral features, the easier the detection. Previous IR RS studies focused on particles which are not water based [8–11] and for which most of the spectral absorption features fall in high SNR areas of the spectrum (1600-1800 cm⁻¹ & 3000-3600 cm⁻¹). The present work focuses on detection of clouds of water-based droplets, which is of great importance for environmental studies, such as for instance monitoring pesticide drift.

To promote detection and quantification of a pollutant cloud, a priori knowledge of the cloud-free (background or ambient) signal is needed. The simplest approach is to compare the average radiance sensed with and without the pollutant cloud [9]. A more complex method [12] uses the cloud-free signal to create an orthogonal operator [14] so that the measured signal can be projected into a subspace in which the cloud-free signal contribution is minimized. The challenge in outdoor environment is that the cloud-free signal is constantly changing, sometimes in a matter of seconds, and this change can be greater than the change caused by the cloud.

Our research aims to improve the cloud detection (and in the future also the quantification) process by estimating the time-varying ambient signals using prior measurements and real-time meteorological data: air temperature, relative humidity and solar irradiance. The hypothesis is that reconstructing a cloud-free ambient signal for each measurement with sufficient accuracy will enable fast detection of pollutant clouds without having to resort to complex radiative transfer models (which are very important for quantification). There are two main goals in this research: estimating the cloud-free signals for each measurement using meteorological data, and successfully detecting clouds of water droplets by comparing the measured signal with the reconstructed ambient signal. In order to examine the feasibility of estimating cloud-free signals, spectral measurements were taken with different types of backgrounds (vegetation, concrete wall and sky). For testing the ability to detect water droplets clouds, additional measurements were taken both in presence and absence of clouds generated using a custom-made sprayer system. The cloud-free signals, estimated using the models developed in the first part of the study were compared to the measured signals. Statistical analysis of the difference between the measured and estimated signals was used to determine whether a cloud was present in the LOS.

2. Methods

2.1. Experimental setups

The experiments included two main setups: controlled experiments and field experiments. The spectral measurements in both setups were conducted using RAM2000 G2 OP-FTIR system (Kassay Field Services Inc.). The spectra were acquired at a resolution of 0.5 cm⁻¹, with the detector gain set to 120 (which give an average voltage of ~4V) and the corresponding acquisition time was approximately 3-4 seconds for each individual signal. To reduce noise interference the signals used for analysis from the controlled experiments (section 2.1.1) were constructed from 12 co-added spectra. Although the range of the measured spectra was 500-5000cm⁻¹, the analysis was restricted to the range of 700-1300cm⁻¹ due to low sensor response at lower wavenumbers, and water vapor noise which increases significantly close to the 1500-1800cm⁻¹ range. Beyond 1800cm⁻¹ the radiation intensity decreases dramatically and no significant signal was measured.

During the experiments, measurements of irradiance (Kipp & Zonen, CMP6 pyranometer, range 285 to 2800 nm), and air temperature and relative humidity (Gill Ltd., 1723-B-2-111) were taken every 0.5 second and 30 second averages were recorded. The sensors were located at a height of 2.5 meters, close to the OP-FTIR.

2.1.1. Controlled experiments

Clouds of water droplets were generated using a custom system designed to resemble commonly-used agricultural sprayers. This system consisted of a centrifugal pump (Pedrollo, 2CPm 160/160) and a series of hollow cone nozzles (HCC) (ASJ-spray jet) aligned with the LOS, about 50 cm below the LOS. The nozzle was positioned 20 meters from the OP-FTIR. The spraying system was operated at a constant pressure of 6.5bars. Three types of HCC nozzles were used (Table 1). All the nozzles sprayed the water upwards creating a hollow water cone with an angle of 80°.

Table 1. Nozzles manufacturer specifications.

View Table | View all tables in this article

Spectral signals were acquired with three types of backgrounds:

A fourth set of measurements was performed in a polyethylene tunnel (length: 20m, width: 2m, height: 2.5m), open on one side to the OP-FTIR and on the other side to the vegetation background. The purpose of this tunnel was to minimize wind interference and achieve more stable clouds. A general summary and description of the controlled experiments is given in Table 2.

Table 2. Description of the controlled experiments which included different backgrounds and environmental combinations.

View Table | View all tables in this article

Experiments 1-4 were conducted for 6 hours each and the OP-FTIR spectra were collected continuously. Experiments 5-8 lasted only ~20 minutes each and during them three clouds were generated with the (a) HCC005, (b) HCC025 and (c) HCC050 nozzle. Each cloud was generated by spraying water for four to five minutes with the specific nozzle. The corresponding water load in the tunnel ranged from 0.05 to 0.5 g_water in the LOS.

2.1.2. Field experiment

The field experiment was conducted to test the performance of the proposed method under field conditions and to determine whether it enables the detection of pesticide drift during sprayer application. Agricultural pesticide solutions are composed mainly of water (active pesticide concentration usually ≤ 0.5 g/L), and hence the resulting spray cloud has water-like optical properties. The field experiment was performed in an orchard of mature peach and apple trees (planting pattern: 2.0m by 4.5m, row length: 150m long) at the Matityahu research station at northern Israel on Sept. 29, 2014. Neat water droplets were sprayed using a standard agricultural blower sprayer (Raz Sprayers Inc., model Spidet) with 16 nozzles (HCC025) working at a pressure of 6.5 bars and covering approximately three fourth of the circumference of the blower, pulled by a tractor. Tractor was driven within two tree rows at a velocity of ~5 km/h. The OP-FTIR was stationed 10 meters from the edge of the orchard with the instrument LOS parallel to the tree line. Due to the natural slope of the orchard, the end of the LOS (background) consisted of soil at a distance of roughly 200 meters from the OP-FTIR. Cloud-free reference spectra and meteorological measurements were taken for an hour prior to the spraying experiment, which lasted 20 minutes.

2.2. Modeling of cloud-free ambient signals

Cloud-free ambient signals were reconstructed from meteorological measurements using two methods: Multiple Linear Regression (MLR) and artificial Neural Network (NN) [15]. In both cases the objective was to create a mapping between one or several meteorological measurements and the perceived radiation at a specific wavelength. Both types of models were calibrated for each of experiments 1-4 using half of the data while the remaining signals were used to estimate the model accuracy. Each model (MLR or NN) used meteorological measurements (temperature and/or relative humidity and/or irradiance) as inputs (independent variables) and the radiative intensity as output (dependent variable). Each modeling method (MLR or NN) was carried out 50 times for each wavelength, each time with different sets of data points for calibration and validation. The prediction error was computed as the Mean Absolute Error (MAE) calculated for each of the models at a specific wavelength:

Where MAE_λ is the mean absolute error in the wavelength λ, n is the number of measurements, pred_λ is the predicted value of the intensity at the wavelength λ, and meas_λ is the measured value of intensity at the wavelength λ. Additionally, the standard deviation (STD) of the prediction errors was calculated at each wavelength:

The NN used in this study was a standard two-layer feed-forward network with a sigmoid transfer function in the hidden layer, a linear transfer function in the output layer, and two hidden nodes. The NN inputs were the meteorological data, normalized between 0 and 1. The training process was performed using the Levenberg-Marquardt optimization algorithm. The stopping criterions were: number of iterations (100), performance error (1·10⁻⁵), minimal gradient (1·10⁻¹⁵), and Marquardt adjustment parameter (1·10¹⁰).

2.3. Cloud detection

Cloud detection was based on the comparison between the measured signal and the reconstructed cloud-free signal, using a methodology inspired from Statistical Process Control (SPC) [16]. In standard SPC the samples aspire to a certain proper average and so-called control limits are based on one, two or three standard deviation values from the computed average. The control limits and rules of the statistical process are defined in order to point out anomalies in the process. SPC is usually used on production floors where constant supervision of products occurs. In the present study the SPC concept was used to detect a cloud by identifying the corresponding measurements as “anomalies”. In the absence of a cloud the measured signal was expected to resemble the estimated ambient, so that “control limits” were determined based on the average errors and standard deviations of the ambient models. Since the ambient signal was modeled separately at each wavelength, the initial step of the SPC-like analysis was conducted at each wavelength. In the second step, cloud detection was achieved by combining the results obtained at the different wavelengths included in the analysis (Fig. 1): At each wavelength three statistical rules were used in order to “flag” a suspicious point as an indicator of a potential cloud:

 

Fig. 1 Diagram of the detection process using the methodology inspired from statistical process control. NN-Neural network. IR-Infrared.

Download Full Size | PDF

As shown in Fig. 1, detection was declared when at least 75% of the wavelengths were flagged according to the above rules.

The performance of the proposed approach was compared with the orthogonal subspace projection (OSP) method suggested by [8, 12]. To ensure a fair comparison the spectral range was the same as above (700-1300cm⁻¹). Principal component analysis (PCA) was applied to the signals of each background/environment experiment (experiments 1-4). The OSP operator $P_U^{\perp }$ was calculated using the first three principal components:

Where I is a unit matrix of appropriate dimension and U denotes a matrix with the first three principal components of the spectra. This operator was used to project new data recorded in the same environment:

where $\widehat{\alpha }\left(\lambda ,n\right)$ is the signal after the orthogonal projection, and $PCA\left(M\left(\lambda ,n\right)\right)$ denotes the matrix formed by the principal components of the examined signal $M\left(\lambda ,n\right)$. Ideally, the contribution of the ambient to the projected signal should be close to zero, making it possible to detect the presence of a cloud.

3. Results and discussion

3.1. Modeling of cloud-free ambient signals

Table 3 shows the validation values of average MAE and its average standard deviation, both normalized relative to the average value of the signal at each wavelength. As expected, different background and environment yield different estimation errors. The signals measured with the concrete background (experiment 1) were reconstructed the most accurately for an outdoor experiment, with average estimation errors and standard deviations of less than 0.5%. Estimation of clear sky signals (experiment 2) was much less accurate, with average errors of ~1.0% or more and standard deviations of ~2.5% or more. Although this estimation error may seem acceptable, one must bear in mind that the signal change caused by a cloud is itself very small so that such estimation error is in fact quite large. The relatively poor results obtained when attempting to estimate the signal that emanates from clear sky is not surprising considering that the meteorological measurements were made only at ground level while the sky-signal measured by the OP-FTIR was affected by the entire atmospheric column.

Table 3. Validation average error (Error) and standard deviation (Std) of estimation of the background signal using multiple linear regression (MLR) and artificial neural networks (NN).

View Table | View all tables in this article

The polyethylene open-ended tunnel caused the atmosphere in the LOS to be more stable and less subjected to rapid changes in temperature and radiation, resulting in more accurate modelling for this environment (experiment 3 vs. experiment 4). However, although the tunnel has a positive influence on the ability to reconstruct the cloud-free spectra, the polyethylene sheets can be expected to make cloud detection more challenging since in this case the radiation scattered into the sensor's LOS by the cloud will originate from a source (polyethylene sheets) hotter than the ambient environment. Accordingly, the thermal contrast between the cloud and the cloud free signals is expected to be lower (see further discussion below).

The nonlinear NN yielded lower error values than the simpler MLR modeling. This is not surprising since the basic relationship between emission and temperature is nonlinear (Planck's law). Solar irradiance is strongly correlated with air temperature but it usually changes faster than the temperature, especially when considering temperature changes of the background (vegetation and concrete), so it contributes valuable real-time information to the model. The relative humidity, besides its direct connection to temperature, is connected to the atmospheric transmission (t):

Where α(λ) is the mass extinction coefficient and ρ is the mass column density. When the air moisture content increases, the expression $\alpha \left(\lambda \right)\rho$ increases as well, resulting in a lower atmospheric transmission and a corresponding change in measured radiation.

In all the experiments, using more than one meteorological variable contributed to a reduction of the average error. The use of all three meteorological variables usually reduced the average error, and hence the NN models based on these three variables were used in the rest of this study.

3.2. Cloud detection

As expected from the above results for reconstruction of the cloud-free ambient signals, the ability to detect a cloud was affected by the background and environment. The detection efforts in experiments 5, 7, and 8 started with choosing the most appropriate wavenumber range for detection. To find the ideal range for detection according to the rules described above, the deviation of each signal from the reconstructed cloud-free signal was calculated. Figure 2 shows the discrepancy between each measured data point and the corresponding estimation. In this figure the discrepancies are expressed in terms of standard deviations of the reconstruction errors of the corresponding model for each wavelength. If the discrepancy amounts to less than the sum of the reconstruction error and one standard deviation, the measured data point is not marked. When the discrepancy is higher, the point is marked in gray-level scale (1-2, 2-3, and >3 standard deviations). The spectral regions most affected by the presence of a cloud can be recognized in this figure as the regions which include a large number of gray/black pixels during the cloud event. These results led to selecting the following ranges:

 

Fig. 2 Difference between the estimated cloud-free signal and the measured signals, at each wavenumber, during an experiment with three spraying sequences. A-Outdoor concrete (experiment 5); B- Outdoor vegetation (experiment 7); C- Tunnel vegetation (experiment 8). The gray-level of each point indicates if the reconstruction error fell within one (lightest shading), two or three standard deviations of the reconstruction error of the cloud-free ambient model.

Download Full Size | PDF

After choosing these ranges, the detection procedure (Fig. 1) was implemented. The detection results of experiments 5, 7, and 8 are presented in Fig. 3.

 

Fig. 3 Detection results of experiments 5 (Outdoor concrete, Frame A), 7 (Outdoor vegetation. Frame B) and 8 (Tunnel vegetation, Frame C). The detection results are compared with the actual cloud events (Gray square wave). The detection threshold (75% flagged wavelengths) is indicated by the dash-dot line. The type of nozzle for each spraying event is indicated on the graph

Download Full Size | PDF

The results of experiment 5 (concrete background) are presented in Fig. 3(a). The detection process yielded three positive cloud detections for the three spraying events. As mentioned above, the spraying order in each experiment was: (a) HCC005. (b) HCC025 (c) HCC050. These successful detections were a direct outcome of the very accurate reconstruction of the ambient signal (experiment 1) which ensured that when a cloud was present in the LOS a large amount of data points deviated from the estimated ambient by one or more standard deviations.

As expected, the first cloud, which was sprayed by the nozzle with the smallest flow rate and the finest particle size distribution (HCC005), was the hardest to detect and the detection result was borderline. The water droplets concentration in the LOS is positively correlated with the flow rate and the particle size. Additionally, finer particles achieve limited elevation because of higher drag force and therefore may not reach the LOS height. The detection results of the 2nd and the 3rd clouds, which were created by nozzles with higher flow rate and larger droplets (HCC025 & HCC050), were highly positive.

The results of experiment 7 (vegetation background) are presented in Fig. 3(b). Compared to Fig. 3(a) (experiment 5), all three clouds led to more similar detection signals and all three clouds were detected successfully.

One of the problems in such experiments is the lack of knowledge about the actual droplet concentration in the cloud. The outlet characteristics (outlet velocity, outlet size distribution ext.) are known but wind (convection), temperature, irradiance and relative humidity (evaporation) influence dramatically on the droplet concentration in the OP-FTIR LOS. In the case of experiment 7 (outdoor – vegetation background), the wind also influenced the background itself by moving leafs and branches and changing the vegetation thickness. Nonetheless all three spraying events were detected.

The results of experiment 8 (tunnel and vegetation background) are presented in Fig. 3(c), which demonstrate the difficulty when performing passive RS in the IR range in such an experimental setup (polyethylene tunnel). Although the tunnel resulted in an increased concentrations of the water droplets in the LOS of the OP-FTIR compared to the previous experiment (due to reduced interference from wind and reduced evaporation process in the highly humid tunnel), the detection results were worse than in the equivalent outdoor experiment. This poorer performance is most likely due to the low temperature difference between the polyethylene sheets and the vegetation background. In an open natural environment the ‘hotter’ background radiation is scattered away from the sensor and the ‘cooler’ atmospheric radiation is scattered into the sensor LOS, causing a partial extinction of the signal due to the temperature difference between the two sources. In the tunnel experiment radiation originating from the polyethylene sheets was scattered by the cloud into the sensor LOS while radiation originating from the vegetation background was scattered away from the sensor LOS. Since both sources were roughly at the same temperature the resulting extinction of the signal was minimal and mostly due to absorption.

The only experiment in which cloud detection was not achieved was experiment 6 (outdoor-sky). The signals recorded with the sky as background (experiments 2 and 6, Fig. 4(a)) differed substantially from the other experiments (Fig. 4(b)). High cloud-free reconstruction errors and standard deviations prevented the clouds from being detected.

 

Fig. 4 Measured signals in the presence and absence of clouds. (a) Measured signals of experiment 6 with the sky as background. (b) Measured signals of experiment 5 with the concrete as background. Each graph shows three clouds signals, one background signal and the reconstruction mean average error. Experiments 7 & 8 had signals similar to experiment 5 and are not shown for clarity.

Download Full Size | PDF

The analysis with the OSP method, which was used for comparison, emphasizes the limitations of that approach. This method relies on stability of the ambient signal over time so that the ambient signal will be effectively subtracted from the cloud signal when projected onto a subspace orthogonal to the reference ambient signal. In their study Ben-David and Ren [12] constructed the orthogonal operator using signals recorded 20 minutes before the cloud appearance. To ensure a fair comparison with the approach developed in the present study, the orthogonal operator was constructed using the ambient signals recorded a few days before the spraying experiment (experiments 3 and 7, respectively). Figure 5(a) shows that as expected the projection effectively eliminated the ambient contribution from the signals used to create the operator. However, Fig. 5(b) shows that this operator did not eliminate effectively the ambient signal during the experiment performed a few days later with the same background. In this case the projection yielded similar results for the signals recorded in the presence and absence of a cloud, so that the cloud could not be detected. The results were similar for experiments 5, 7, and 8.

 

Fig. 5 A-OSP-projected cloud-free ambient signals used for creating the OSP operator (experiment 3; outdoor vegetation). B-OSP-projected signals of the experiment 7 (outdoor vegetation) using the same OSP operator as in A.

Download Full Size | PDF

3.3. Field experiment

As mentioned above the purpose of this experiment was to test the ability of the reconstruction and detection processes under real field conditions. A model for the reconstruction process was developed (according to section 2.2) using the 120 ambient spectra recorded prior to the spraying experiment. The reconstruction training process, which used meteorological measurements of irradiance, air temperature and relative humidity together with NN, yielded an average validation error of 0.31% and an average standard deviation of 0.25%, which were smaller than those obtained in the controlled experiments. The wavenumber range chosen was 750-1050 cm⁻¹. The relatively short time of ambient measurements contributed to lower errors and standard deviations (lower variability in the meteorological conditions) but the validity of the resulting model is limited accordingly. Increasing the number of ambient spectra acquired in different meteorological conditions would generate a more robust model which could successfully reconstruct ambient spectra for a wider range of meteorological conditions. Nevertheless, for our short agricultural spraying experiment the measured ambient spectra were sufficient to yield good cloud detection results as can be seen in Fig. 6.

 

Fig. 6 Detection results of the field experiment. The detection results are compared with the actual cloud events (Gray square wave). During measurements 3-20 the sprayer was at a distance of 12m from the OP-FTIR LOS (two rows of tree between the sprayer and the OP-FTIR LOS) and during measurements 23-48 the sprayer was at a distance of 6m from the OP-FTIR LOS (one row of trees between the sprayer and the OP-FTIR LOS). The detection threshold (75% flagged wavelengths) is indicated by the dash-dot line.

Download Full Size | PDF

Figure 6 corresponds to the measurements recorded while the sprayer was traveling within the two rows nearest to the OP-FTIR. Detection was achieved during both spraying events, although, as expected, the detection signal was strongest when the sprayer was traveling in the row closest to the OP-FTIR. In accordance to standard practice, during the transition between the rows the driver briefly stopped the operation of the sprayer and this interruption is clearly visible in Fig. 6 (around measurements 21-25). During the spraying of the first row (further from the LOS) the cloud was not visible to the naked eye.

4. Conclusions

When using passive OP-FTIR measurements, knowledge of the ambient cloud-free signal is essential for the detection of pollutant clouds. The main challenge, especially in prolonged experiments, is that the outdoor ambient environment and background radiation are constantly changing. The method developed in this study is simple and straightforward to implement, and can assist in the field of passive RS using devices such as OP-FTIR. The proposed method can improve the ability to detect pollutant clouds, even when the pollutant spectral features are limited.

Still, there are obvious limitations to the method described in this work. At present, the sensitivity of this method is suitable for applications involving high particle load (such as detecting significant drifts during agricultural spraying), but is not suitable for monitoring low concentrations such as ambient aerosols concentrations.

The performance of the method was investigated both in semi-controlled and real-life measurement campaigns. The first type of measurements showed the feasibility of the approach and emphasized the effect of the background at the end of the LOS on the cloud detection capability. The field experiment demonstrated the method's potential for practical applications as the proposed approach was able to detect pesticide spray drifting out of an orchard in real-time during the spraying operation. Such measurements are of great practical importance. Pesticide drift, both from orchards and open fields, is an important source of air pollution and an increasing source of concern, mainly for communities close to agricultural areas. Despite extensive research, measurements of airborne pesticides are limited and estimations of pesticide drift are still lacking. Information from real-time continuous measurements such as obtained by the OP-FTIR, would be very useful law-enforcing agencies as well as for scientists working on atmospheric transport of pesticides and human exposure to them. In addition, such pesticide drift measurements, which can be performed at various heights, can be used for comparing sprayers’ configurations in order to improve application techniques (i.e., achieve better canopy coverage with lower pesticide drift).

The next step will be to investigate the use of the method described in this work as a pre-processing step for quantifying the particulate pollutants in the cloud. The present method can isolate the cloud spectral signature which, together with appropriate radiative transfer model, could enable quantification of the concentration and size distribution of the aerosols.