Many weather features such as precipitation and snow depth can be recorded using automatic surface observation systems. However, automatically observing dew and frost presents several problems. Many studies have used various wetness sensors and passive microwave devices to detect dew. Unfortunately, several of these sensors are complex, and only a few are capable of detecting frost. This paper proposes a novel method for indirectly detecting dew and frost based on computer vision. The setup is simple, inexpensive, and only requires images of several glass substrates near the underlying surface. Images taken during dew or frost formation exhibit distinct changes in hierarchical visual features. These changes are detected by tracking the variations of several low-level statistical features that are extracted from the images in time. Additionally, an effective texture analysis method is proposed to describe the morphology of frost. Field experiments were conducted at several weather stations in Beijing, China. The validation of the method for measuring the onset and duration of dew/frost on short grass shows that 1) the proposed computer-vision-based algorithm achieves an accuracy of approximately 90% in discriminating among dewy, frosty, and dry nights based on the hourly manual observations of the grass surface and 2) the algorithm is also capable of measuring the duration of dew and frost on grass with about 70% accuracy.
The presence of dew or frost influences many aspects of the ecological environment, such as crop growth, the water balance of forests, and plant protection (Malek et al. 1999; Beysens et al. 2005; Cittadini et al. 2006). When near-surface atmosphere cools as a result of radiative surface cooling, while the relative humidity of the surrounding air increases, the moisture in the air condenses directly onto cold surfaces when the dewpoint temperature is higher than the surface temperature (Monteith 1957; Shank et al. 2008). The condensed water is called dew or frost when the surface temperature is above or below the freezing point, respectively. Dew is an important source of water for biological soil crusts, plants, insects, and small animals, and also has a beneficial effect on the stabilization of sand in desert areas (Kidron et al. 2002; Agam and Berliner 2006). Frost events in the late fall and early spring can potentially impact agricultural production and cause serious damage to fruit and trees (Tait and Zheng 2003; Prabha and Hoogenboom 2008).
Daily observations of dew and frost are an important component of surface observations, but they must be manually obtained in China. Observers must examine the vegetated ground surface using touch and visual inspection methods at different locations in an observation field. This work should be performed several times every night without fixed time intervals (China Meteorological Administration 2007). Therefore, developing an automated method for observing both dew and frost would be practical.
Several types of drosometers, including container-type drosometers, absorbent paper-type drosometers, and flat drosometers (Barradas and Glez-Medellín 1999; Takenaka et al. 2003) are commonly used to observe the onset and amount of dew. However, these instruments are often used for manual observations, and some of them require maintenance after every use. Recently, attention has been focused on techniques for modeling the relationship between the amount and duration of dew and leaf wetness (Heusinkveld et al. 2008; Cosh et al. 2009; Kabela et al. 2009). Several wetness sensors have been used to determine the onset and duration of dew (Wilson et al. 1999; Moro et al. 2007; Dalla Marta et al. 2007). Some of these sensors, such as the leaf wetness sensor designed by Decagon Devices, Inc. (Decagon Devices 2009), are capable of automatically detecting dew and frost. A practical issue is that any contact surface that is placed near the ground can quickly be contaminated by dirt, and frequent maintenance may be required during long-term monitoring.
Computer vision for surface observations has gained significant attention recently, such as ground-based cloud classification (Calbó and Sabburg 2008; Zhuo et al. 2014). Similarly, observing dew and frost using image-based approaches can have significant advantages. In contrast to the one-dimensional data of wetness that is produced by typical wetness sensors, image data could potentially be used to explore other properties of dew or frost, such as dew patterns. However, obtaining observations of dew and frost is still a challenging task for computer vision. Note that Duvdevani (1947) dew gauges identify the appearance of dew using a set of standard photographs. Although the observations are made by manual comparisons within 1 h after sunrise, these devices indicate the potential of performing dew and frost observations with a computational visual approach. Our previous work (Zhu et al. 2011) analyzed the growth of frost on a glass plate and employed an effective feature descriptor to represent natural frost textures. In this paper, we propose an innovative method to indirectly detect both dew and frost on a glass plate using computer vision. Several experiments of dew and frost observations were conducted at several standardized World Meteorological Organization (WMO) weather stations in Beijing, China, and the proposed method was validated against manual observations of common types of short grass.
2. Instrumentation and methods
Dew and frost are best detected on leaves directly, which is a difficult task for computer-vision-based approaches. Dew and frost can be recognized by several perceptual methods, such as visual inspection from different viewpoints, but a camera only obtains images of leaves from a fixed direction. In addition, wind-induced flapping makes it difficult to locate leaves. This problem is more complex when motion blurring occurs.
The setup for the field work is shown in Fig. 1. Three flat glass plates are placed horizontally over the ground, and the camera captures a 704 × 576 image once every 10 min (the spatial resolution on the glass plates is 0.1–0.2 mm). A light-emitting diode (LED) with a white cold light1 illuminates the glass plates vertically from above at night. The illumination is provided only when capturing images, thus, no artificial illumination exists during image acquisition intervals, which eliminates the illumination-induced influences on dew or frost formation. The workflow is depicted in Fig. 2. As shown in Fig. 2c, the three glass images are cropped from the original image. These images are analyzed using two detection algorithms (Fig. 2d) that are executed independently. Note that, the upper part of each plate is transparent with a smooth treatment on both surfaces, while the lower part is opaque with a frosted treatment on the ground-facing side (the upper and lower parts are described as the smooth part and the frosted part, respectively). Interestingly, the frosted glass surface becomes transparent when it is covered with water but becomes opaque again when it dries. During a dew formation period, the frosted side holds water and gradually becomes transparent, which decreases the differences between the images on the two parts. Therefore, dew events can be identified by detecting the changes between the different parts of the glass plate. Notably, the orientation of the frosted side is not arbitrary. We observed that dew is usually deposited on the ground-facing side slightly earlier than on the other side. Orienting the frosted side toward the ground reduces the error in estimating the onset of dew formation. The droplet contact angles of the smooth and frosted sides are approximately 23.0° and 21.4°, respectively. In addition, there is no hydrophobic coating on either side because the coating would only affect the type of dew pattern on the glass surface.
The first flat drosometer was introduced by Duvdevani (1964) and Takenaka et al. (2003), and they discussed measuring the amount of dew on flat drosometers with different materials. Several studies of dew patterns have used flat plates to measure dew growth. Nikolayev et al. (1998) conducted a laboratory experiment that used an optical method to measure the average radius and the surface coverage of dewdrops growing on a flat glass plate, which is similar to the method used in our setup. The difference between the methods is that Nikolayev et al. (1998) coated the glass plate with a hydrophobic layer that acts as a substrate from the optical point of view.
Experiments are conducted at several weather stations in Beijing; the geographical distribution is shown in Fig. 3. Our setup was installed in the observation fields with different types of short grass; an exception is a nearly bare area chosen at the Changping station. The images and corresponding manual records for evaluating the dew observations are from the Beijing, Changping, Pinggu, and Daxing stations. The largest subset is obtained at the Beijing station over 14 months (349 days), while the data from the others are obtained over 2 months (a total of 120 days). The data for evaluating the frost observations are from nine stations shown in Fig. 3. The subset obtained from the Beijing station covers 5 months (129 days), while the data from the other stations are distributed over 3 months (a total of 222 days). Observations are not made on several days and we discuss this in section 4.
Images to be processed by the algorithm for detecting dew and frost on the glass plates are acquired every 10 min at all of the stations. Manual observations of the grass surface are made every hour to reduce the chance of missing dew and frost. Both the proposed algorithm and manual observations are validated over an extended period from 30 min before sunset to 4 h after sunrise [see discussion in section 4c(2)]. The evaluation includes two steps. First, the representativeness of the glass plate in lieu of real leaves is validated against the hourly manual observations of the grass surface. In this step, the onset and duration of dew/frost on the glass plates are manually derived from each image by visual inspection during the period described above. Second, our algorithm is validated against results that are obtained by manually inspecting the images of the glass plates.
The numerical dew and frost detection algorithms follow a temporal–spatial approach. In the dew detection algorithm, we establish the relationship between dew formation and the image features that are associated with changes in the optical properties of the glass surface. In the frost detection algorithm, we first measure the similarities between the continuously captured images of a glass plate. These similarities are then fitted to a numerical function that describes the frost formation on the glass plate. In addition, a classification-based image analysis method is used to classify the texture information associated with frost.
An illumination normalization procedure is performed before the detection to prevent illumination variances. For the gray value I(i, j) associated with the coordinate (i, j) in the image of a glass plate I, we remove the global illumination as follows:
where avg(·) and std(·) are the average and standard derivation operators, respectively. Illumination variances influence our algorithm, as discussed in section 4c(2).
a. Dew detection
Two observations can be made by analyzing an image acquired on a representative night without dew (see Fig. 4a): 1) the smooth part of each glass plate appears transparent, while the frosted part appears opaque; and 2) the frosted part is brighter than the smooth part. Dew causes two gradual changes on the plate surface: the smooth part becomes more opaque when it is covered with light dew, while the frosted part becomes more transparent (see Fig. 4b). In contrast, heavy dew makes both parts appear rough (see Fig. 4c). In both cases, the smooth and frosted parts appear significantly different when they are dry, while they are similar when they are wet. Based on these observations, we use the intensity and gradient to describe the two dew patterns.
Let Ik(i, j) denote the grayscale image of the kth glass plate. The normalized difference of the intensity between the smooth and frosted parts of the kth glass plate is
where N is the number of glass plates, and Ak and Bk represent the regions of the smooth and frosted parts of the kth plate, respectively. NDk is large when the plate is dry, and it decreases when dew condenses on the plate. Figure 5a shows a typical change in NDk during a dew formation period. A rapid decrease in NDk begins at the 17th frame and lasts for several frames, which indicates the onset of dew formation. The dew detection is then converted to a slope detection problem that can be solved by determining the most significant descending slope. For each plate, we compute the maximum difference in NDk between the current frame and the historical frames over a particular period of time to reveal the dew process:
where t is the number of received frames. refers to NDk of the tth frame, and M indicates the largest possible formation time of dew. Figure 5b shows the maximum difference that corresponds to Fig. 5a.
Dew formation can also be described by the change in roughness on the plate surface. As shown in Fig. 6a1, the vegetated ground surface appears less homogeneous than the frosted part in the image. Similar to the changes in intensity, there are two change patterns in roughness (see Figs. 6b1 and 6c1). In our method, the average gradient magnitude is used to measure the roughness of the plate surface. We adopt the first-order derivative of the 2D Gaussian function (Canny 1986) as the kernel of our gradient operator:
For the grayscale image of the kth glass plate Ik(i, j), the normalized difference of the gradient magnitude between the smooth and frosted parts is
where the asterisk (*) is the convolution operator. Similar to Eq. (3), represents the maximum change between the current frame and the historical frames over a particular period for the kth plate. Finally, the appearance or absence of dew can be identified as follows:
where ThND and ThNG are the thresholds that ensure that significant changes in the corresponding features can be detected immediately. Theoretically, both thresholds should depend on the ground surface conditions. However, experimental results demonstrate that the thresholds are not sensitive to surface variations in most cases.
and exhibit similar trends during a dew formation period. Although each feature can be used independently, incorporating them into practical work is recommended. For instance, was less effective in the experiments that we conducted in October and November 2010 because the vegetation had almost died, and only soil and rocks were left on the ground; thus, NDk was very small or even negative. In addition, it is not appropriate to exclusively use because is more sensitive to changes, which ensures a rapid response to the onset of dew formation.
b. Frost detection
Dew and frost appear differently on the glass plates. Dew is always uniformly distributed on the entire surface, while frost may only form in specific areas. Frost may exist on either (or both parts) of a glass plate; thus, distinguishing changes between the two areas is less effective in frost detection. Nonetheless, frost can be recognized by several optical properties: 1) frost crystals are opaque and white; 2) rapid changes occur in the images at the initial stage of frost formation and then frost remains unchanged or only changes slightly over a particular period of time; and 3) different frost patterns usually have common textures. Based on these observations, we propose a two-stage approach for detecting frost.
1) Detecting changes in surface optical properties
A representative frost formation period (see Fig. 7) can be divided into the following three subperiods: 1) “free period” (Figs. 7a–c): no frost is present on the plate surface; 2) “accumulation period” (Figs. 7d–f): frost accumulates rapidly on the plate surface; and 3) “stationary period” (Figs. 7g–k): the morphology of frost remains stable and changes slightly. Note that frost continues to accumulate for the rest of the night, but the rate of accumulation is much slower than that in the accumulation period.
Negligible differences occur between the images from the free and stationary periods, while significant changes occur between the images from the accumulation period. We employ a correlation method to measure the similarities between the frames:
where and are both m × n matrices. Let denote the current grayscale image of the kth glass plate. The correlation coefficient between and each historical one is calculated as
where t is the number of received frames, and L refers to the number of frames that are associated with the maximum possible length of the accumulation period. Larger values of indicate stronger similarity between and . Let pi, i = 1, 2, 3 denote the set of in the three subperiods described above. The variance in the values within p1 or p3 is small because of the strong similarities between the images within the corresponding subperiod. Additionally, the values in p3 are much larger than the values in p1 because of the weaker similarities between the tth frame and the frames in the free period. However, the values in p2 increase monotonically. Considering the acquisition time of the images, as the images are captured closer to the tth frame, the images become more similar to those in the stationary period and less similar to the images in the free period. The numerical analysis of shows that it best fits an arctangent function of the following form:
where , , , are the model parameters. When L frames are received, the calculated is used to solve these parameters using the Levenberg–Marquardt algorithm (Levenberg 1944). For example, Fig. 8a shows the fitting result of the frost formation in Fig. 7.
A potential frost event can be identified by reasonably evaluating the slope detection model [Eq. (9)] using the following conditions:
where and refer to the upper and lower asymptotes of , respectively, and and are preset thresholds. Equation (10) ensures a strong similarity between the observation data and the prediction data . The values in p2 increase monotonically, which leads to Eq. (11). Equations (12) and (13) indicate that the values in p3 should be greater than the values in p1, and all of the values in should be positive. Equation (9) also implies a method for estimating the occurrence time of frost formation tH as
where T(t) refers to the acquisition time of the tth frame, ⌊·⌋ is the floor function that maps a real number to its largest previous integer, Tint refers to the acquisition interval, and is the so-called step point (see Fig. 8a), which has the maximum slope in .
A successful fitting occurs when Eqs. (10)–(13) hold. Because the appearance of frost changes slightly in p3, successful fittings should be achieved continuously. Assuming Eq. (9) has been estimated well N times, the step point of the nth successful fitting is
where and are the model parameters of the nth successful fitting. Note that all independent successful fittings correspond to the same frost event. In other words, the estimated occurrence time of frost in all of the fittings should be equivalent; thus, a linear relationship between and n forms
where is the prediction of , and m and p are the estimated parameters. The following conditions ensure a good estimation of Eq. (16):
where round(·) is the round-off method and Thline is a threshold. Incorporating Eq. (17) can effectively improve the slope detection model. Figure 8b plots the step points of six successful fittings that are sequentially achieved in a frost formation. The plot shows that only the results from the fourth fitting to the sixth fitting hold for Eq. (17), while the others are less clear.
2) Coding the texture information of frost
Disturbances other than frost can produce similar changes that fit our proposed model. To detect frost reliably, a validation step is introduced to recognize frost on still images. The different appearances of frost on a single image have similar textures2 according to our visual perception. In this study, we employ the completed local binary pattern (CLBP) (Guo et al. 2010) as the basic texture descriptor to encode the texture information of frost. CLBP is an extended version of the local binary pattern (LBP) feature (Ojala et al. 2002), and it uses three binary patterns to encode the local structure of a pixel of the image I as follows:
where M and N are the width and the height of I, respectively; nc and np are the gray values of a pixel and its pth neighbor, respectively; and R is the Euclidean distance between the pixel and its neighbors. As shown in Fig. 9, the morphologies of frost (Fig. 9, first row) appear different from each other, but their corresponding CLBP histograms (Fig. 9, second row) are similar.
Frost is sometimes only present in particular regions (Fig. 10, first row), which causes pixels of dry regions to be included in the feature computation. Thus, the CLBP feature is less discriminative in representing frosty and nonfrosty regions. An alternative method for limiting this influence is to employ the “spatial pyramid” scheme (Lazebnik et al. 2006), which represents an image by dividing the entire area into blocks and then computing the distributions of local features at increasingly higher resolutions. By connecting the features at different scales, one can obtain a better description of both large-scale and local structures. As shown in the second row of Fig. 10, a spatial pyramid with three scales is built to reconstruct the CLBP feature. We first resize all of the blocks at different scales to the same size as the original image and then compute the CLBP feature of each block. Finally, the features are serially connected over different blocks and scales to obtain an overall descriptor.
Note that the dimension of the CLBP feature is greatly increased by incorporating the spatial pyramid scheme. For example, by using the riu2 mode (see the caption of Fig. 9) to generate three binary pattern maps and then using 3D mapping to compute the histogram at different scales of a three-level pyramid, the number of bins in the final histogram is 26 × 26 × 2 × (1 + 4 + 16) = 28 392. Directly using the feature would increase the computational complexity and cause difficulties when training a classifier. An effective way to solve this practical problem is principal component analysis (PCA) (Hotelling 1933), which uses an orthogonal transformation to convert a set of variables into another set of linearly uncorrelated variables. PCA helps us reduce the dimension by extracting the most important components of the features. Experimental results show that using 100 eigenvectors with the highest eigenvalues in each bin at different scales is sufficient to obtain high classification accuracy.
c. Measurement of dew and frost durations
The duration of dew or frost is also an important factor that can have a positive or negative effect on a crop or ecosystem (Jacobs et al. 1990). Theoretically, the changes that appear on a glass plate that are caused by the transition between the wet state and the dry state are nearly symmetrical. The dew and frost durations can be measured simply by using the strategies for identifying the onset of dew or frost formation to determine its termination. For example, we can reverse the decision conditions in Eq. (6) to identify the drying of dew. However, the estimated termination of dew and frost may be disturbed by strong variations in illumination, which is discussed in section 4c(2).
4. Evaluations and discussions
The setup and our proposed image-based algorithm are separately validated for the datasets that are introduced in section 2. As mentioned in section 2, several days are removed from the dataset because of device failure and/or the occurrence of precipitation. Raindrops or snow can produce changes on a glass plate that resemble dew or frost and can lead to false positive results in the detection. For example, in June and July 2011 at the Changping station, 10 dew events are reported by our algorithm, but 15 rainy nights actually occurred. We observe that “rain patterns” that contain regular raindrops can be adequately recognized using the CLBP-based texture descriptor that is introduced in section 3b(2) if the classifier has been trained by the corresponding image samples. However, identifying a rain event using our method is still problematic because dew and raindrops cannot always be distinguished. The same problem occurs when discriminating between snow and frost. Consequently, false positives that occurred during a precipitation event, or that were caused by precipitation-related moisture on the glass plate are not taken into account in our experiments.3
Frost can form from the deposition of water vapor or from freezing dew. In the latter case, the dew duration in our experiment was usually very short. This dew can be manually observed on the glass plates and can be detected by our algorithm due to the short interval of image acquisition (10 min). However, this dew is sometimes missed in the hourly manual observations of the grass surface. For a fair comparison, only frost is recorded in this case.
a. A glass substrate’s representativeness of real leaves
Our proposed sensor is an indirect method for observing dew and frost; therefore, we first discuss how well the glass plate represents real leaves. This representativeness can be evaluated in terms of the translation rate of dew and frost from the grass to the glass plates. We consider an observed night as dewy or frosty when dew or frost, respectively, was manually observed on the grass that night; otherwise, the night is considered dry. A successful translation means that dew/frost can be identified in the images of the glass plates by visual inspection on a dewy/frosty night, while the images should remain unchanged on a dry night. Table 1 shows that our setup successfully translates 93% of the 469 observed nights, including 233 dewy nights that are determined by hourly manual observations of the grass surface. A significant decrease in the translation rate (65%) occurred in June 2011 at the Changping station, where our setup was placed over a nearly bare soil surface. The reason is that grass surfaces are insulators and therefore cool more rapidly than bare soil surfaces at night. The frost observations are compared in Table 2, which shows that 92% of the 453 observed nights, including 130 frosty nights that are determined by hourly manual observations of the grass surface, are correctly translated by our setup.
The dew and frost durations were also observed at the Beijing station. Figure 11a compares the dew durations observed by manually comparing the images of the glass plates with hourly manual observations of the grass surface, and the frost durations observed by the two methods are compared in Fig. 11b. Of the 1395 observed dew hours and 148 frost hours on the grass surface, 86% of the dew hours and 82% of the frost hours are successfully observed on the glass plates.
Interestingly, the dew and frost durations on the glass plates are height dependent. Figure 12a shows the dew durations that were observed by manually inspecting the images of the three glass plates in September 2010 at the Beijing station. On most dewy nights, the dew duration is longer on the lower glass plate. However, the Beijing station data from October 2010 to March 2011 indicate that the frost durations do not exhibit the same regularity (Fig. 12b). Of the dew hours obtained by hourly manual observations of the grass surface, 94%, 78%, and 71% are manually observed on the three glass plates, respectively, which are sorted by ascending height (Fig. 12a). The corresponding percentages for the frost durations are 70%, 61%, and 62% (Fig. 12b). Using the earliest onset and the latest presence of wetness on the three glass plates, the percentages for dew and frost are 95% and 82%, respectively. Clearly, considering all of the plates improves the translation of both events.
b. Evaluations of the dew and frost detection algorithm
In this section, the performance of the algorithm is validated against manual observations of the glass plates. Similarly to section 4a, we classified an observed night as dewy or frosty if dew or frost, respectively, was manually observed in the images of the glass plates acquired that night. Meanwhile, the detection of the onset of dew or frost formation is validated by evaluating the ability of our algorithm to discriminate among dewy, frosty, and dry nights. Additionally, we provide the detection accuracy when the results obtained by the hourly manual observations of the grass surface are considered to be the ground truth data.
1) Algorithm’s performance in detecting dew on the glass plates
All of the following parameters are kept the same in the evaluation: M = 10 in Eq. (3), ThND = 17 in Eq. (6), and ThNG = 11.5 in Eq. (6). A dew event is identified if it is recognized by our algorithm on at least one of the three glass plates. Table 3 shows that our algorithm achieves an accuracy of 91% in discriminating between dewy nights and other nights, which were classified by manually inspecting the images of the glass plates. The false positive rate and false negative rate are 2% and 18%, respectively. The false negatives contain important information on the algorithm’s performance in discriminating between dew and frost. Only 6 of the 225 dewy nights listed in Table 3 (approximately 3%) are mistakenly recognized as frosty nights. It is challenging for our algorithm to precisely measure the dew duration on the glass plates. Figure 11c shows that 75% of the 1298.04 dew hours that are manually observed from the images of the glass plates are detected successfully.
2) Algorithm’s performance in detecting frost on the glass plates
The evaluations of frost detection are divided into two parts. First, the effectiveness of the proposed feature is evaluated by a classification test on a subset of images that are manually selected from the entire dataset. Second, the proposed algorithm is validated by evaluating its performance in deriving frost information from the glass plates.
The image samples for the training and testing are selected from the nine stations described previously. First, image patches corresponding to the smooth and frosted parts of each glass plate are separated into two groups. Second, for each group, the patches are manually classified into three categories: dew,4 frost, and none. The distributions of samples in the smooth/frosted groups are 1) dew: 635/545; 2) frost: 2561/2215; and 3) none: 4029/3203. Figure 13 shows several typical samples from the smooth group.
For each group, a support vector machine (SVM) with a radial basis function (RBF) kernel is trained using 400 samples that are randomly selected in each category. SVMs are a set of supervised learning approaches with associated kernel functions for pattern recognition, and RBF is usually a reasonable first choice in kernel functions (Keerthi and Lin 2003). The kernel parameters are optimized by cross validation on the training set in the grids: log2 C ∈ [−5, 15] and log2 γ ∈ [−15, 3] (with a grid step size of 2). Table 4 lists the classification results using the proposed feature and other LBP-based descriptors. In Table 4, u2 indicates that the binary pattern codes are uniform, while riu2 means that the uniform codes are also invariant to rotation. Terms P and R have the same meanings as in Eq. (18). The classification results show that our proposed feature outperforms the others for both groups. Note that, all of the descriptors achieve higher classification accuracies in the group associated with the frosted parts of the plates than in the other group. This finding is likely attributable to the opaqueness of the frosted part of the glass; that is, the corresponding images are free of the influence of vegetation variation.
Our proposed method for frost detection includes two steps. First, a potential frost event is explored by the slope detection model [see section 3b(1)]. Second, frost is recognized using the classifier (SVM) that was trained in the first part of this evaluation. The same parameters are used in the evaluation as follows: L = 10 in Eq. (8), ThcorrVal = 0.95 in Eq. (10), ThupdnGap = 0.15 in Eq. (12), and Thline = 0.96 in Eq. (17). A frost event is identified when frost is recognized on either the smooth or frosted parts of one glass plate. Table 5 shows that our algorithm achieves an accuracy of 91% in discriminating between frosty nights and other nights that is observed by manually inspecting the images of the glass plates. The false positive rate and false negative rate are 2% and 22%, respectively. In the false negatives, 6 of the 151 frosty nights listed in Table 5 (approximately 4%) are mistakenly identified as dewy nights. Figure 11d evaluates the algorithm’s performance in detecting the frost duration on the glass plates and shows that 77% of the 127.06 frost hours that are manually observed from the images of the glass plates are detected successfully.
3) Overall performance of our method in detecting dew/frost on grass
The evaluations in sections 4a, 4b(1), and 4b(2) provide an indirect comparison between the hourly manual observations of the grass surface and the image-based detection. The results show that our algorithm has an accuracy of 89% when discriminating between dewy nights and other nights that is observed by manually examining the grass surface, while 3% of dewy nights are falsely recognized as frosty nights. The accuracy of the frost detection is 90%, and 6% of frosty nights are falsely recognized as dewy nights. Additionally, 71% of the dew duration and 68% of the frost duration on the grass are detected successfully.
The evaluations in section 4a show that our setup is capable of representing the onset of dew and frost formation on short grass, but it has problems translating their durations. We believe that the differences in the thermodynamic properties of leaves and the glass substrate are the leading cause of the decrease in the translation rate. Additionally, slight visual changes on a plate sometimes cannot be observed immediately when dew or frost forms, and visually determining the precise moment that dew begins to freeze is also difficult.
The proposed algorithm plays a crucial role in the sensor’s performance. The experimental results [sections 4b(1) and 4b(2)] show that our algorithm performs well when detecting the occurrence of dew or frost on the glass plates, but it produces less precise results when measuring its duration. Two common problems are identified by analyzing the failures and are discussed in detail below.
1) Difficulty in identifying low levels of wetness
Compared with the false positive rate, our method has a higher false negative rate in detecting the onset of dew and frost formation on the glass plates because our method is less effective in the presence of weak wetness. Such a dew event was observed on 14 November 2010 at the Beijing station, and Fig. 14a shows three images of a glass plate that were acquired around the onset of dew formation. However, no distinctive changes occur between the two parts of the third image, except for some blurring on the smooth part. Similarly, changes caused by a weak frost event may not fit the slope detection model well because Eq. (12) cannot always be established during the very slow accumulation of frost. For example, in the failure that occurred on 24 November 2010 at the Huairou station (see Fig. 14b), frost is only present in the areas surrounded by red ellipses, while the other areas remain unchanged. Additionally, frost patterns that are produced by a weak frost event may not be recognized correctly. These unusual patterns, which are formed by small frost crystals that are scattered across the glass plate, were observed on 4 February 2011 at the Pinggu station (see Fig. 14c).
2) Algorithm’s performance in different illumination conditions
Our algorithm works best during homogeneous lighting conditions. However, variations in illumination can be produced by changing cloud cover or, more generally, by variations in solar elevation during particular periods around sunset and sunrise. The variance mainly causes a uniform intensity change on the entire surface of the glass plate because of its small area. In practice, this influence has a minor impact on our algorithm because it can be effectively eliminated by the illumination normalization technique [see Eq. (1)]. Additionally, our proposed texture descriptor for recognizing frost is robust to illumination variance. However, influences that are caused by local illumination changes, such as shading or strong reflection on partial regions of a glass plate, are still a problem.
As introduced in section 4, the detection begins 30 min before sunset so that local illumination changes can be effectively avoided, though a few very early events might be missed by our algorithm. We found that of the 233 dewy and 130 frosty nights that were identified by the hourly manual observations of the grass surface, only nine and two false negatives, respectively, were induced by early events. Such early events may also delay the estimated onset of dew or frost formation. Our algorithm may fail to identify changes on glass plates that are already wet, while the occurrence time of wetness is more likely to be obtained from other plates in this case. In the evaluation, the onset of dew and frost formation is delayed 19 and 6 times, respectively. The delays are irregular because the time interval of the appearance of dew or frost on different plates may vary from 10 min to 1 or 2 h.
The detection is terminated 4 h after sunrise for the purpose of measuring the dew and frost durations. Based on our observations, the illumination variance can have a significant influence on identifying the drying of a glass plate. Figure 11c shows that the detected dew duration is shorter than that manually derived from the images of the glass plates in most months. Most of the missing hours are caused by false negatives, while others are due to the incorrect prediction of dew termination. Normally, both parts of a glass plate start to become wet at nearly the same time dew forms (see Fig. 14d, left). However, because of the uneven sun exposure, dew may evaporate asynchronously on different parts of a plate and produce a recognizable difference (see Fig. 14d, right). In this case, the calculated termination time of dew is normally 30–60 min in advance of dew termination.
Conversely, the algorithm often produces prolonged frost duration because of the delayed prediction of frost termination. Figure 11d shows that the detected frost duration is longer than the manually derived duration from the images of the glass plates over many months. We found that most failures are caused by false positives in the classification. And shown in Fig. 14e, melted water may produce strong diffuse reflections of sunlight (see Fig. 14e, right) that resemble a frost-covered surface (see Fig. 14e, left). In this case, the calculated termination time of frost is normally delayed 10–40 min after frost termination.
d. Comparison with other flat drosometers for daily monitoring
A practical problem for daily dew and frost observations is that all types of contact surfaces, such as ours and those on artificial leaf wetness sensors (Decagon Devices 2009), would be contaminated by dirt when they are placed near the ground surface. An advantage of our approach is that the regular maintenance can be performed less frequently compared with other techniques. On the one hand, our algorithm is not sensitive to the accumulation of small amounts of dirt or small foreign objects. As shown in Fig. 15a, the glass plate on the right is partially covered by dirt, which is difficult to visually identify when dew forms (see Fig. 15b). In addition, Figs. 15c and 15d show three glass plates that are covered by several dead branches, which are different from dirt because they remain visible in the images when dew forms. However, the branches only slightly affect our dew detection method, which is based on statistical features. In the computation, the disturbances in small areas only weakly contribute to the overall description of the entire glass plate. On the other hand, the side of the glass plate that faces the camera is quite smooth, so dirt on the surface can easily be washed away by rainfall.
Additionally, compared with the one-dimensional time series data that are recorded by wetness sensors, intuitional image data contain a wealth of information that could potentially be used to explore other properties of dew and frost, such as dew patterns.
e. Scope for future improvement
Because this study represents the first attempt to use a computer-vision-based approach for detecting dew and frost, there are still some points left for improvement. In future research, we plan to improve our work in two ways. First, the performance of classical offline training and online classification schemes for frost detection are influenced by the selection strategies of the training set. A straightforward solution might be achieved using an online learning procedure. Second, a decision support system will be developed by combining this method with other essential meteorological parameters. For example, dew and frost events greatly depend on the dewpoint temperature and the amount of water vapor in the surrounding air. Because this information can be obtained automatically, its integration into our approach should produce further improvements.
5. The limitations of our approach
The quantification of dew amounts has been an essential aspect of several studies. However, our approach cannot currently perform this calculation. Theoretically, the amount of dew can be quantified using an image-based method if one can correlate dew patterns with the amount of dew on a substrate. Inspired by Duvdevani’s work (Duvdevani 1947), a similar approach might work if two improvements are made to our setup: 1) replacing our glass substrate with one that is similar to the wooden blocks that are used in Duvdevani dew gauges, such that various dew patterns can be categorized; and 2) improving the spatial resolution of the image, so that different dew patterns can be visually classified. In addition, a discriminative classifier should be trained to distinguish dew patterns.
Measuring the amount of frost represents a greater challenge. In addition to the improvements discussed above, several other problems should be considered. First, frosty areas on a substrate should be precisely measured because frost may only form in particular localized areas (see section 3b). Second, frost patterns with significant visual differences have similarly distributed CLBP histograms (see Fig. 9); specifically, they have common attributes in their inherent textures. Compared with dew patterns, frost patterns are more difficult to classify because they have very small interclass distances in the feature space.
In this article, we proposed a new sensor for dew and frost detection that is based on computer vision. The sensor’s performance was validated by manual observations of short grass at several weather stations in Beijing, China. The experimental results demonstrate that more than 90% of dewy, frosty, and dry nights that are identified by hourly manual observations of the grass surface are successfully identified on the glass plates by manually inspecting their images. More than 80% of the dew and frost durations on grass are also correctly identified. The proposed algorithm produces an accuracy of more than 90% when discriminating among dewy, frosty, and dry nights that were identified by manually inspecting the images of the glass plates. Precisely measuring the dew and frost durations on the glass plates is challenging, and our algorithm correctly detects approximately 70% of the durations. Therefore, there is still room for improving our algorithm. Finally, we noted the limitations of our method and analyzed the prospects for future work.
This work was supported by the China Meteorological Administration under the R&D Special Fund for Public Welfare Industry (meteorology): GYHY200906032. The authors are grateful to Kejun Wu, Xiaobing Zhang, Xiangang Wen, Yang Xiao, Xiyao Duan, Yi Xiong, Yi Zheng, Zhenghong Yu, Xiaodong Bai, Mengni Ye, Liang Ye, Yanan Li, Hao Lu, Zhiwen Fang, Chunhua Deng, Xiaoxia Li, Bo Du, Fa Tao, Xiaoyan Qiao, Jinyue Zhang, Aiqun Zhang, Shujie Li, Dongmei Wang, Shen Chai, Minfeng Sun, Haiyan Yu, Mingshui Liao, and Naijun Zhao for their valuable work in obtaining the hourly dew and frost observations.
The power consumption and the light-output power of the LED source are 12 and 6 W, respectively; and the LED viewing angle is 30°. The LED source was manufactured by Mei Sai De Technology Co., Ltd. (http://www.hhyelec.com.cn) and the product code is SD-BG8060ZR.
Texture is a visual feature that describes the spatial arrangement of intensities in an image and is not affected by variations in color or illumination.
The precipitation information was obtained by an infrared forward scatter present in the weather sensor 1.5 m above the ground. The sensor was manufactured by CAMA (Luoyang) Environmental Measurement Co., Ltd. (http://www.camaem.com) and the product code is CJY-2C/T.
Dew patterns with significant dewdrops may be falsely recognized as frost. Therefore, an additional category composed of the image samples with large dewdrops was added to make the classifier more discriminative.