Validation of leaf area index measurement system based on wireless sensor network

Study area

With the advanced observational techniques, abundant data accumulation, and ability to carry on multi-scale experiments, the Huailai Remote Sensing Station and around (for short Huailai Station), located in Huailai, Hebei province, China (40.349°N, 115.785°E), becomes one of the ideal study areas for the observation and validation of the LAI²⁷. The Huailai Station is mainly covered by corn and some weeds. So, we mainly use LAIS to monitor the growth cycle of corn (in April 2015, we submitted an application for plant collection permission to Huailai Remote Sensing Station and obtained approval.)

Huailai WSN vegetation monitoring system includes 6 sets of monitoring equipment, and its distribution is shown in Fig. 1 as follows, in which red dot represents LAIS Node, purple frame represents MODIS pixel, red frame represents observation area. The observation system is designed for the application of remote sensing pixel scale authenticity tests. The observation scale is a 1 km MODIS pixel on the pixel scale, and the actual coverage area is 2 km * 1.5 km. The six sets of equipment cover the core area of the test station and the surrounding typical growth plot, which is a good representative of the 1 km pixel scale.

Each piece of equipment consists of two cameras which were only one camera with two different angles in previous work²³ set up at a height of 2.5–4 m above the ground (Fig. 2), one for vertical downward observation and the other for inclined observation, which can take canopy photos regularly every day at its fixed position. The observation system obtained the photos of the corn canopy from May to August, but the corn did not grow in August. Therefore, in this study, we selected the photos taken by the vertical observation camera of the corn sample plot in the experimental station from May to July 2015.

Related work—data acquisition

Data collection using LAIS

The data collection complies with the plant guidelines statement: “LAI-2000 Plant Canopy Analyzer Instrution Manual” (Supplementary Information 2) (https://www.licor.com/env/, Last visit time: 21 October 2021). Existing facilities such as the high poles and the wireless sensor network in the experimental station have proved convenient for the installation of the LAI measurement system. LAIS uses the GEO001 digital serial camera that is suitable for a variety of embedded image acquisition modes. The specification of the camera includes: the total field of view is 120°, the maximum image size is 2176 * 1920 (approximately 5 million pixels), mounted at a height of 3 m, the spatial resolution at ground level is about 3 mm. The acquired image is simultaneously stored in a flash card in two formats: the JPEG format merits in less file size thus suitable for quick wireless transfer; the RAW format, which is the user data in our analysis, contains 3 channel binary image in 10 bits bit-depth. Compared to our previous work, an important new feature of this camera is the programmable cut-off filter. As we know, unlike scientific sensor which has the precise spectral response to each band, the digital camera is cheap and can only acquire the so-called RGB image. Usual digital cameras have one NIR cut-off filter to exclude the near-infrared light. The GEO001 camera, which was a commercial camera produced by Zhongshan Yunteng Photographic Equipment Co., Ltd, has two cut-off filters: one is the NIR cut-off filter, another is a blue cut-off filter. Switching on the NIR cut-off filter results in an ordinary color image as in a usual household digital camera. While the blue cut-off filter is switched on and NIR cut-off filter is switched off, near-infrared light is allowed to reach the detector array and blue light is blocked, resulting in false-color images as in Fig. 3b. Adding near-infrared light can increase illumination in the shadow area, and blocking blue light can alleviate the disturbance of sun glint, so, switching to a blue cut-off filter helps to improve the image quality when the direct sunlight is strong such as around noon time.

To acquire an image in the best illumination condition and avoid the influence of rain or other unsuitable weather, the image acquisition device based on WSN was set up to acquire images three times per day: 5:30 a.m., 1:30 p.m., and 6:30 p.m. According to our experience, when the canopy is open (sparse vegetation), usually images acquired at 6:30 p.m. are the best for classification because the direct sunlight is weak; when the canopy is closed (dense vegetation), the illumination on the soil background is very poor in all time, and classification is difficult. So, the camera is programmed to switch to a blue cut-off filter when acquiring images at 1:30 p.m., while the images acquired at other times were with NIR cut-off filter, resulting in true color images, as shown in Fig. 3.

LAILLW data and LAI2000 data

To evaluate the accuracy of the improved finite length averaging method proposed in this study, a field experiment was carried out to measure LAI by manual sampling (Supplementary Information 3,4). A field sampling scheme covering the corn growing season (late May to early July) was designed (Supplementary Information 1). The LAI of corn in the experimental area was measured by the quadrat harvesting method, and the validation data of LAI of corn in each growth period were obtained. Considering the rapid growth of the corn, the sampling experiment period was set as 1 week, but due to the actual work in summer and the influence of rainfall, six effective measurements were carried out in the field experiment: May 30, June 7, June 13, June 20, July 4 and July 16.

The LAILLW method, which is also known as the shape factor method, involves outdoor and indoor measurements. The formulas are:

where L represents the leaf area index, S refers to the area of a single plant, and N refers to the number of plants in a unit area. The shape factor ƒ is the ratio of the S to the value multiplied by the length and width of all leaves in the plant.

To reduce measurement errors, 10 plants were selected in the sample, and the length and width of each leaf on each corn were recorded with a ruler. To obtain the shape factor, representative corn plants were cut next to the sample (not in the image coverage area) and the true area of each leaf was obtained by software, and the shape factor was derived from this²³. Through the length and width of 10 strains measured in the field, and the shape factor obtained, the total leaf area of 10 corns can be calculated, and the average leaf area of one plant is finally obtained. The LAI value under the LAILLW method is obtained.

Using the difference between the solar radiation values of the upper and lower canopies, the LAI2000 canopy analyzer can obtain LAI and set up a corresponding point folder to save the measured data for subsequent collation. 10 measurement points were selected for each site, and the average value was the final result for each site. To reduce the effects of the solar altitude angle on measurement accuracy, the experiments were repeated every two hours.

To make it easier to record the date of data acquisition, the data were summarized in the order day of the year (DOY). For example, 30 May 2015 is the 150th day in the year and its DOY is 150. The DOY information of data acquisition using the LAILLW method and LAI2000 is specifically shown in Table 1.

MODIS LAI data

MODIS leaf area index data was downloaded from the United States Geological Survey (https://modis.gsfc.nasa.gov/data/dataprod/mod15.php), named MCD15A2Hv006. It is an 8-day composite dataset with a 500-m pixel size. The algorithm chooses the best pixel available from all the acquisitions of both MODIS sensors located on NASA’s Terra and Aqua satellites from within the 8 days.

In the comparison of MODIS LAI data, as the pixel of the satellite product is in 500 m resolution, it is not recommended to directly compare single node LAIS measurement with the MODIS LAI product because of the scale mismatch. Though complicated upscaling approaches have been discussed and implemented in Huailai station for other parameters²⁸, it is not the purpose of this study So, we simply averaged the LAI in all the LAIS nodes to compare to the average MODIS LAI product in the 3 * 3 nearest pixels (1.5 km * 1.5 km), referred to as MODIS LAI_Mean in a later context, which approximately covers the area of all LAIS nodes. Time matching was carried out by selecting the date of the MODIS product closest to the date of the handheld LAI2000 measurement. The following Table 2 is obtained by taking 3 * 3 pixels closest to the LAIS Nodes.

Improved LAIS methods

In previous work, we have deployed sensors and cameras, and also have an automatic image processing and preliminary method of calculating LAI²³. Figure 4 is a flow chart of our work. The previous articles focused on hardware and system implementation but did not pay much attention to performance. On this basis, we upgrade the image classification method and LAI calculation method, which will be explained in detail below.

Image preprocessing and classification methods

Because of weather-related factors such as water vapor and dust or inaccurate exposure, a small number of the photographs are not clear. Besides, some of the image data cannot be decoded because of unstable communications and other factors. Therefore, it is necessary to check and select the photographs that meet the processing requirements before binary image processing. Currently, the selection process is carried out by human visual inspection based on the following principles: (1) when the canopy is open (sparse vegetation), the image at 6:30 p.m. is preferred, when the vegetation the canopy is closed (sparse vegetation), the image at 1:30 p.m. is preferred; (2) if the preferred image is not clear, other clear image acquired on the same day should be used; if all the images are not clear, then this day is marked as a failure.

If we decided to use the image acquired at 1:30 p.m. It is also necessary to convert it from a false-color image to a true-color-like image (as shown in Fig. 3b) in which the leaves are shown in green color. The conversion is carried out by multiplying the vector of DN (digital number) of 3 bands with a coefficient matrix which is provided by the camera manufacturer. Another preprocessing is to choose the near nadir-view area of the image for further processing. As the off-nadir-view area of the image is subject to large geometric distortion as well as saturation of fraction of vegetation cover (FVC), they are not used in this study. The images are clipped to an ROI (region of interest) of about 2 * 2 square meters in ground area, with a maximum view zenith angle less than 30°.

The study of the color spatial distributions of the crop images is helpful for the classification of the images and extraction of the image information. The color of the image pixel is the most direct and effective element that can be used to describe the image²⁹. Because the red–green–blue (RGB) color space has the characteristic of a clear and convenient expression of information. When corn leaves are small, the crops in the fields are sparse, and most of them are soil background in the images. The soil in a lower hue is similar to the corn in terms of R and B components, while it has an overlap with the corn in G components when soil is in a higher hue. This makes it difficult to classify sparse corn scenes only by RGB space, so it is necessary to consider the characteristics of hue, luminosity, and saturation (HLS) spatial components.

Statistical analysis showed that the component values of the crop leave in the RGB color space were in the ranges of G > R and G > B while the corresponding values for the soil follow the law that B < G < R³⁰. In the HLS color space, the H component has a specific distribution law³¹. Therefore, the distributions of R, G, B, and H components were used as an important criterion for crop image classification³².

In this study, two threshold classification methods were proposed, as shown in Table 3 The default thresholds (t1, t2, t3, t4, and t5) were recommended by the system when operating each method.

The condition “t1 < H < t2” corresponds to the typical range of H component of green leaves in HLS color space, and “S > t3” is for the occasion when the junction of green leaves and soil in the H component is not very high. The intersection of two conditions can accurately distinguish the general green leaves. “G > t4” can distinguish green leaves with high illumination and “R < t5 and B < t5” can obtain green leaves in the case of dry soil background. The three conditions can get higher classification accuracy. In the experiments, the crop canopy images that were obtained when the leaves are small and the corn is sparse in the field were suitable for use with method 1 and the first four corresponding default thresholds were 80, 160, 18, and 240. When “t5 = − 1” the soil is moist and when “t5 = 40” the soil is a little dry.

When the leaves were larger and they occluded each other, changes in the chlorophyll content and enhancement of the sun-light resulted in a low contrast between the leaf and the background in the true-color images when the growth period reached two months into the panicle stage, so it was no longer appropriate to use classification method 1. The default thresholds in method 2 were 5, 5, − 1, 80, and − 1. False-color images are acquired in raw format. First, need to convert the RAW format to BMP format. Then replace the order of the image bands of G and B in professional software such as ENVI.

In the classification of crop canopy images, only the sub-region in the center of the photograph was used because of the geometric distortion of the camera photographs, and the corresponding area in each binary image had dimensions of approximately 2 × 2 m. Figure 5 shows the original and binary images of site 1 at 1:30 p.m. and 6:30 p.m. on July 1. By visually comparing the original image with the binary image (Fig. 5), it can be seen that the classification result of the false-color image is better when the leaves are large and occluded, the contrast between the lower leaves and the soil is low. This is also another significant finding in this validation experiment and it’s of practical significance for the application of the proposed system to farmland monitoring.

During the application, it is necessary to select the appropriate image as well as classification methods by the practical situation such as the crop type, the crop growth stage, and the image quality. We have mentioned in “Data collection using LAIS” section that when the canopy is sparse, we prefer to use the image at around 6:30 p.m.; if the illumination on soil is too poor as the result of the dense canopy, we prefer to use the false-color image at around 1:30 p.m., and the above color space transformation is required. Sometimes the weather is bad and the preferred images are not of good quality, we need to choose one of the three images at 5:30 am 1:30 p.m., and 6:30 p.m. according to the situation. Currently, this choice needs human visual inspection, but it is a quick and simple operation, and won’t waste much time as in the field measurement. Although most of the image processing is automated, human supervision is still essential, as the image quality is important to distinguish leaves from soil background, which is a prerequisite for calculating LAI. Figure 6 shows the binary images for the different dates in site 1.

The improved finite length averaging method for LAI estimation

The classification discriminates pixels of green leaf from soil background in the LAIS acquired image. So, the fractional vegetation coverage (FVC) and gap probability, which equals 1-FVC, can be derived by dividing the number of leaf pixels by the total number of pixels in the ROI, assuming the change of view zenith angle within the ROI can be neglected. Theoretically, LAI can be related to FVC through the widely used Beer-Lambert law.

However, the original Beer-Lambert law only applies to uniformly distributed and infinitesimal leaves. For the real canopy, leaf angle distribution and clumping index (CI) should be considered^1,24. The finite length averaging method was proposed in 1986 to simultaneously estimate LAI and CI from field measurement of canopy gaps. It also applies to gap data generated from canopy photography and still is the recommended method to estimate LAI and CI for crop canopy up to now. The formula can be summarized as:

where L denotes LAI and (Omega) denotes CI; ({P}_{i}(theta )) is the gap probability when observation zenith angle is (theta), which is 0 in this case; (G(theta )) is related to leaf angle distribution and is normalized projection area of leaf in the observation direction; the footnote (i) denotes the ({i}{th}) sample line/sample rectangle, and ({text{m}}) is the total number of sample lines/sample squares. When dealing with gap data generated from canopy photography, usually the sample square is adopted instead of the sample line.

However, Lang and Xiang²⁴ also pointed out two limitations with this approach: (1) If the gap probability is equal to 0, then the estimation is meaningless as infinity occurs. This is often the case of a high LAI. (2) the method assumes that the leaf size should be sufficiently small relative to the side of the sample square. If the sample square is too small, then the precondition cannot be satisfied, and usually, an overestimation of LAI occurs.

As the camera of LAIS is fixed on the top of a pole, the acquired sample image is of limited area. Then we face the problem of either using a small sample rectangle or the total number of sample rectangles is insufficient. To solve the problem of a small sample square, we proposed empirical formula to replace the log function in characterizing the relationship between gap probability and LAI in the sample square based on computer simulations³³. The new formulas correct the shortcomings of over-estimation and instability of log function when the canopy is dense and the side length of the sample square is short. The revised formula is:

where D denotes the equivalent leaf length, which is defined as the square root of the average area of the single leaf; and (f) is the proposed empirical formula, in the form of:

where ({a}_{ij}) ((mathrm{i}=1,…,6; j=mathrm{1,2},3)) are empirical coefficients with values in Table 4. W is the side length of the sample squire. Our study also found that the optimal setting for the side length of the sample square is about 3 times of equivalent leaf length in most cases of crop or grass scenes.

Source: Ecology - nature.com