Comparison Between Vegetation Index Obtained by Active and Passive Proximal Sensors

Proximal sensors have been used to characterize the crop biophysical properties by reflectance values and/or using Vegetation Indices (IV). Our goal with this work is to compare NDVI (Normalized Difference Vegetation Index) spectra-temporal profiles obtained by active (GreenSeeker 505 Handheld) and passive (FieldSpec4 model Standard-Res) proximal sensors to monitor soybeans and beans. It was monitored agricultural fields with soybeans Nidera 5909RG variety and beans Imperador variety, located in the municipality of Cascavel, Parana state, Brazil. The proximal sensors were used to monitor the crop's conditions on different Days After Sowing (DAS). NDVI from FieldSpec4 (NDVI FS) showed a higher correlation with GreenSeeker NDVI (NDVI GS) in the wavelengths of 649 nm and 771 nm for soybeans (rs = 0.9105) and 646 nm and 792 nm for beans (rs = 0.9382). The inter-calibration of NDVI GS values in function of NDVI FS, considering the entire phenological cycle, resulted in RMSE = 0.0520 and dr = 0.8630 for soybeans and RMSE = 0.0636 and dr = 0.8890 for beans. NDVI values showed saturation during the major vegetative development of the crops, interfering in the inter-calibration process. In general, the NDVI GS and NDVI FS were similar in terms of their spectral-temporal pattern. According to our results, the active sensor could be used to crop monitoring, resulting in a lower cost and less climatic interference.


Introduction
Food security is a global concern due to the continued population growth and the stagnant number of food production areas (Harfenmeister, Spengler, & Weltzien, 2019). In this context, soybean and beans, present in many people's food base, are two of the more important crops related to the food demands. There are efforts from farmers, technical, and direct stakeholders to guarantee the maximum food production sustainably. Therefore, accurate spatial and temporal crop information is necessary (Kenduiywo, Bargiel, & Soergel, 2016). Remote sensing data, mainly through spectro-temporal profiles, could be related to the crop biomass, allowing monitoring crop phenological development (Johann, Vieira, Lamparelli, & Duft, 2013). However, due to the crop dynamics (Atzberger, 2013), temporal data frequency is important for periodic monitoring (Atzberger, 2013;Johann, Becker, Uribe-Opazo, & Mercante, 2016). The Vegetation Indexes (VI) are spectro-temporal examples that could be used to express the phenological crop cycle (Almeida et al., 2015;Formaggio & Sanches, 2017;Johann et al., 2016). NDVI (Normalized Difference Vegetation Index) is a well know VI, which was proposed by Rouse et al. (1973). Using the contrast between the red and infrared channels, NDVI is related to agronomics variables (e.g., plant height, leaf area index, dry matter, yield, intercepted photosynthetically active radiation) (Baret & Guyot, 1991). Most of the sensors can provide NDVI values. However, the wavelengths and bandwidth could be different (Kim, 2010), resulting in different patterns.
FieldSpec4 sensor costs tens of thousands of dollars more than the GreenSeeker 505 Handheld sensor. Thus, our goal is to compare the spectro-temporal NDVI profiles obtained from both sensors for soybean and beans fields monitoring. This comparison allows us to identify the possibility of using a more affordable sensor to provide crop monitoring and propose an inter-calibration process between NDVI sensors values.

Study area
In the study, two fields were monitored in the municipality of Cascavel, Parana state, Brazil ( Figure 1). This municipality is recognized due to the high technology on the farmer level (Souza, Mercante, Johann, Lamparelli, & Uribe-Opazo, 2015) and as one of the largest national producers of soybeans, wheat, corn, and beans (IBGE, 2020). Soybeans and beans fields have areas of 78.32 ha and 40.20 ha, respectively. According to Koppen classification, both fields have a Humid subtropical climate (Cfa) (Aparecido, Rolim, Richetti, Souza, & Johann, 2016). The soybeans field has a 6% slope (range of 1% to 15%) and the beans field has a 7% slope (range of 2% to 20%) (Brasil, 2008;Embrapa, 2009). Eutrophic Red Nitosol soil type is present in the beans field and Red Dystrophic Latosol soil type is present in the soybeans field (Santos et al., 2011).

Field data
Soybeans, Nidera 5909RG variety, was sown on 2015 October 6 th and 7 th and harvested during 2016 on February 3 rd and 10 th . Beans, Imperador variety, was sown on 2016 March 3 rd and 4 th . However, due to a frost incident on May 1 st and 2 nd (DAS 59 and 60), beans were not harvest. Both crops had 0.45 m of rows spacing. In the study, were monitored 24 points on the soybeans field and 20 points on the beans field. The sampling data field was collected with 34, 45, 71, 76, and 110 Days After the Sowing (DAS) for soybean and 0, 10, 25, 43, 55, and 70 DAS for beans ( Figure 2). The field missions were influenced by the climatic conditions (rainy, cloud frequency, and frost incident) as shown in

Terrestrial sensors
The active sensor GreenSeeker 505 Handheld (Trimble Agriculture, Westminster, CO, USA) and the passive hyperspectral sensor FieldSpec4 Standard-Res model (ASD, Boulder, CO, USA) were used in the study. Few studies used these proximal sensors at the farm level in this study region (Cattani et al., 2017;Prudente et al., 2019;Viana et al., 2018), but they did not compare the NDVI values from different sensors. GreenSeeker 505 Handheld register the reflectance on two regions of the electromagnetic spectrum, red wavelength (RED GScentered at 0.66 µm ±0.0125 µm) and near-infrared (NIR) wavelength (NIR GScentered at 0.78 µm ±0.0125), both with 0.025 µm of spectral resolution. This sensor also provides the NDVI values with these bands, which was called NDVI GS (TRIMBLE, 2010;Yao et al., 2013). FeldSpec4 has a field of view (FOV) of 25° and can provide data with intervals of 0.001 µm in a range of 0.35-2.5 µm in 0.001 intervals (ASD, 2015). To compare the NDVI for both sensors were used the same GreenSeeker 505 Handheld bands regions (red: 0.645 µm to 0.675 µm, and NIR: 0.765 µm to 0.795 µm) to calculate the FieldSpec4 NDVI values (NDVI FS), this way has 961 different NDVI FS with the NDVI GS.
In both sensors, with a high of 0.80 m above the canopy, were covered 0.6 m horizontally on the ground (Cattani et al., 2017;Prudente et al., 2019;Stocker et al., 2019;TRIMBLE, 2010). It was provided three replicates between different crop rows for each point in each field mission (e.g, each DAS). For the GreenSeeker 505 Handheld, it was walked approximately 10 m, with a speed average of 0.5 m.s -1 , recording 10 samples per second, totalizing between 150 to 200 samples for each point (Cattani et al., 2017;Sharma, Bu, Denton, & Franzen, 2015;Zheng et al., 2016). Considering the FieldSpec4, three samples were collected for each replicate in the time (one-second peer sample), totalizing nine samples at each point (Prudente et al., 2019;Viana et al., 2018Viana et al., , 2017. As FieldSpec4 is a passive sensor, it was used a perfectly diffuse surface for the calibration process every 10 minutes (ASD, 2015;Viana et al., 2018).

Analysis
The data from the entire phenological cycle (all the DAS) for each crop were used in the statical approach. First, the Spearman correlation coefficient (rs), at 5% of significance, compares the NDVI from both sensors for each crop. After, the NDVI FS with the highest correlation value with NDVI GS was used to provide the inter-calibration process, using Simple Linear Regression. Also, it provided a statical summary of the NDVI values, where the T-student test (5% of significance) was used to analyze that the average of NDVI values was different from each sensor for each crop. To analyze the inter-calibration process, using the NDVI FS values to estimate the NDVI GS, it was used the determination coefficient (R² ), Mean Error (ME), Root Mean Square Error (RMSE). To verify the performance process, it was used the improved concordance coefficient of Willmott (Willmott, Robeson, & Matsuura, 2012) (dr).

Results
All Spearman correlation (rs) among NDVI GS and NDVI FS were significant (at 5% of significance) for soybeans and beans. The values presented no outliers for both crops ( Figure  4), with a rs range of 0.897 to 0.910 for soybean and 0.925 to 0.938 for the bean. NDVI FS, using the bandwidth of 0.648 µm and 0.649 µm on the RED spectrum region and 0.771 µm to 0.773 µm on the NIR spectrum region had the highest rs values with NDVI GS to soybean, with the best rs (0.9105) value to 0.649 µm (RED) and 0.771 µm (NIR). For beans, were identified that the bands from 0.645 µm and 0.646 µm on the RED spectrum and 0.790 µm to 0.795 µm on the NIR spectrum had better NDVI FS correlation with NDVI GS. The best rs Journal of Agricultural Studies ISSN 2166-0379 2021, Vol. 9, No. 2 value (0.9382) was achieved with 0.646 µm to the RED and 0.792 µm to the NIR. Therefore, were used NDVI FS with RED 0.649 µm and NIR 0.771 µm to soybean and RED 0.646 µm and NIR 0.792 µm to beans.

Figure 4. Spearman correlations (rs) between NDVI GS and each NDVI FS wavelength for soybean (a) and bean (b)
NDVI GS and NDVI FS had a similar spectro-temporal pattern for each crop ( Figure 5). However, each crop had different NDVI patterns. Moreover, the NDVI FS has a higher range than NDVI GS for both crops. Analyzing each DAS, soybean had the highest NDVI values for the DAS 71 in both sensors ( Figure 5), and in the DAS 34 and DAS 110 occurred the lower NDVI values from NDVI GS and NDVI FS, respectively. For beans, the higher NDVI values identified on the DAS 55 the lower on the DAS 0 and DAS 10 for both sensors ( Figure  5). These lower NDVI values for beans were expected due to the straw in the soil, without vegetation contribution.
Positive and significant rs values were found between NDVI GS and NDVI FS for the entire soybean (0.9105) and bean (0.9382) crop cycle, as shown in Figure 5. For each DAS ( Figure  5), the NDVI was correlated for the beginning and end of the soybean crop cycle (DAS 34 and DAS 110). Between the DAS 45 and 76 occurred the NDVI saturation, with a small value variation, resulting in a not significant rs. The beans showed more significant rs for DAS than soybean, with a higher value from the DAS 25 and DAS 43. Besides, the not significant NDVI rs for beans occurred on the first (DAS 0) and last (DAS 70) fieldwork, corresponding to the sown and after frost event, respectively. According to the T-student test ISSN 2166-0379 2021 at 5% of significance, NDVI GS and NDVI FS values for the entire phenological cycle were different to the soybean and equal to the bean. Considering the NDVI means for each DAS, soybean had equal NDVI values only to the DAS 34 and DAS 110. To the beans, the last two DAS (DAS 55 and 70) were different for NDVI GS and NDVI FS. Figure 5. Spectro-temporal profiles with standard error bars and spearman correlation for NDVI from both sensors to soybean (a) and bean (b). *5% significance; rs: Spearman correlation

Journal of Agricultural Studies
The NDVI FS adjustment in the NDVI GS function shown R² of 0.8619 to the soybeans and 0.9168 to the beans ( Figure 6). Moreover, the dr values around 0.9 for both crops indicate that the NDVI values registered for both sensors are related. Beans show higher R² e dr and lower ME and RMSE values than soybean ( Figure 6). The NDVI GS adjusted values ( Figure 6) had lower errors and the highest dr values ( Figure 6). This means that the adjusted and observed NDVI GS values were similar. It was found three clusters between the observed and adjusted NDVI GS values (Figure 6), corresponding to the DAS 45, DAS 71, and DAS 76. During these DAS, the soybean canopy fully covers the soil between the rows, minimizing the soil response, become the majority contribute to the NDVI values near 0.90. Besides, had a sudden NDVI variation during the DAS 45 to DAS 76, with a coefficient of variation lower than 2%. ISSN 2166-0379 2021 Figure 6. Scatterplot between the NDVI GS and NDVI FS collected in the field from soybean (a) and bean (b), on the left, and scatterplot between the NDVI values observed and adjusted to soybean (c) and bean (d), on the right. R² : Determination coefficient; ME: Mean Error; RMSE: Root Mean Square Error e dr: improved concordance coefficient of Willmott

Journal of Agricultural Studies
Considering the DAS (Table 1), the concordance between observed and adjusted soybean NDVI GS values was lower than beans. The negative dr values to the DAS 71 and 76 to the soybean (Table 1), representing a discordance between the observed and adjusted NDVI values for some DAS. That could be due to the saturation factor, mentioned before. For beans, the first and last two DAS had lower dr values (Table 1), which could be related to the insignificance of crop presence (e.g., small plant height, sown period, etc., to the DAS 0 and DAS 10) and the frost occurrence (DAS 70). In general, was not observe the NDVI saturation in the bean fields. This because the NDVI values were influenced by lower biomass, less dense canopy, and the soil contribution between rows (Figure 2).

Discussion
Soybeans and beans had similar active and passive NDVI patterns. At the beginning and end of the crop cycle (Figure 5a and Figure 5b) the soil fraction contribution is more evident, resulting in lower NDVI. This soil fraction is more present in the bean field, resulting in small NDVI values than the soybean field. Besides, for beans, monitoring started during the sown period and for soybean, it started on the DAS 34 when the canopy has grown up and attenuated the soil contribution.
In the inter-calibration process, the NDVI values showed more similar to the beans than to the soybeans, in both sensors. Yao et al. (2013) studying winter wheat in China, also found a strong relation between NDVI from active (GreenSeeker) and passive (FieldSpec4) terrestrial sensors. These authors found the equation y = 1.076x -0.362, with R² : 0.896 and RMSE: 0.056, been similar to our results. However, due to the NDVI saturation, the soybean development is not well represented for the higher NDVI values. This NDVI limitation in the dense canopy as soybean is well recognized in several studies (Feng et al., 2016;Liu et al., 2018;Sun, Fang, Liu, & Ye, 2017;Yao et al., 2013;Zheng et al., 2016). As beans had a small canopy than soybean, not covering the space between rows, was not possible to identify this limitation on the bean field. Besides, the number of rainy days (Figure 3) during the soybean vegetative crop cycle was a limitation to have more DAS and improve the spectro-temporal characterization.
This could be due to the data acquisition scheme, where FieldSpec4 was collected statically, against the dynamic way from the GreenSeeker 505 Handheld. NDVI FS is related to statics points, NDVI GS is related to tracks/lanes, being more subject to the field heterogeneity (e.g, soil presence, and different canopy sizes). Thus, as bigger the canopy more significant was the difference between NDVI FS and NDVI GS. The highest differences between NDVI FS and NDVI GS were found from DAS 45 to DAS 76 in the soybean fields, and from DAS 55 to DAS 70 in the bean fields, resulting in a statical different average of NDVI values in these DAS. Moreover, each sensor's Spectral Response Function is different and results in different reflectance registered in the same environmental conditions. Therefore, has different reflectance values for the same canopy and, consequently, different NDVI.
The results have shown that it is possible to monitor the soybean and bean with the NDVI from the different sensors with similar results. GreenSeeker 505 Handheld and FieldSpec4 were used in several studies (Cattani et al., 2017;Prudente et al., 2019;Viana et al., 2018Viana et al., , 2017 to monitor the spectro-temporal development of different crops and estimate biophysical parameters. According to Zheng et al. (2016) and Prudente et al. (2019), the active sensor shows better results to monitor the crop phenological stages and biophysical parameters. One of the factors that could influence is that the passive sensor depends on the solar radiation, and may suffer the calibration process's interference. GreenSeeker is way cheaper than the FieldSepc4 sensor, corresponding an affordable way to monitor the soybean and bean through the NDVI profiles. However, FieldSpec4 has hundred of data in different spectral regions and can provide more crop information.

Conclusion
The NDVI GS and NDVI FS had similar spectro-temporal patterns and were found a strong concordance correlation, around 0.9, between the NDVI from the different sensors from soybean and beans. Hence this indicates the possibility of providing the intercalibrate process, which can be used NDVI from one sensor to estimate the NDVI values to the other sensor. However, this process was less effective after occurs the soybean closed canopy with no visible soil, due to the NDVI saturation process. Therefore, the GreenSeeker 505 Handheld is useful in providing the soybeans and beans monitoring with the NDVI and had a more affordable cost. Future researches with all FieldSpec4 and more crop types are encouraged.