Phosphorous management based on an on-line visible and near infrared (vis-nir) spectroscopy sensor

Transcription

1 Ref: C0378 Phosphorous management based on an on-line visible and near infrared (vis-nir) spectroscopy sensor Abdul. M. Mouazen, Boyan Kuang, and Graham Halcro, Department of Environmental Science and Technology, Cranfield University, Cranfield, MK43 0AL, UK. Abstract Soil available phosphorous (P) is an essential element for crop roots, seeds and canopy development. Conventional soil sampling methods of one sample per ha followed by laboratory analysis are tedious, time consuming, expensive and does not allow exploring spatial variation in P at a desired fine scale. Visible and near infrared (vis-nir) spectroscopy has proven to be a robust, quick and relatively cost effective approach to measure key soil properties at appreciable accuracy. On-line visible and near infrared (vis-nir) spectroscopy measurement of available phosphorous (P) showed large within field variation. This paper aims to utilise P data generated with an on-line vis-nir spectroscopy sensor for site specific management of P 2 O 5 fertiliser. The final aim expected from the variable rate (VR) P application is to ensure uniformity of P distribution across the field, which is hoped to optimise and homogenise crop growth and yield. On-line measurement was carried out for three successive years of 2011, 2012 and 2013 after crop harvest in a 21 ha field in Duck end farm, Bedfordshire, the UK. The P management experiment was for a crop rotation of wheat, oil seed rape and spring barley, respectively. Variable P was only applied in year 2 after crop harvest, where the field was divided into 4 P-index zones of index 0, index 1, index 2 and index 3, according to the RB209 recommendation, provided by the UK Department of Food and Rural Affairs (DEFRA). Indexes 0 and 1 received 140 kg/ha and 70 kg/ha P 2 O 5, respectively, whereas indexes 2 and 3 received no P 2 O 5 fertiliser. The purpose of this VR P application was to attempt unifying the entire field to index 2, which is considered the optimal P level for crops. After the three on-line measurements of soil spectra of the three successive years, a previously developed calibration model of P was used to predict P levels and generate comparison and full-point maps using ArcGIS software. Results showed that the on-line measurement accuracy was acceptable with coefficient of determination (R 2 ), root mean square error of prediction (RMSEP) and residual prediction deviation (RPD) of 0.60, 0.60 mg/100 g and 1.5, respectively. However, accuracy was larger with soil samples scanned under laboratory non-mobile conditions with R 2, RMSEP and RPD of 0.75, 0.51 mg/100 g and 1.8, respectively. The VR application of P 2 O 5 in year 2 after crop harvest led to improving the uniformity of the spatial distribution of P measured in year 3 with the on-line soil sensor. The number of zones of P-index was decreased from 4 indexes before P 2 O 5 VR application to a uniform P index e.g. index 2. The coefficient of variation (CV) of P in the field was reduced from 26% in 2011, and 25% in 2012, to 16% in The online measured P map of year 3 showed significantly more uniform P distribution across the field, comparing to previous years. It was concluded that the on-line vis-nir soil sensor is an effective tool to manage and minimise within field variation in P in arable crops. Keywords: On-line measurement, soil phosphorous, environment, variable rate application. Proceedings International Conference of Agricultural Engineering, Zurich, /8

2 1 Introduction Soil available phosphorous (P) is an essential element for crop roots, seeds and canopy development. Phosphorous deficiency is considered to be one of the major limitations of crop production particularly in low-input agriculture systems around the world (Raghothama, 2005). It is estimated that 5.7 billion hectares of land worldwide is deficient in P for achieving crop production (Batjes, 1997). Although the shortage of P in many parts of the world negatively affect crop growth and yield, excess application of manure has become a significant sources of soil and water pollution in the developed countries, particularly in areas with high rates of run off and soil erodability (Sharpley, 2001). However, agriculture and environmental impacts of P starts at within or subfield scale, where input is applied homogenously by the majority of farmers worldwide. Even farmers adopting precision farming technologies for variable rate (VR) applications of fertilisers do not intend to manage smaller field units than 1 ha, over which one average sample is considered as representative of the underlying variability. Therefore, within field management of P should be targeted at fine resolution, so as management at larger scales could be achieved. In order to fulfil this requirement, proximal soil sensors that quantify and map P spatial distribution are key success for within field management of P and beyond. Recent review report by Kuang et al. (2012) discussed the potential of different technologies used for proximal soil sensing in agriculture. The review revealed that the majority of these technologies can be successfully implemented for mapping the spatial variability, with limited capability in isolating and quantifying sources of the variability. Among different techniques discussed the visible and near infrared (vis-nir) spectroscopy was concluded to be the most successful technique to achieve this goal, particularly for field applications under both mobile and non-mobile measurement conditions. However, the use of this technology should be made with a specific attention to the fact that user should distinguish between directly and indirectly spectrally active properties. This is true, as properties with direct spectral responses in the near infrared (NIR) range are generally measured with higher accuracy than those with indirect spectral responses (Stenberg et al., 2010; Kuang et al., 2012). Properties with direct spectra responses are moisture content, organic carbon, clay, and total nitrogen, whereas all other soil properties are with indirect spectral responses, among which P is a good example. In spite of the fact that P has no direct spectral response in the NIR range, literature shows some successful cases (e.g. Bogrekci and Lee, 2005; Maleki et al., 2006; Mouazen et al., 2009). The conclusion is that when P is successfully measured with vis-nir spectroscopy, this is more likely to be through co-variation with other soil properties that have direct spectral responses e.g. moisture, clay or organic carbon. Experience also demonstrates relationship between soil colour and P. So far, there is no clear explanation of the success cases. The majority of VR P fertilization is based on manual soil sampling of limited number of samples, which is successively followed by laboratory analysis of P and development of recommendation maps. It is obvious that this method is tedious, time consuming, expensive and does not allow exploring spatial variation in P at a desired fine scale so as to allow successful management of P at smaller unit than one ha. Only limited work was reported on the use of proximal soil sensing for VR P application. Among these, probably the most successful example is the sensor-based VR P fertilisation reported by Maleki et al. (2008), which was based on vis-nir real time sensing and control of P. In this study, it was found that the average P 2 O 5 applied on plots was kg/ha, which was 1.25 kg/ha less than the uniform rate fertilisation (30 kg/ha), recommended according to the standard soil test. The overall profit was about 30 per ha, by only applying variable rate P 2 O 5. However, the study was conducted for one cropping season, where no follow up study was conducted to conclude on the fertilisation efficiency from agronomic and environmental point of views. This is particularly true for the evaluation of the resultant spatial homogeneity or heterogeneity of P obtained after the VR P application, which is expected to affect crop growth and yield. The aim of this paper is to utilise spatial data on P generated with an on-line vis-nir spectroscopy sensor for site specific management of P 2 O 5 fertiliser. The final aim expected Proceedings International Conference of Agricultural Engineering, Zurich, /8

3 from VR P application was to ensure uniformity of P distribution cross the field, which is hoped in the long term to optimise and homogenise crop growth and yield. 2 Materials and methods 2.1 Experimental site The study area was a 22 hectare Horn End Field at Duck End Farm, Wilstead, Bedfordshire, U.K. (Latitude; 52d 05m 51s N, Longitude; 0d 27m 19s W), (Fig. 1). The field is normally under an annual three crop rotation system of winter wheat, winter barley and winter oil-seed rape. The soil type was defined as Haplic Luvisols (Soil Survey of England and Wales, NSRI, UK). The textures of selected soil samples indicated the presence of clay, clay loam, sandy clay loam and loam (United States Department of Agriculture (USDA) classification). The topography of the area is rather flat with an elevation that varies between m, determined by differential global positioning system (DGPS) equipment (EZ-Guide 250, Trimble, USA). The study took place over three cropping seasons ( ). In study year 2, the very wet winter caused standing water in the field and the farmer chose to plant spring barley in Horns End field rather than the planned crop of winter wheat. Figure 1: Location of Duck End Farm and study Horns End field. 2.2 On-line sensor and measurement The on-line multi-sensor platform designed and developed by Mouazen et al. (2005) was used in this study. It consists of a subsoiler that penetrates the soil to the required depth, making a trench, whose bottom is smoothed due to the downwards forces acting on the subsoiler. The optical unit was attached to the backside of the subsoiler chisel to acquire soil spectra from the smooth bottom of the trench. The retrofitted subsoiler was attached to a frame, which was mounted onto the three point linkage of a tractor. An AgroSpec mobile, fibre type, vis-nir spectrophotometer (Tec5 Technology for Spectroscopy, Oberursel, Germany) with a measurement range of nm was used to measure soil spectra in diffuse reflectance mode. The data acquisition consisted of a tec5 analogue to digital data converter (tec5 AG, Oberursel, Germany) and AgroSpec (tec5 AG, Oberursel, Germany) data logging software. A semi-rugged laptop (Toughbook, Panasonic UK Ltd., Bracknell, UK) was used to collect data. All hardware including the laptop was enclosed in an IP-65 metal box during measurement so as to protect against dust and rain. The AgroSpec software logged DGPS and spectrophotometer readings at 1 Hz. The spectrometer system, laptop and DGPS were powered by the tractor battery. On-line measurement with the multi-sensor platform was carried out in parallel transects at an average speed of 2 km h -1. A constant gap of 20 m was kept between neighbouring transects. During the one-line measurement soil samples were collected for the evaluation of the measurement accuracy of P. Proceedings International Conference of Agricultural Engineering, Zurich, /8

4 2.3 Laboratory chemical and optical analyses Each sample was divided into two parts; one part was dried for 24 hours at 105 O C and the other part was left fresh (wet). The dried soil sample was analysed to determine P, by spectrometric determination of phosphorus soluble in sodium hydrogen carbonate solution (BS 7755, Section 3.6, 1995). Particle size distribution analysis was also conducted to ascertain soil texture type (BS 7755 Section 5.4., 1998). Each fresh soil sample was dumped into a glass container and mixed well. Big stones and plant residue were excluded (Mouazen et al., 2005). Then each soil sample was placed into three Petri dishes, which were 2 cm in depth and 2 cm in diameter. The soil in the Petri dish was shaken and pressed gently before levelling with a spatula. A smooth soil surface ensures maximum diffuse light reflection and high signal-to-noise ratio (Mouazen et al., 2005). The soil samples were scanned in diffuse reflectance mode using the same mobile, fibre type, vis-nir spectrophotometer (AgroSpec from Tec5 Technology for Spectroscopy, Germany) used during the on-line measurement. A 100 % white reference was used before scanning. A total of 10 scans were collected from each container, and these were averaged in one spectrum. 2.4 Spectra pre-treatment and development of calibration models Pre-treatment of the soil spectra was conducted using Unscrambler 9.8 software (Camo Inc., Oslo, Norway). Spectra at wavelengths of nm were selected for the calibration to eliminate noise at the edges of each spectrum. After the noise was removed, the spectra were reduced by averaging wavelengths nm by 3 and wavelengths nm by 6. Maximum normalisation was followed, which is typically used to get all data to approximately the same scale, or to get a more even distribution of the variances and the average values (Mouazen et al. 2005). After normalisation, the first derivative was calculated, using Savitzky and Golay (1964) polynomial. The polynomial order of 2 was selected to estimate new spectral values with two points on both left and right sides of a certain point. At last the smoothing methods of Savitzky and Golay (1964) were applied. The smoothed values with second-order polynomials were used with three points at the right side and three points at the left side of a certain point. A total of 383 samples were used in this study (Table 1) including samples from Horns End field. They were collected from 13 fields in the UK. Samples were divided into either calibration (70%) or prediction (30%) sets. The calibration spectra with corresponding P values were subjected to a partial least squares regression (PLSR) with leave-one-out crossvalidation. Samples located individually far from the zero line of residual variance were considered to be outliers and were excluded from the analysis. About 5% of samples were excluded during calibration. The resultant calibration model of P was used to predict P values using the laboratory and on-line collected spectra following the same spectra pre-treatment of the calibration. A total of 21 samples collected during the on-line measurement were used as the prediction set (Table 1). Model performance in cross-validation and prediction was evaluated by means of coefficient of determination (R 2 ), root mean square error of prediction (RMSEP) and residual prediction deviation, which is the ratio of standard deviation (SD) to RMSEP. Table 1 Sample statistics for calibration set and prediction set Calibration set prediction set Nr Min Max Mean SD Nr Min Max Mean SD P (mg / ) Sd is standard deviation Proceedings International Conference of Agricultural Engineering, Zurich, /8

5 2.4 Mapping The full-point maps based on all on-line predicted points were developed with ordinary kriging using ArcGIS ArcMap (ESRI ArcGIS TM version 10, CA, USA). Semi-variograms analysis was carried out to produce the full-point maps using Vesper 1.63 software developed by the Australian Centre for Precision Agriculture (Minasny et al., 2005). An exponential model was adopted to calculate semi-variance, since it resulted in the lowest RMSEP. Erorr and comparison maps between on-line predicted and laboratory measured P based on 21 prediction samples were developed with the inverse distance weighing interpolation method. 2.5 Applying of P fertiliser Phosphorous application map was created based on the calculated requirements of 2012 online measurement, using the RB209 guidance leaflet (DEFRA, 2010) (Fig. 2). A homogenous application at 50 kg ha -1 of P 2 O 5 was applied to the entire field. This was followed by a further 30 kg ha -1 on the index 1 areas and 60 kg ha -1 on the index 0 areas. Soil of index 3 required no remediation. P 2 O 5 was applied in 24 m wide treatments by the Kuhn Aero spreader. The existing tramlines were followed to ensure the crop was not damaged by the machinery. 3 Results Figure 2: Phosphorus application schema for Horns End field in 2012 Model performance in cross-validation shown in Table 2 indicates moderate prediction accuracy (e.g. RMESP = 0.55 mg / & RPD = 1.93). This result in similar to that reported by Maleki et al. (2006) for fresh soil samples collected from several fields in Belgium (RMESP = 1.15 mg / & RPD = 2). The model performance for P prediction using laboratory collected spectra is not as good as that of the cross-validation. Although a smaller RMSEP was calculated, larger intercept and smaller RPD were obtained for the laboratory prediction. For the on-line prediction, smaller accuracy as compared to the cross-validation and laboratory prediction was observed. A larger RMSEP and a lower RPD was calculated (Table 2). Particularly the smaller slope and larger intercept values of the 1:1 linear line indicate deterioration of model performance for the on-line prediction. This is expected, as during on-line measurement there are more source of error, as compared to the laboratory based calibrations. Table 2: Summary of model calibration and online validation results Validation R 2 Slope Intercept RMSEP RPD (mg/100g) (mg/100g) Cross-validation prediction Laboratory validation On-line validation R 2 is coeffecient of determination, RMSEP is root mean square error of prediction (RMSEP), RPD is residual prediction deviation = standard deviation / RMSEP. For example, noise, vibration, possible interferences of ambient light, stones and plant roots, variation of soil-to-sensor distance and mismatch of sample position of laboratory analysis Proceedings International Conference of Agricultural Engineering, Zurich, /8

6 sample and corresponding spectrum, all contribute to the overall error of the on-line measurement (Mouazen et al., 2007; Stenberg et al. 2010). However, Mouazen et al. (2009) reported very similar results (R 2 = 0.62, RMSEP = 1.07 mg/100g and RPD = 1.42) to those of the current work (Table 2). The study by Maleki et al. (2006) showed that for a model to be accurate enough, it must achieve values for R 2 of >0.7, and RPD >1.75. From this definition it can be established that the P model in cross-validation and in prediction using laboratory scanned spectra both achieve this target (Table 2). However, the on-line validation of the model developed shows deterioration in model performance (R 2 = 0.42; & RPD = 1.50). This may be attributed to the ambient conditions encountered during the on-line measurement. Comparison maps between measured and on-line predicted P in 2011, shown in Fig. (3) demonstrate spatial similarity, where zones with high and low concentration are clearly located in the field. Both maps indicate low P concentration in the north western corner, whereas high P concentration can be observed in the south eastern part of the field. The error map indicates high error at few points in the field, which might be attributed to chemical analysis error, position mismatch between measured and predicted sample or to the moderate prediction accuracy of the vis-nir spectroscopy model of P having indirect spectral response NIR range (Stenberg et al., 2010; Kuang et al., 2012). Figure 3: Comparison between laboratory chemical (left), on-line (middle) and error (right) maps. Change in the spatial variability shown with the full-point maps between 2011 and 2013 indicates significant differences between years (Fig. 4). In the first measurement in 2011, the field could be divided into two halves, with the north western half having low P concentration as compared to the south eastern part. The latter part has always received the largest amount of manure as it is easily accessed as compared to the other part. Almost similar spatial distribution can be observed in 2012 to 2011, although the area with high concentration in the south eastern part becomes smaller. This can be attributed to the fact that P levels are likely to be at their lowest at harvest (e.g. harvest of 2012), after having been removed from the soil by crop growth (Styles and Coxon, 2007). It is also interesting to note the high concentration of P to appear at the north western corner of the field, which can be attributed to the farmer injecting manure in 2011 after seeing the P map of Figure 4: Full-point maps measured in 2011 (left), 2012 (middel) and 2013 (right). The resultant P map of 2013, measured after Phosphate application in spring 2013 shows a completely different pattern, with more homogeneous spatial distribution, as compared to Proceedings International Conference of Agricultural Engineering, Zurich, /8

7 2011 and 2012 maps. Achieving spatial homogeneity within a field is probably one of the most important goals of precision agriculture. This improved homogeneity is supported by the lower SD and coefficient of variation (CV) calculated for 2013 map, as compared to 2011 and 2012 maps (Table 3). Table 3 Statistical detailes of predicted phosphorous (P) of the full-points maps Year Number of data points Min (mg/100g) Max (mg/100g) Mean (mg/100g) SD (mg/100g) SD is standard deviation; CV is coefficient of variation This is an interesting feature thanks to the on-line sensor used in this study, which proves to be a good tool not only to map soil P, but also to provide a good source of quantitative information to enable management of P site specifically in the field. Although the results of the on-line validation was not encouraging, the improved homogeneity detected with the map of 2013 follows correctly a logic of preceded site specific phosphate application guided by the on-line measurement of P spatial distribution over two cropping seasons. However, there are still areas with low P index that should be corrected with more precise application of phosphate, which are planned for next cropping seasons. 4 Conclusions An on-line visible and near infrared (vis-nir) spectroscopy sensor was used for mapping available phosphorous (P) in a 22 ha field through three cropping seasons. Map of year 2 (e.g. 2012) was used for informing site specific phosphate application in The results obtained in this work allowed the following conclusions to be drawn: On-line measurement and mapping of soil P based on vis-nir spectroscopy is possible, although P has indirect spectral response in the NIR range. Although accuracy of P measurement is not as good as that reported for other soil properties having direct spectral responses, mapping of soil P provided correct information about the spatial distribution of P that match farmer practices preceding the on-line measurement. The use of P data measured with the on-line soil sensor can be useful for guiding site specific application of phosphate that led to improved homogeneity of P spatial distribution remarkably. Further work is needed to study the influence of site specific Phosphate application on crop growth and yield. The study should also assess the environmental impact as well as the economic impacts 5 Acknowledgements Authors acknowledge the financial support received from Home Grown Cereal Authority (HGCA) in the UK through two funded projects (RD , & RD ). 6 References Batjes, N. H. (1997). A world data set for derived soil properties by FAO-UNESCO soil unit for global modeling. Soil Use and Managment, 13, Bogrekci, I., Lee, W. S. (2005). Spectral phosphorus mapping using diffuse reflectance of soils and grass. Biosystems Engineering, 91(3), CV (%) Proceedings International Conference of Agricultural Engineering, Zurich, /8

8 British Standards (1998). Soil quality: BS 7755: Section 5.4: Part 5: Physical methods. Section 5.4: Determination of particle size distribution in mineral soil material - method by sieving and sedimentation. British Standards Institution, UK. Kuang, B., Mahmood, H. S., Quraishi, Z., Hoogmoed, W. B., Mouazen, A. M., van Henten, E. J. (2012). Sensing soil properties in the laboratory, in situ, and on-line: a review. In S. Donald (Ed): Advances in Agronomy, 114, (pp ), AGRON, UK: Academic Press. Maleki, M.R., Van Holm, L., Ramon, H., Merckx, R., De Baerdemaeker, J., Mouazen, A.M. (2006). Phosphorus sensing for fresh soils using visible and near infrared spectroscopy. Biosystems Engineering, 95(3), Maleki, M. R., Mouazen, A. M., De Keterlaere, B., Ramon, H., De Baerdemaeker, J. (2008). On-the-go variable-rate phosphorus fertilisation based on a visible and near infrared soil sensor. Biosystems Engineering, 99(1), Mouazen, A. M., De Baerdemaeker, J., Ramon, H. (2005). Towards development of on-line soil moisture content sensor using a fibre-type NIR spectrophotometer. Soil & Tillage Research, 80(1-2), Mouazen, A. M., Maleki, M. R., De Baerdemaeker, J., Ramon, H. (2007). On-line measurement of some selected soil properties using a VIS-NIR sensor. Soil & Tillage Research, 93(1), Mouazen, A. M., Maleki, M. R., Cockx, L., Van Meirvenne, M., Van Holm, L. H. J., Merckx, R., De Baerdemaeker, J., Ramon, H. (2009). Optimum three-point linkage set up for improving the quality of soil spectra and the accuracy of soil phosphorous measured using an on-line visible and near infrared sensor. Soil & Tillage Research, 103(1), Minasny, B., McBratney, A. B., Whelan, B. M. (2005). Vesper 1.62 Spatial prediction software for precision agriculture nd ed. Australian Centre for Precision Agriculture, McMillan Building A05, The University of Sydney, NSW Raghothama, K. G. (2005). Phosphorous and plant nutrition: an overview. In K. A. Barbarick, C. A. Roberts, & W. A. Dick (Eds), Phosphorous: Agriculture and the Environment, (pp ), Madison, USA: ASA, Inc, CSSA, Inc, & SSSA, Inc. Savitzky, A. & Golay, M. J. E. (1964). Smoothing and differentiation of data by simplified least squares procedures. Analytical Chemistry, 36, Sharpley, A. N., McDowell, R. W., & Kleinman, P. J. A. (2001). Phosphorous loss from land to water: Integrating agriculture and environmental managment. Plant and Soil, 237, Stenberg, B., Viscarra Rossel, R., Mouazen, A.M., & Wetterlind, J. (2010). Visible and near infrared spectroscopy in soil science. In S. Donald (Ed): Advances in Agronomy, 107, (pp ), AGRON, UK: Academic Press. Styles, D., & Coxon, C. (2007) Meteorological and management influences on seasonal variation in phosphorus fractions extracted from soils in western Ireland. Geoderma, Proceedings International Conference of Agricultural Engineering, Zurich, /8