Estimating soil organic carbon of sown biodiverse permanent pastures in Portugal using near infrared spectral data and artificial neural networks
Introduction
Grasslands are one of the most important terrestrial ecosystems for humans due to their contribution to food systems through the production of feed for grazing animals (Naidoo et al., 2008, Sloat et al., 2018). They also play a relevant role in providing ecosystem services (Egoh et al., 2016) and in some cases, they can be biodiversity hotspots (Davidson et al., 2012, Isbell et al., 2011, Soliveres et al., 2016). Nevertheless, inadequately managed grasslands or grasslands converted from natural land use (e.g., tropical forests) can also be responsible for environmental degradation, e.g. loss of climate regulation, biotic production and biodiversity (Abreu et al., 2017, Parton et al., 1993).
An example of enhancing the positive effects of grasslands, developed in Portugal since the 1960 s, is the system of sown biodiverse permanent pastures rich in legumes (SBP). SBP improve grass yield and allow increasing animal stocking rates in semi-natural pastures (Teixeira et al., 2011, Teixeira et al., 2008a, Valada et al., 2012). Their installation requires sowing a mixture of up to 20 species or cultivars of legumes and grasses that provide quality animal feed (Morais et al., 2018a, Teixeira, 2010, Teixeira et al., 2015, 2018). Currently, more than 90% of the total Portuguese SBP areas are located in the agricultural region of the Alentejo, mostly in areas of “montado/dehesa” agro-silvo pastoral ecosystems (Pereira et al., 2009, Teixeira et al., 2015). Nevertheless, those SBP areas represent less than 10% of the permanent grassland area in this region.
High grass yields in SBP lead to a substantial input of fresh soil organic carbon (SOC) (Teixeira et al., 2011), associated to CO2 removal from the atmosphere (Morais et al., 2018b, Teixeira et al., 2019) and soil carbon sequestration. Between 2008 and 2014, the Portuguese Carbon Fund (PCF – funded by the Portuguese Government), through the Terraprima – PCF project, supported the installation and maintenance of SBP using a system of payments proportional to carbon sequestration to assist Portugal in complying with the Kyoto Protocol goals under the Agriculture, Forestry and Other Land Uses activities (APA, 2018, Martins et al., 2015). More than 1,000 farmers established SBP in an area of more than 4% of the country’s agricultural land due to this project (Teixeira et al., 2015). However, it was not possible to monitor the real C sequestration in all these farms since conventional laboratory methods for SOC assessment were too costly and time-consuming. Farmers were paid on the basis of pre-determined sequestration factors established from data obtained in past studies (Teixeira et al., 2011) and not based on the carbon content increase measured on the farm. Indirect monitoring of a combined set of management practices was carried out to ensure that the sequestration factors were applicable. All farmers complying with the set practices received an equal payment. However, C sequestration, like most costs and benefits of grasslands, is highly site-specific and depends on variables such as soil properties, vegetation composition, climate, presence or absence of tree cover and grazing management practices (van Oijen et al., 2018). Therefore, the management of these ecosystems towards maximizing C sequestration requires better knowledge of grassland properties. Future initiatives to promote C sequestration, including subsidies based on the amount of C sequestered, should consider cost-effective methods for SOC quantification to survey large areas more effectively and differentiate performance between farmers.
Among cost-effective methods, visible (400–700 nm), near-infrared reflectance (NIR – 700–1300 nm) and shortwave-infrared reflectance (SWIR – 1300–2500 nm) spectroscopy has been used to indirectly estimate soil properties as an alternative to laboratory methods (Gomez et al., 2008, Srivastava et al., 2015). Various soil characteristics, e.g. SOC, clay content and soil moisture (Kuang et al., 2012, Stenberg et al., 2010), are well correlated with visible-NIR (VNIR), in particular with the NIR wavelengths (Qiao et al., 2017, Stevens et al., 2013).
Several methods can be used to correlate spectral data with SOC content, partial least-squares regression (PLSR) being the most widely used (Gogé et al., 2012, Gomez et al., 2008, Mouazen et al., 2010, Serrano et al., 2021, Shi et al., 2014). Recently, a wide range of alternative machine learning methods have been increasingly applied (Baligh et al., 2020, Fernandes et al., 2019, Morellos et al., 2016, Mouazen et al., 2010), artificial neural networks (ANN) being one of the most frequently used (Kuang et al., 2015). Advantages of ANN are that they can (i) estimate complex, high-dimensional and nonlinear relationships (Liu et al., 2016); and (ii) use both quantitative and qualitative data as inputs (Yang et al., 2018). Nevertheless, the use of these models for land use systems monitoring is still underexplored (Brown et al., 2006). ANN are increasingly used to estimate SOC content. Using Scopus® to search in the titles and abstract of published studies and with the following search string: “artificial neural network” AND (“soil organic matter” OR “soil organic carbon”)), we found 115 hits. In these, more than 75% were published since 2016.
The present study aimed to assist with expedited and efficient monitoring of SBP through the combination of spectral data with machine learning methods, namely ANN. For this, an ANN algorithm was calibrated for the estimation of SOC (at 20 cm depth) in soil samples from SBP soils using 800–2800 nm (laboratory) spectral data as an input. This work is intended as a first step in automating SOC estimation using progressively more remote sources of data. Testing the decreased spectral resolution through the use of Sentinel-2 (S2) aggregation enables us to prepare for future use of remotely sensed spectral data, as we explored how accurately data collected in those bands can estimate SOC for SBP. Here, we provide an ANN-based algorithm for estimating SOC spectroscopically and paves the way for future development of algorithms based on S2 that enable remote estimation of SOC. The calibration used conventional (laboratory) SOC analyses from the same soil samples. We tested two data reduction methods: (i) principal component analysis (PCA), and (ii) aggregation of spectral data based on S2 bandwidths. In both instances, we used an 8-fold leave-one-location-and-year-out approach to estimate the generalisation capacity of the model.
Section snippets
Study area and soil sampling design
This work used soil samples collected in five farms in south and central Portugal (farms 1, 2, 4 and 5 in the Alentejo and farm 3 in Beira Interior), across latitudes and longitudes ranging respectively between 38°10′ and 40°30′N and 7°40′ and 8°30′ W (Fig. 1). Farms 1 and 5 are located near Évora, farm 2 near Cabeço de Vide, farm 3 near Covilhã, and farm 4 near Grândola. All farms are in the hot-summer Mediterranean climate region, according to the Köppen climate classification system (IPMA,
Laboratory analysis of soil organic carbon
The main trend between folds that can be observed in the values of SOC determined with analytical techniques (combination of farm and year) is that, in general, SOC variation within the same fold tends to be higher than that between folds (Fig. 3). For each of the two years, the differences in SOC between farms are small. For example, Farm 3–2019, Farm 4–2019 and Farm 5–2019 have similar SOC. In Farm 3–2019 the median is 7.48 g kg−1 and interquartile range is 3.16 g kg−1, in Farm 4–2019 the
Discussion
In this paper, we have combined spectral images with ANN for measuring SOC in SBP systems in Portugal. We have shown that two different spectral dimensionality reduction approaches performed as well as the raw spectra (Table 4). PCA is a dimensionality reduction based on the statistical correlation between different wavelengths that uses the actual measurements, and we therefore expected it to perform better than S2. Further, the considered first five PC explains more than 99% of the variance
Conclusion
We provide an algorithm for estimating SOC in Portuguese SBP using spectroscopy. We concluded that spectroscopical methods are a viable solution for expediting SOC monitoring. The method applied laboratory analysis using high-resolution NIR spectral data. Moreover, it provided evidence that those spectroscopical methods can be replaced by different data products or data collection methods with lower spectral resolution. A comparison of ANN modelling results between two reduced spectra methods
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgments
This work was supported by Programa de Desenvolvimento Rural (PDR2020) through “Viabilização de pastagens semeadas biodiversas através da otimização da fertilização fosfatada” (PDR2020-101-030690) and “GO SOLO: Avaliação da dinâmica da matéria orgânica em solos de pastagens semeadas biodiversas através do desenvolvimento de um método de monitorização expedito e a baixo custo” (PDR2020-101-031243). The work was also supported by Fundação para a Ciência e Tecnologia through projects “Animal
References (72)
- et al.
Global soil characterization with VNIR diffuse reflectance spectroscopy
Geoderma
(2006) - et al.
Estimation of soil organic matter content by modeling with artificial neural networks
Geoderma
(2019) - et al.
Optimization criteria in sample selection step of local regression for quantitative analysis of large soil NIRS database
Chemom. Intell. Lab. Syst.
(2012) - et al.
Soil organic carbon prediction by hyperspectral remote sensing and field vis-NIR spectroscopy: An Australian case study
Geoderma
(2008) - et al.
Sensing soil properties in the laboratory, in situ, and on-line
A review, in: Advances in Agronomy.
(2012) - et al.
Comparison between artificial neural network and partial least squares for on-line visible and near infrared spectroscopy measurement of soil organic carbon, pH and clay content
Soil Tillage Res.
(2015) - et al.
Digital soil mapping algorithms and covariates for soil organic carbon mapping and their implications: A review
Geoderma
(2019) - et al.
Estimating the crop leaf area index using hyperspectral remote sensing
J. Integr. Agric.
(2016) - et al.
The use of machine learning methods to estimate aboveground biomass of grasslands: a review
Ecol. Indic.
(2021) - et al.
Machine learning based prediction of soil total nitrogen, organic carbon and moisture content by using VIS-NIR spectroscopy
Biosyst. Eng.
(2016)