With the needs of efficient acquisition of soil information, Digital Soil Mapping (DSM) has been greatly developed and widely applied for over the past two decades. The spatial estimates of soil properties produced with diverse methods over various study areas, have been often seen as the main output of DSM, as they play an important role in environmental modelling and policy. However, compared with the soil property maps, their prediction uncertainty is still less emphasized, which may potentially lead to mis-uses of results and inappropriate decisions if the uncertainty is not assessed, reported, and taken into account by end-users.
In this communication, we present a preliminary review of the sources of prediction uncertainties in DSM coming from learning soil data (data source, sampling in space and time, measurements), covariates, and models. We also summarize the methods used to estimate the uncertainty, and to assess the reliability of the uncertainty estimates. We also consider the propagation of uncertainties when several soil attributes are combined to derive information and/or used as inputs for modelling. Furthermore, we discuss some strategies for mitigating the uncertainty, challenges, and future perspectives. This review aims to consolidate the understanding of DSM uncertainties and to contribute to reliable DSM practices, facilitating more informed decision-making in soil-related research and management. 

How to cite: Chen, Q., Richer-de-Forges, A., Chen, S., Vaudour, E., Bispo, A., and Arrouays, D.: Uncertainty in Digital Soil Mapping at broad-scale: A review, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-6005, https://doi.org/10.5194/egusphere-egu24-6005, 2024.

Artificial neural networks (ANNs) have proven to be a useful tool for complex questions that involve large amounts of data, for example, predicting soil classes on various scales. Our use case of predicting soil maps with ANNs is in high demand by government agencies, construction companies, or farmers, given cost and time intensive field work.
However, there are two main challenges when applying ANNs. In their most common form, deep learning algorithms do not provide interpretable predictive uncertainty. This means that properties of an ANN such as the certainty and plausibility of the predicted variables, rely on the interpretation by experts rather than being quantified by evaluation metrics validating the ANNs. This leads to the second challenge: these algorithms have shown a high confidence in their predictions in areas geographically distant from the training area or areas only sparsely covered by training data.

To tackle these challenges, we use the Bayesian deep learning approach “last-layer Laplace approximation”, which is specifically designed to quantify uncertainty into deep networks, in our explorative study on soil classification. It corrects the overconfident areas without reducing the accuracy of the predictions, giving us a more realistic uncertainty expression of the model's prediction.  In our study area in southern Germany we divide the soils into typical soils of valleys, the Swabian Jura and the Black Forest. As a test case, we then explicitly exclude the soil types of Swabian Jura and Black Forest in the training area but include these regions in the prediction. These two regions are characterized by very different soil types compared to the rest of the study area due to their considerably different geology, climate, and terrain.

Our findings emphasize the need to address the issue of overconfidence in ANNs, particularly for distant regions from the training area. Moreover, the insights gained from this research are not only limited to addressing overconfidence in ANNs, but also offer valuable information on the predictability of soil types and identifying knowledge gaps. By analysing regions where the model has limited data support and, consequently, high uncertainty, stakeholders can recognize the areas that require more data collection efforts.

How to cite: Rau, K., Eggensperger, K., Schneider, F., Hennig, P., and Scholten, T.: How can we quantify, explain, and apply the uncertainty of complex soil maps predicted with neural networks?, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-11463, https://doi.org/10.5194/egusphere-egu24-11463, 2024.

In large parts of Switzerland, there are up to the present day no soil maps in a more precise scale than 1:25’000. In 2020 the Swiss government adopted a soil strategy to promote soil mapping and the required tools and guidelines. Furthermore, authorities are legally obligated to delineate areas for the high-quality arable land inventory. In the future, soil maps will be available at high resolution for increasingly large areas. However, farmers are not used to work with this type of soil information as it was so far not available and data products are unknown.

In a pilot study of 1’000 ha in the canton of Bern a large number of soil properties have been mapped at high resolution using digital soil mapping techniques based on 1’500 newly surveyed observations. The content of the maps ranges from continuous or classified basic soil properties such as texture or soil organic matter content to aggregated soil properties such as water storage capacity. From these baseline maps certain applications and examples were assessed for specific agricultural use and the corresponding maps were discussed with local farmers. The examples include options for precision farming, proposals for new sampling sites for the legally required soil survey, the generation of more meaningful routine measurements and planning bases for irrigation systems at regional scale. Feedback was diverse and interest in soil maps differs largely according to farm management, specialization and digital affinity. Our study shows the importance of stakeholder involvement and training as well as familiarization of the farmers with digital soil maps.

How to cite: Kellermann, L., Tanner, S., Oechslin, S., Nussbaum, M., and Burgos, S.: Useful digital soil mapping products for farmers , EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-17128, https://doi.org/10.5194/egusphere-egu24-17128, 2024.

One of the main challenges in digital soil mapping is the imbalanced datasets for soils classification. For these datasets, machine learning techniques use to overestimate the majority classes and underestimate the minority ones. In general, this generates maps with poor precision and unrealistic results. Considering these maps for land use decision-making can have dire consequences. This is the case of acid sulfate (AS) soils, a type of harmful soil that can generate serious environmental damage when drained in agricultural or forestry activities. In the study area, the probability of finding AS soils is very high. Furthermore, some of the most hazardous AS soils in Finland are located there [1]. Therefore, it is necessary to create high-precision maps to avoid environmental damage. Since the dataset for this region is highly imbalanced, the first step in creating accurate maps is to balance the dataset. Although most  soil class datasets in nature are imbalanced, this problem has been hardly studied. In this work, we analyze different techniques to address the problem of imbalanced datasets. The methods considered to balance the dataset are under- and oversampling techniques and the combination of both. For the oversampling of the minority class, we create a kind of artificial samples from the quaternary geological map. The method used for the modeling is Random Forest, one of the best methods for the classification of AS soils [2,3]. Balancing the dataset improves the performance of the model in all the studied cases, where the values of the metrics for both classes are above 80%. Furthermore, we create AS soil probability maps for the four balanced datasets and the imbalanced dataset. A detailed comparison between the maps is made. In addition, the extent of the AS soils obtained in all the cases is compared with the extent of the AS soils in the conventionally produced occurrence map [1]. The modeled probability maps created from the balanced datasets have a high precision. The results of this study confirm the importance of balancing the dataset to improve the prediction and classification of AS soils.

[1] Geological Survey of Finland. Acid Sulfate Soils–map services http://gtkdata.gtk.fi/hasu/index.html 

[2] V. Estévez et al. 2022.  “Machine learning techniques for acid sulfate soil mapping in southeastern Finland”. Geoderma 406, 115446.

[3] V. Estévez et al. 2023. “Improving prediction accuracy for acid sulfate soil mapping by means of variable selection”. Front. Environ. Sci. 11:1213069.  doi: 10.3389/fenvs.2023.1213069

How to cite: Estévez, V., Mattbäck, S., Boman, A., Liwata-Kenttälä, P., Björk, K.-M., and Österholm, P.: Acid sulfate soil mapping in western Finland: How to work with imbalanced datasets and machine learning, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-1375, https://doi.org/10.5194/egusphere-egu24-1375, 2024.

Being a fundamental indicator of soil health and quality, soil bulk density (BD) plays an important role in plant growth, nutrient availability, and water retention. Due to its limited availability of BD in databases, pedotransfer functions (PTFs) has been widely used in predicting BD, while the impact of PTFs’ accuracy on soil organic carbon (SOC) stock calculation has not been explored. Herein, we proposed a local modeling approach for predicting BD across EU and UK using LUCAS Soil 2018. Our approach involved a combination of neighbor sample search, Forward Recursive Feature Selection (FRFS) and Random Forest (RF) model (local-RF_FRFS). The results showed that local-RF_FRFS had a good performance in predicting BD (R² of 0.58, RMSE of 0.19 g cm^-3), surpassing the traditional PTFs (R² of 0.40-0.45, RMSE of 0.22 g cm^-3) and global PTFs using RF with and without FRFS (R² of 0.56-0.57, RMSE of 0.19 g cm^-3). Interestingly, we found the best traditional PTF (R²=0.84, RMSE=1.39 kg m^-2) performed close to the local-RF_FRFS (R²=0.85, RMSE=1.32 kg m^-2) in SOC stock calculation using BD predictions. However, the local-RF_FRFS still performed better (ΔR²>0.2 and ΔRMSE>0.1 g cm^-3) for soil samples with low SOC stock (<3 kg m^-2). Therefore, we suggest that the local-RF_FRFS is a promising method for BD prediction while traditional PTFs would be more efficient when BD is subsequently utilized for calculating SOC stock.

How to cite: Chen, S., Chen, Z., Zhang, X., Luo, Z., Schillaci, C., Arrouays, D., Richer-de-Forges, A., and Shi, Z.: Machine learning based pedotransfer function improves soil bulk density prediction but not for soil organic carbon stock, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-1755, https://doi.org/10.5194/egusphere-egu24-1755, 2024.

Maps of soil pH are an important tool for making decisions in sustainable forest management. Accurate pH mapping, therefore, is crucial to support decisions by authorities or forest companies. Soil pH values typically exhibit a distinct distribution characterized by two frequently encountered pH ranges, wherein aluminium oxides (Al₂O₃) and carbonates (CaCO₃) act as the primary buffer agents. Soil samples with moderately acid pH values (pH CaCl₂ of 4.5-6) are less commonly observed due to their weaker buffering capacity. The different strength of buffer agents results in a distinct bimodal distribution of soil pH values with peaks at pH of around 4 and 7.5. Commonly used approaches for spatial mapping neglect this often observed characteristic of soil pH and predict unimodal distributions with too many moderately acid pH values. For ecological map applications this might result in misleading interpretations.

This study presents a novel approach to produce pH maps that are able to reproduce pedogenic processes. The procedure is suitable for bimodal responses where the response distribution is naturally inherent and needs to be reproduced for the predictions. It is model-agnostic, namely independent from the used statistical prediction method. Calibration data is optimally split into two parts corresponding each to a data culmination, i.e. for soil pH values belonging to the ranges of the two principal buffer agents (Al₂O₃ and CaCO₃). For each subset a separate model is then built. In addition, a binary model is fitted to assign every new prediction location a probability to belong either to Al₂O₃ or CaCO₃ buffer range. Predictions are combined by weighted mean. Weights are derived from probabilities predicted by the binary model. Degree of smoothness is chosen by sigmoid transform which allows for optimal continuous transition of the pH values between Al₂O₃ and CaCO₃ buffer ranges. For each location uncertainty distributions may be combined by using the same weights.

We illustrated application of the new approach to a medium and strong bimodal distributed response (1) pH in 0–5 cm and (2) pH in 60–100 cm of forest soils in Switzerland (2 530 calibration sites). While model performance measured at 354 validation sites slightly dropped compared to a common modelling approach (drop of R² of 0.02–0.03) distributional properties of the predictions are much more meaningful from a pedogenic point of view. We were able to demonstrate the benefits of considering specific distributional properties of responses within the prediction process and expanded model assessment by comparing observed and predicted distributions.

How to cite: Nussbaum, M., Zimmermann, S., Walthert, L., and Baltensweiler, A.: Hierarchical modelling in digital soil mapping: Advantages for mapping bimodal soil pH, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-6172, https://doi.org/10.5194/egusphere-egu24-6172, 2024.

Digital Soil Assessment (DSA) methods are increasingly being applied as integrative step in which DSM outputs are incorporated into crop and land management decision support systems, often through the vehicle of process models. The value of the acquisition, interpretation of any soil data, and the additional effort in developing and producing a DSM product resides in its use and utility. The purpose of DSA is the development of the value argument for the conversion of quantitative data on soil properties obtained through DSM to a spatial assessment of the capacity of a soil to fulfil a particular function. Here we present the use of DSM/DSA to i) demonstrate the discovery new soil process knowledge on rice production in Bangladesh and the conditions which can determine heavy metal accumulation; ii) determine and illustrate DSA under data sparce conditions through mapping areas in West Africa under which Fe toxicity in the soil limits agricultural production and iii) integration of DSA in a wider decision support system to support decision making in a semi-arid region in Morocco, illustrating the trade-offs between agricultural production and ecosystem services. Through these case examples, we demonstrate the flexibility of DSA approaches in addressing both the generation of new knowledge around soil processes, in justifying the acquisition of new soil data and in helping landowners to manage their crops and soils sustainably.

How to cite: Corstanje, R., Oulaid, B., Zawadzka, J., Ingram, B., Milne, A., Kirk, G., and Hannam, J.: Integrated Digital Soil Assessment across South East Asia and Africa., EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-12959, https://doi.org/10.5194/egusphere-egu24-12959, 2024.

Digital soil mapping relies on statistical relationships between soil profile observations and environmental covariates at the sample locations. However, inherent limitations of legacy soil profiles, such as inaccurate georeferencing and inconsistent sampling techniques, frequently introduce location errors into these soil profiles that greatly affect the quality of digital soil mapping. To address this challenge, this study focuses on reducing the location error of legacy soil profiles and evaluating the resulting impact on digital soil mapping. We enhanced the consistency between environmental covariates (i.e., elevation, slope and land use) with relative high accuracy and detailed descriptive information of legacy soil profiles to reduce the location error of legacy soil profiles. We constructed quantile regression forest models to predict soil properties and their uncertainty at different depths using soil profiles before and after location error correction. Our results demonstrate that for the majority of soil variables, correcting positional errors in legacy soil profiles significantly enhances the accuracy of the digital soil mapping. The largest improvement was found for soil organic carbon at 5 cm depth, with 21% increase of   R^2. The impact of reduced location error is particularly noteworthy in regions characterized by complex terrain or sparse sampling. In addition, the accuracy and details of the predicted maps are significantly improved, which better represent the spatial variation of soil attributes across China. Besides, we also found that elevation was the primary controlling factor for correcting location error of legacy soil profiles, followed by land use and slope. This research presents a significant step towards producing high-resolution and high-quality spatial soil datasets, which can provide essential support for soil management and ensure future soil security.

How to cite: Shangguan, W. and Shi, G.: Reducing location error of legacy soil profiles leads to significant improvement in digital soil mapping, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-2255, https://doi.org/10.5194/egusphere-egu24-2255, 2024.

The adaption of our land use and agricultural practices requires more detailed and more reliable spatial soil physical data. The LUCAS topsoil database is an up to date collection of soil physical data, however it is spatially scarce. The soil physical data of the Hungarian Soil Information and Monitoring System (SIMS) is spatially denser and has data from multiple layers from 1992. Harmonizing and combining the two datasets can lead to the creation of better resolution and more accurate maps. Before combining the databases, we must make sure, that the sample points represent the area in the same way. The comparison of the data began with the cleansing of the datasets, followed by the conversion of the many sampling depths of the SIMS data to 0-20 cm using mass preserving splines and the conversion of the particle size limit from FAO/WRB to USDA standard. To make sure, that the sum of the sand, silt and clay fractions was 100% additive log ratio (ALR) transformation was applied on both LUCAS and SIMS. Mapping was carried out using random forest kriging with 10-fold cross-validation on a 100 m * 100 m grid using 28 environmental covariates. The ALR maps were converted back, resulting in the sand, silt and clay maps. Using the three maps, soil texture classes were calculated for both datasets using the USDA soil texture triangle. The soil texture classes were compared to each other pixel-by-pixel using the taxonomical distances of the texture classes. The particle fraction maps were compared to each other also pixel-by-pixel using linear regression. The results let us conclude that the LUCAS and SIMS databases produce very similar maps of both sand, silt a clay. The soil texture class comparison also resulted in a very close match with the majority of the country producing very close or perfect matches.  The two soil monitoring systems produce very similar results when mapping sand, silt, clay and soil texture for the whole country and can safely be combined together for future use and mapping.

How to cite: Benő, A., Szatmári, G., Laborczi, A., Kocsis, M., Bakacsi, Z., and Pásztor, L.: Comparison of predicted soil physical property maps based on (i) LUCAS topsoil database and (ii) Hungarian Soil Information and Monitoring System, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-18218, https://doi.org/10.5194/egusphere-egu24-18218, 2024.

Soil organic carbon (SOC) is a critical factor influencing global carbon cycling. Accurate estimates of its spatial distribution are essential for addressing global climate change. Digital soil mapping has demonstrated significant potential in providing precise and high-resolution spatial information about SOC across various scales. We conducted an evaluation of two soil mapping approaches for SOCD estimates in France: the direct approach (calculate-then-model) and the indirect approach (model-then-calculate). Our study utilized 916 topsoil samples (0−20 cm) from the LUCAS Soil 2018 dataset and 24 environmental covariates. We employed a random forest model and forward recursive feature selection to build spatial predictive models of SOCD using both the direct and indirect approaches. The results revealed that, with the random forest model and full covariates, both approaches demonstrated moderate performance (R² = 0.28−0.32). Through the use of forward recursive feature selection, the number of predictors was reduced from 24 to 9, leading to enhanced model performance for the direct approach (R² of 0.35), while no improvement was observed for the indirect approach (R² of 0.28). The mean SOCD of French topsoil was estimated at 5.29 and 6.14 kg m⁻² using the direct and indirect approaches, respectively, resulting in SOC stocks of 2.8 and 3.3 Pg, respectively. Notably, the indirect approach exhibited better performance in estimating high SOCD. These findings serve as a valuable reference for SOCD mapping.

How to cite: Chen, Z., Shuai, Q., Shi, Z., Arrouays, D., Richer-de-Forges, A., and Chen, S.: Comparison of direct and indirect approaches for mapping soil organic carbon stock, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-1764, https://doi.org/10.5194/egusphere-egu24-1764, 2024.

Detailed soil spatial information is a worldwide need for monitoring soil quality, especially in agricultural region. Remote sensing technology has evolved as a powerful tool for characterizing spectral reflectance of bare soil, which is the perquisite for retrieving soil information. However, existing methodologies were mostly designed to extract bare soil information from single satellite image, which is prone to cloud contamination and phenological variation. Although some composite algorithms based upon multitemporal images were proposed for soil mapping, they were all designed for coarse-resolution satellite dataset; besides, their generalization ability over a large scale (e.g., national) remains poorly explored. To fill the knowledge gap, we proposed a new framework, namely Two-Dimensional Bare Soil Separation (TDBSS), for extracting continuous bare soil information at 10-m spatial resolution based on multi-temporal Sentinel-2 images for cropland across China. The TDBSS used Soil Adjusted Vegetation Index and Green-Red Vegetation Index as two-dimensional indicators. The optimal thresholds for these two indicators were further obtained across two dimensions based upon ecoregion-specific samples. These thresholds were further applied for nine primary agricultural zones in China and subsequently adapted for the entire country. We also compared the framework with three widely used bare soil detecting algorithms (i.e., Geospatial Soil Sensing System (GEOS3), soil composite processor (SCMaP), and Barest Pixel Composite (BPC)) using the spatial accuracy. The TDBSS performed the best with an overall accuracy (OA = 78.28%), while SCMaP showed the lowest OA of 29.25%. The results showed the TDBSS was an effective method for a large-area mapping of bare soil. The resultant bare soil composite map holds great significance for further retrieving soil properties for Chinese cropland. TDBSS is computationally efficient and readily applied for a broad spatial scale, which is practically crucial to the food security, land management, and precision agriculture policymaking.

How to cite: Xue, J., Zhang, X., Chen, S., Su, Y., and Shi, Z.: A new framework for retrieving bare soil information using multi-temporal Sentinel-2 images across China, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-4622, https://doi.org/10.5194/egusphere-egu24-4622, 2024.

Evaluating the effectiveness of remote sensing-based vegetation indices in estimating the spatio-temporal distribution of Nitrogen rates

One of the main objectives of precision agriculture is to optimize nitrogen fertilizer application management. This study aimed to assess the efficacy of different vegetation indices derived from satellite data in estimating soil ammonium and nitrate values during the corn growing season. To achieve this, multi-temporal Sentinel-2 images and ground data including ammonium and soil nitrate values measured at specific ground stations throughout the corn growing season for Hunter field, Canada, were utilized. Firstly, various vegetation indices including NDVI, EVI, MSAVI, ARVI, GNDVI, and OSAVI were calculated for different dates throughout the crop growing season. Subsequently, the Pearson correlation between these vegetation indices and temporal variations in soil ammonium and nitrate values during the growing season was examined. Moreover, the relationship between vegetation indices at the crop growth peak and the amount of fertilizer applied to the soil during planting was investigated. The findings indicated that the average correlation coefficients between total soil nitrate and ammonium values throughout the growing season and the NDVI, EVI, MSAVI, ARVI, GNDVI, and OSAVI indices were -0.67, -0.72, -0.69, -0.68, -0.73, and -0.70, respectively. Furthermore, the average correlation coefficients between these indices at the growth peak and the cumulative ammonium and nitrate applied at planting were 0.60, 0.60, 0.64, 0.60, 0.68, and 0.64, respectively. The correlation coefficient and root mean square error (RMSE) between the measured and modeled sum of ammonium and nitrate, based on the six vegetation indices in a multivariate form, were 0.89 and 17.3 mg, respectively.

How to cite: Biswas, A., Fathololoumi, S., and Sulik, J.: Evaluating the effectiveness of remote sensing-based vegetation indices in estimating the spatio-temporal distribution of Nitrogen rates, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-4671, https://doi.org/10.5194/egusphere-egu24-4671, 2024.

The knowledge about soil hydraulic properties (i.e., water retention and hydraulic conductivity characteristics) is of relevance for accurate determination of land surface fluxes. Yet, the laborious and time-consuming process of measuring the soil water retention curve (SWRC) with conventional laboratory methods poses a challenge. In addition, the measured soil water content and matric potential pairs obtained with standard methods are often fragmentary and consist of only a limited number of measurements across the desired soil water content range. Proximal and remote sensing methods are rapid and cost-efficient alternatives to quantify soil attributes across different scales. However, past studies that centered around proximal and remote sensing of soil hydraulic functions mainly rely on statistical relationships and a physically-based method is still lacking. In this presentation, we introduce an innovative physics-based laboratory method that allows the direct estimation of the complete SWRC across the entire range from saturated to dry conditions. The inputs to the model include measured data pairs of soil water content and reflectance within the shortwave infrared domain. The fundamental hypothesis behind the new method is that the soil reflectance spectra are a function of both soil water content and the pore scale distribution of capillary and adsorbed soil water. The performance of the proposed model was evaluated for 21 soils that vastly differ in physical and hydraulic properties. The RMSE and R-squared between retrieved and measured water contents at various matric potentials were found to be 0.03 m³ m⁻³ and 0.96, respectively, indicating the good performance of the proposed method. The results suggest that the new method is a rapid and efficient alternative to established laboratory measurement methods.

How to cite: Norouzi, S., Sadeghi, M., Tuller, M., Ebrahimian, H., Liaghat, A., Jones, S. B., and de Jonge, L. W.: A Physics-Driven Spectroscopic Approach for Rapid Estimation of the Soil Water Retention Curve, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-3053, https://doi.org/10.5194/egusphere-egu24-3053, 2024.

The current agricultural system allows farm machinery to operate randomly, thus compacting around 23-33% of the land with critical levels in Europe. To tackle this issue, the European Union (EU) has launched the mission ‘A Soil Deal for Europe’ under the Horizon Europe program, this mission aims to have healthy soil by 2030. Mitigating soil compaction to improve soil structure has been selected as one of the eight objectives of this mission. It is evident from past research that with the increasing size and weight of farm machinery, soil compaction has become a significant threat to top and sub-soil; however, subsoil compaction is even more persistent and cumulative than topsoil. Therefore, within the SOLGRAS project, we emphasize both surface and sub-surface soil compaction. The ability of soil to withstand the soil compaction is governed by its soil strength. This soil strength depends on many soil properties, such as soil water content, bulk density, texture, and organic matter content. We aim to map these soil properties as a first step before predicting soil strength at the field level. Since proximal sensors provide rapid, low-cost, non-destructive measurements, they have significantly enhanced digital soil mapping. In this project, we have used geophysical sensors based on electromagnetic induction and gamma-ray radiometric principles to predict the above-mentioned soil properties at the field level in Denmark. The geophysical survey and collection of soil samples were performed on the same day for each of the three chosen farmer fields. Here, we present our results of predicted soil properties obtained via the proximal sensors and a limited number of laboratory-measured values. Samples of 100 cc undisturbed soil cores and bags were collected from the surface (15-cm depth) and sub-surface (40-cm depth) at 23 sampling sites for each field. Data from these sampling sites are used to train and validate our models for predicting soil properties associated with soil strength. These predicted high-resolution maps produced at the field level will enable us to set up optimum vehicular specifications and routes before entering the field.

How to cite: Khatkar, A., Beucher, A., Koganti, T., Munkholm, L. J., and Lamandé, M.: Mapping Soil Strength Markers at Field Scale Using Proximal Sensors, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-1659, https://doi.org/10.5194/egusphere-egu24-1659, 2024.

Remote sensing of saline soils has been an active area of research in the past few decades. This is particularly so as soil salinity is a major environmental geo-hazard in both agricultural lands and arid and semi-arid regions. Saline soil adversely affect soil and play a major role in soil erosion, dispersion and degradation. Furthermore, saline soils in arid and semi-arid regions lead, in certain situations, to land subsidence, and ground upheaval. In agricultural lands saline soils lead to reduced agricultural productivity, interference with plant nutrition and soil erosion.

Mapping saline soils is carried out using various techniques and procedures ranging from direct field observations and sampling to space based remote sensing techniques. Traditional methods of measuring soil salinity are time-consuming and labor intensive, making remote sensing techniques an attractive alternative. Remote sensing provides a less costly procedure due to the large global spatial coverage, continuous repetitive coverage and high-quality earth observations. Most of the remote sensing of saline soils previous research have focused on the broad band remote sensing part with primary focus on the spectral ranges in the near infra-red and the short-wave infra-red. However, higher levels of detection accuracy has not been widely achieved.

In comparison to broad band remote sensing data, hyperspectral remotely sensed data provides an alternative approach that is much more accurate in detection levels of saline soils and their spatial distribution. This research employed a hand-held hyperspectral sensor specifically the SVC-XHR-1024i to collect reflectance data over various samples of soils collected in western United Arab Emirates. The collected data were then processed to derive spectral indices that are sensitive to soil salinity. Laboratory measurements of electrical conductivity (EC) of soil water extracts were carried out for the corresponding soil samples. The study established a statistical relationship between measured soil hyperspectral reflectance and EC salinity values. The study findings indicate that the spectral ranges in the shortwave infrared (SWIR) and near-infrared (NIR) are crucial and optimal in detecting soil salinity, in comparison to any other spectral ranges. An accuracy of 71% in detecting saline soils, including salt flats, was achieved through the use of narrow band hyperspectral data at the SWIR and NIR ranges. This by far exceeds accuracy levels that were previously achieved using broad band remote sensing data. The study also, highlights the potential of hyperspectral remote sensing as a cost-effective and efficient tool for monitoring soil salinity and identifying areas at risk of salinization, which can inform land management strategies for sustainable agriculture and future land development.

How to cite: Abuelgasim, A. and Aldhaheri, A.: Hyperspectral Remote Sensing of Saline Soils in Arid and Semi-Arid Environments, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-4248, https://doi.org/10.5194/egusphere-egu24-4248, 2024.

An increasing number of hyperspectral satellite platforms are becoming available as exemplified by the DESIS platform.  This generates an immense amount of spectral information about the earths surface which is unlabelled.  In this work we use DESIS imagery to compare two deep learning approaches for utilising all of this unlabelled data for predicting topsoil properties. The first is transfer learning from laboratory based spectral libraries.  The second is a novel self-supervise learning approach which employs a transfer-based autoencoder architecture for unsupervised learning, analyzing spectra patterns and cultivating powerful latent representations for the downstream task of soil property analysis. Using a dataset from eastern Australia we show that the self-supervised learning approach gives superior predictions than transfer learning.

How to cite: Bishop, T., Hu, K., Filippi, P., and Wang, Z.:  A Deep Learning Approach for Improving Soil Property Prediction with Unannotated Hyperspectral DESIS Imagery, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-11994, https://doi.org/10.5194/egusphere-egu24-11994, 2024.

Soil moisture, despite its crucial role in various agricultural processes, acts as noise in retrieving properties such as texture and soil organic carbon through spaceborne hyperspectral data. High spatiotemporal variability in moisture reduces the capability of soil monitoring. Soil moisture determines a reduction of the reflectance over the entire spectrum, which is not linear and its magnitude varies depending on the spectral region and the soil type. Within the framework of TEHRA project (an Italian Space Agency research initiative), a study was carried out to explore the combined use of MARMIT-2, a multilayer radiative transfer model of soil reflectance to estimate soil water content, and Machine Learning methods to address this challenge. Two local soil spectral libraries (SSLs), including both dry/wet samples and SMC (soil moisture content) values, collected over different locations in Italy between 2021 and 2022 (Maccarese-Pignola-Castelluccio and Jolanda di Savoia, respectively), have been used to investigate two different approaches.

The first one is devoted to the retrieval of soil moisture content. By performing the inversion and the calibration of MARMIT-2 it is possible to increase the dataset by adding further wet spectra (and SMC values) for each sample of the original spectral library. The wet soil reflectance is expressed in terms of dry soil reflectance and three free parameters: the thickness of the water layer L, the surface fraction of the wet soil ε, and the volume fraction of soil particles in the water layer δ. Given a dry sample and the corresponding wet measurement, the Nelder-Mead algorithm is used to minimize a cost function.  The calibration, instead, is performed by fitting a sigmoid function following the soil-by-soil approach. The dataset is generated by varying (L, ε, δ) to simulate wet reflectances and the corresponding SMC is calculated using the sigmoid and the parameters found during the calibration. A Machine Learning Regression Model (a Multilayer Perceptron) has been trained using Maccarese-Pignola-Castelluccio plus additional libraries made available by authors of MARMIT and tested using Jolanda di Savoia. Results are very promising: MAE: 5.088; R2 score: 0.844; RMSE: 6.165. The model has been applied also to different PRISMA images proving to be coherent with respect to the values measured in laboratory included in the SSL.

The second approach is to train a deep convolutional autoencoder capable of extracting the corresponding dry spectrum from a wet one. The dataset is composed by couples of wet and dry reflectances, resampled to PRISMA bands configuration and cleansed of water vapor absorption bands. The autoencoder consists of several blocks of convolutional layers, batch-normalization, and ReLU-activation functions. The downsampling is performed by average pooling and the upsampling with inverse convolutions. The model has been trained on Maccarese-Pignola-Castelluccio SSL, with additional samples added thanks to the inversion of MARMIT-2. Jolanda SSL has been kept aside to be used for testing the model (MAE: 0.04469, MSE: 0.00317, CS: 0.98). The autoencoder has been applied also to PRISMA images; however, further developments need to be carried out given the remarkable difference between simulated and real spaceborne data.

How to cite: Tricomi, A., Bruno, R., Casa, R., Mirzaei, S., Pascucci, S., Pignatti, S., Rossi, F., and Guarini, R.: Soil moisture effects: integrating physically based models and machine learning for enhanced retrieval and dryification strategies, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-15629, https://doi.org/10.5194/egusphere-egu24-15629, 2024.

Farmers, agricultural decision-makers, and land-use planners require updated, reliable, and accurate assessments of soil characteristics for sustainable management of natural resources. Saline and calcareous soils are a significant threat to crop production and can significantly reduce agricultural productivity, especially in arid and semi-arid regions. The accumulation of salt in the root zone restricts soil processes including nitrification, denitrification, and residue decomposition due to diminishing microorganism activity and soil biodiversity. On the other hand, in calcareous soils, the availability of some nutrients such as P, Fe, and Cu is limited due to the high pH. Given that traditional approaches to monitoring and mapping are costly and time-consuming, a rapid and efficient estimation of soil properties has been given attention by researchers using remote sensing data with field measurements. In this study, we have explored the performance of the PRISMA hyperspectral imagery satellite (prisma.asi.it) for estimating spatial variations of electrical conductivity (EC) and calcium carbonate equivalent (CCE). The study took place in the marginal lands of Sirjan Playa, southeast of Iran (Lat. 55°32′E, Lon. 29°23′N), which are mainly under pistachio cultivation. A total of twelve PRISMA L2D (BOA reflectance) images, acquired from June 2020 to December 2023; co-registered with the closest Sentinel-2 image (of about 0.5 pixel of RMS) and smoothed by the Savitzky-Golay filter (frame size of 7 and 3rd degree polynomial), were used. Field campaigns were performed to collect 250 soil samples from the top 15 cm of surface soil. Furthermore, the EC (min = 1.25%, max = 44.75%, std = 5.65%). and CCE (min = 1.25%, max = 44.75%, std = 5.65%) was measured using HCl. Both gaussian process regression (GPR) and the partial least squares regression (PLSR) algorithms were tested to predict EC and CCE from PRISMA 2D image spectra The results revealed that GPR achieved good prediction for CCE with a R² of 0.75, a root mean square error (RMSE) of 4.09%, and a ratio of performance to interquartile distance (RPIQ) of 2.75. The PLSR model, instead, showed the highest performance (R² = 0.63, RMSE = 44.8, RPIQ = 1.5) for predicting EC. These models were validated by the K-fold cross-validation approach (k = 10). The reason for the weaker salinity (EC) prediction by the PLSR could be attributed to the non-linear spectral behavior with respect to the salinity level. Furthermore, it seems that the presence of significant amounts of gypsum in the area could mask the accuracy of the EC prediction. Moreover, salt (i.e., the dominant salt is halite in the study area) diagnostic absorption bands occur in the atmospheric water vapor absorption region of soil spectra. Therefore, hyperspectral remote sensing appears to be a valuable resource for monitoring the spatiotemporal variation of EC (a fair prediction model with an RPIQ of 1.5) and CCE (a very good model with an RPIQ of 2.75). Further analysis should be done to better understand the effect of the external parameter (e.g., gypsum abundance) on the EC prediction.

How to cite: Mirzaei, S., Rasooli, N., and Pignatti, S.: Evaluating the spatial variations of soil salinity and calcium carbonate in marginal lands of Sirjan playa (Iran), using PRISMA hyperspectral imagery, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-20310, https://doi.org/10.5194/egusphere-egu24-20310, 2024.

Imaging spectroscopy is commonly used for many applications like soil, water, and vegetation. Digital soil mapping, especially by space-borne sensors, has become advantageous and promising due to its high efficiency. By using multispectral or hyperspectral images, topsoil properties could be estimated efficiently and accurately on a large area scale. Moreover, deep learning has been explored in the remote sensing community and achieved excellent performances in many remote sensing tasks. In this work, we explore deep learning methods to retrieve Soil Organic Carbon (SOC) value from DESIS hyperspectral images for the whole Bavaria state in Germany. For the hyperspectral data, we use all available DESIS images in Bavaria, which is 560 in total. Regarding the soil data, we combine SOC data from LFU (Bavarian State Environment Agency) and LUCAS 2018 (Land Use and Coverage Area frame Survey). Following a rigorous data selection process, we opted to include 1200 soil samples in our experiments. Starting from the raw hyperspectral images, we conduct a few preprocessing steps such as land cover masking, filtering by NDVI, building temporal composite, and then extracting patches surrounding each soil sample. These preprocessed patches are fed into deep learning models such as 1D CNN and 2D CNN, which are trained to predict the SOC value. To better interpret the model's performance, we also compute the SHAP(Shapley Additive Explanations) value for both frameworks. Specifically, we explore the SHAP value in spectral dimension for 1D CNN and analyze digital elevation features with 2D CNN in spatial dimension. During experiments, we split the whole dataset into train, validation, and test. To evaluate the performance, RMSE, R², and RPID are computed. For the specific structure of the models, many different parameters are investigated in parameter tuning. For each trial, 5 cross-validation is applied. In the end, we visualize the prediction results by a soil map.  From the results, the best-performed model could get RMSE 0.62 and R² 0.40 on the test set. Moreover, we find that the first-order derivative of the spectrum is the most important feature for predicting SOC, while 1D CNN is capable of extracting useful information from it and achieving excellent regression results with RMSE 0.66 and R² 0.32. Additionally, spectrums between 530 nm - 570 nm and 730 nm - 780 nm are the most informative according to SHAP analysis.

How to cite: Zhao, X., Heiden, U., Karlshöfer, P., Xiong, Z., and Zhu, X. X.: Soil Organic Carbon Retrieval from DESIS Images by CNN, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-9731, https://doi.org/10.5194/egusphere-egu24-9731, 2024.

Soil organic carbon (SOC) plays an important role in sequestering CO2 and assists in reducing atmospheric greenhouse gases in addition plays a critical role in maintaining the sustainability of grasslands. The valuable roles of SOC, make its accurate measurement critical however temporal changes in SOC are small and spatially vary.  Therefore, a large number of samples are required to detect the SOC changes which makes it a complex and costly task. Stratification is capable of improving the efficiency of sampling by reducing the number of samples and increasing the accuracy of SOC measurement. Stratification relies on assessing the relationship between SOC and environmental factors. Vegetation has the potential to be used as a proxy to spatially predict SOC.

This experiment aimed to assess the relationship between SOC and vegetation characteristics as a key factor in small areas with uniform climate and soil type. The three study sites were located in southern Queensland with subtropical climate. Short-term data was collected using the BOTANAL method and biomass harvesting over two years period in different seasons which included biomass, pasture composition, and vegetation type. Long-term data was extracted from various satellite images for up to 30 years which indicate the long-term effect of vegetation on SOC. Remote sensing data contained vegetation and soil indices.

The kriging method was applied to both soil and vegetation data to interpolate unsampled points for the study areas, then K-means clustering was used to cluster the data. Spearman rank-order correlation coefficient was used to assess the correlation between SOC clusters and vegetation factor clusters.

While some of the vegetation parameters have a significant correlation with SOC, the correlation is not consistent between different sites and different seasons. It can be concluded from this study that vegetation factors are not capable of using landscape clustering for SOC sampling on small scale.

How to cite: Ahmadi, S., Bonner, M., Mitchelle, E., Grace, P., and Rowlings, D.: Predicting the spatial distribution of SOC using remotely sensed data and vegetation data in southern Queensland’s grasslands , EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-13365, https://doi.org/10.5194/egusphere-egu24-13365, 2024.