Ruhollah Taghizadeh

An understanding of the key factors and processes influencing the variability of soil organic carbon (SOC) is essential for the development of effective policies aimed at enhancing carbon storage in soils to mitigate climate change. In recent years, complex computational approaches from the field of machine learning (ML) have been developed for modelling and mapping SOC in various ecosystems and over large areas. However, in order to understand the processes that account for SOC variability from ML models and to serve as a basis for new scientific discoveries, the predictions made by these data-driven models must be accurately explained and interpreted. In this research, we introduce a novel explanation approach applicable to any ML model and investigate the significance of environmental features to explain SOC variability across Germany. The methodology employed in this study involves training multiple ML models using SOC content measurements from the LUCAS dataset and incorporating environmental features derived from Google Earth Engine (GEE) as explanatory variables. Thereafter, an explanation model is applied to elucidate what the ML models have learned about the relationship between environmental features and SOC content in a supervised manner. In our approach, a post hoc model is trained to estimate the contribution of specific inputs to the outputs of the trained ML models. The results of this study indicate that different classes of ML models rely on interpretable but distinct environmental features to explain SOC variability. Decision tree-based models, such as random forest (RF) and gradient boosting, highlight the importance of topographic features. Conversely, soil chemical information, particularly pH, is crucial for the performance of neural networks and linear regression models. Therefore, interpreting data-driven studies requires a carefully structured approach, guided by expert knowledge and a deep understanding of the models being analysed.

Salinization is a threat to global agricultural and soil resource allocation. Current investigations of global soil salinity are limited to coarse spatial resolution of the available datasets (\textgreater250 m) and semiqualitative classification rules (five ranks). Based on these two limitations, we proposed a framework to quantitatively estimate global soil salt content in five climate regions at 10 m by integrating Sentinel-1/2 remotely sensed images, climate, parent material, terrain data, and machine learning. In hyper-arid and arid region, models established using Sentinel-2 and other geospatial data showed the highest accuracy with R2 of 0.85 and 0.62, respectively. In semi-arid, dry sub-humid, and humid regions, models performed best using Sentinel-1, Sentinel-2, and other geospatial data with R2 of 0.87, 0.80, and 0.87, respectively. The accuracy of the global models is considerable with field validation in Iran and Xinjiang, and compared with digitized salinity maps in California, Brazil, Turkey, South Africa, and Shandong. The proportion of extremely saline soils in Europe is 10.21%, followed by South America (5.91%), Oceania (5.80%), North America (4.05%), Asia (1.19%), and Africa (1.11%). Climatic conditions, groundwater, and salinity index are key covariates in global soil salinity estimation. Use of radar data improves estimation accuracy in wet regions. The map of global soil salinity at 10 m provides a detailed, high-precision basis for soil property investigation and resource management.

Soil acidification is an ongoing problem in intensively cultivated croplands due to inefficient and excessive nitrogen (N) fertilization. We collected high-resolution data comprising 19,969 topsoil (0–20 cm) samples from the Land Use and Coverage Area frame Survey (LUCAS) of the European commission in 2009 to assess the impact of N fertilization on buffering substances such as carbonates and base cations. We have only considered the impacts of mineral fertilizers from the total added N, and a N use efficiency of 60 %. Nitrogen fertilization adds annually 6.1 × 107 kmol H+ to European croplands, leading to annual loss of 6.1 × 109 kg CaCO3. Assuming similar acidification during the next 50 years, soil carbonates will be completely removed from 3.4 × 106 ha of European croplands. In carbonate-free soils, annual loss of 2.1 × 107 kmol of basic cations will lead to strong acidification of at least 2.6 million ha of European croplands within the next 50 years. Inorganic carbon and basic cation losses at such rapid scale tremendously drop the nutrient status and production potential of croplands. Soil liming to ameliorate acidity increases pH only temporarily and with additional financial and environmental costs. Only the direct loss of soil carbonate stocks and compensation of carbonate-related CO2 correspond to about 1.5 % of the proposed budget of the European commission for 2023. Thus, controlling and decreasing soil acidification is crucial to avoid degradation of agricultural soils, which can be done by adopting best management practices and increasing nutrient use efficiency. Regular screening or monitoring of carbonate and base cations contents, especially for soils, where the carbonate stocks are at critical levels, are urgently necessary.

Digital soil mapping (DSM) can be used to predict soils at unvisited sites, but problems arise when predictions are needed in areas without any soil observations. In such situations, DSM can still extend the results from reference areas with soil data to target areas that are alike in terms of soil-forming factors and obey the same rules. Such DSM methods have low accuracy due to the complexity of spatial variation in soil, and the difficulty of matching soil-forming factors exactly between reference and target areas. A new approach for extrapolating soil information from reference to target areas is proposed in the current research. We evaluated the ability of a semi-supervised learning (SSLR→T) approach compared to a supervised learning (SLR→T) approach for extrapolating soil classes in two areas (reference and target areas) in central Iran. The SSLR→T used soil observations from the reference area and covariates from both areas. Then, the learned knowledge produced by SSLR→T was transferred to the target area to estimate soil classes. The findings revealed that SSLR→T resulted in higher overall accuracy (0.65) and kappa index (0.44) in the target area compared to the SLR→T (overall accuracy = 0.40 and kappa index = 0.18). Furthermore, the SSLR→T produced the lower values of the confusion index (mean = 0.66) compared to the SLR→T (mean = 0.80). This indicated that the SSLR→T could not only increase the accuracy but also decrease the uncertainty of the soil class predictions, compared to the spatial extrapolation predictions derived from the SLR→T. Generally, these findings indicated that leveraging covariate information from the target area during the training of DSM models in the reference area could successfully improve the generalization power of the models, indicating the effectiveness of SSLR→T for spatial extrapolation.

Knowledge about the spatial distribution of soil particle size fractions (PSF) is critical for sustainable management and resource assessment of the agricultural regions. Although conventional machine learning algorithms, such as random forest (RF) or support vector machine, have been extensively used in digital soil mapping to predict the PSF, less research examined the potential of state-of-the-art deep learning approaches for such processing. Importantly, deep learning approaches such as convolutional neural networks (CNNs) are able to incorporate contextual information about the landscape, which is of great use for DSM analysis. Accordingly, this study addresses this much-needed investigation by using a patch-based, multi-task CNN for predicting PSF of clay, sand, and silt contents at six standard layers given as soil depth increments as recommended by the GlobalSoilMap.net (i.e., 0–5, 5–15, 15–30, 30–60, 60–100, 100–200 cm). The depth functions were derived from equal-area smoothing splines in a region covering large parts (~140,000 km2) of central Iran. The robustness of the proposed architecture is evaluated against RF. Additionally, to allow a fairer comparison between RF and CNN models, we used simple smoothing (mean) filters to effectively reproduce the auxiliary data which are then fed in the RF (RF*). To evaluate the three models, we established a training (75%) and test set (25%). According to the test set, for all soil depths and all PSFs, the results demonstrate that CNN consistently outperforms RF and RF* in terms of root mean square error (RMSE) and coefficient of determination (R2). At the top layer, for example, CNN decreased the RMSE values for clay, sand, and silt contents compared to the RF (22.4%, 18.9%, and 10.7%) and RF* (18.0%, 7.4%, and 9.6%). These findings indicate that even the use of feature-engineered auxiliary data did not enable the RF* models to reach the performance of CNN. The resulting maps can be used as valuable baseline soil information for the effective management of agricultural and environmental resources in the study area and beyond.