Applications of Machine Learning and Remote Sensing in Soil and Water Conservation

1. Introduction

Soil and water play a pivotal role in various ecological processes, including nutrient cycling, water filtration, and habitat provision, which collectively support biodiversity and ecosystem stability. Soils contribute to the cycling of carbon, nitrogen, and phosphorus critical for plant growth and ecosystem productivity [1,2,3]. Water is essential for maintaining hydrological cycles, regulating temperature, and sustaining terrestrial and aquatic habitats [4,5]. These resources are fundamental to ecosystem resilience and functionality, impacting not only natural processes, but also human activities such as agriculture and urban development [6,7].

In addition to their ecological significance, soil and water resources are crucial for sustainable agricultural practices and food security [8]. However, their integrity is increasingly compromised by anthropogenic factors including climate change, population growth, deforestation, and unsustainable land use practices [9]. Climate change can exacerbate soil erosion, disrupt nutrient cycles, and affect water availability by altering precipitation patterns and increasing frequencies of extreme weather events [10]. Population growth and urban expansion place additional pressures on these resources, leading to overexploitation, pollution, and habitat loss [11]. Deforestation can further undermine soil structure and reduce land’s capacity to retain water, while poor land management practices can accelerate soil degradation and water contamination [12,13].

Addressing these challenges requires a comprehensive approach to soil and water conservation that encompasses a range of strategies aimed at mitigating negative effects of these stressors [14,15]. Effective soil conservation involves practices such as erosion control, moisture retention through irrigation management and organic amendments, and sustainable land use planning. Similarly, water conservation encompasses measures to enhance water quality, improve storage capacity, and promote efficient usage. These conservation practices are important not only for sustaining ecosystem health and agricultural productivity, but also for supporting broader environmental management goals, including wildfire mitigation and recovery. To address this, numerical models have traditionally been important tools in soil and water conservation [16,17]. However, their reliance on a limited set of variables and specific assumptions often results in prediction accuracy being heavily dependent on given input data [18,19,20,21]. Additionally, they may fail to account for uncertainties in the detection of climate change [18]. Thus, advanced tools that can complement or even replace traditional numerical models are needed. Integration of machine learning (ML) and remote sensing (RS) data presents a promising solution to limitations of traditional methods in soil and water conservation [22]. Advances in usage of RS data provide extensive spatial and temporal data, capturing environmental changes with high precision across large areas [23,24]. RS techniques such as multispectral and hyperspectral imaging, LiDAR, and synthetic aperture radar (SAR) enable the collection of data on various environmental parameters, including soil moisture, vegetation cover, land surface temperature, and water quality. When combined with ML techniques such as Random Forest (RF), Support Vector Machines (SVM), Convolutional Neural Networks (CNNs), and Gradient Boosting Machines (GBMs), these data can be analyzed to identify patterns, make predictions, and develop more effective conservation strategies. Combined with ML, these data can be analyzed to identify patterns of spatial and temporal data, make predictions, and establish more effective conservation strategies [25,26]. The ability of ML to process vast amounts of RS data enables extraction of valuable insights from complex, unstructured datasets, leading to improved accuracy in predicting soil and water resource changes. Additionally, the fusion of ML and RS allows for real-time monitoring and assessment, offering dynamic and responsive tools for decision-making in conservation practices [27,28]. This integrated approach can enhance resource optimization, increase precision of policy implementations, and facilitate data-driven decision-making in soil and water management.

The objective of this review is to explore the potential of ML and RS technologies in advancing soil and water conservation. This review aims to provide an overview of current research areas within soil and water conservation and to present how a combination of ML and RS can overcome limitations of traditional methods. The review will highlight applications of various ML techniques and RS data in different subfields of soil and water conservation, demonstrating their effectiveness in improving prediction accuracy and resource management. Additionally, this paper will discuss future research directions, focusing on integrating ML with RS and development of innovative solutions for sustainable soil and water conservation. The scope includes a detailed examination of how these advanced tools can be applied across diverse conservation challenges and implications for future policy and practice.

2. Materials and Methods

2.1 Searching/Classification Methodology

The literature review focused on collecting published journals that applied ML algorithms and RS data to conduct research related to soil and water conservation, as summarized in Figure 1.

To identify relevant publications, we utilized the web-based bibliographic database ‘ScienceDirect’ to identify relevant publications, using specific keywords such as “machine learning,” “remote sensing,” “soil conservation,” and “water conservation”. This search resulted in the identification of 200 studies conducted across 47 countries, as shown in Figure 2. Studies conducted in China, the United States, and Iran, with smaller but notable frequencies in Australia, Canada, Russia, and several European countries showed the highest frequencies. Studies conducted in other locations around the world, including parts of Africa, South America, and South Asia, were also marked, indicating their relative frequency in the data. Figure 3 illustrates trends in reviewed publications across four key research areas from 2006 to 2024. Wildfire management exhibited a dramatic increase in publications, peaking in 2022, likely due to the increasing impact of climate change, advances in ML and RS technologies, and increased global awareness driven by initiatives such as the UN’s SDGs and the Paris Agreement. Research studies on soil properties, hydrology, and water resources also peaked around the same time, although they showed a more gradual rise and fall. Frequency of research on biomass-vegetation remained relatively low in comparison, with minimal fluctuations throughout the period.

Research subjects of 200 studies were then identified and categorized into a total of 37 specific research topics, referred to as “subcategorized subjects,” including above-ground biomass, grassland biomass, ground biomass, soil conductivity, soil salinity, soil organic carbon (SOC), soil aggregate stability, soil chemistry, soil degradation, soil erodibility, soil matric potential, soil mercury, soil moisture, soil nutrients, soil total nitrogen, soil respiration, soil stiffness, soil texture, soil types, soil organic matter, soil water content and evapotranspiration, groundwater level, streamflow, surface water, water storage, sediment concentration, algal blooms, Secchi disk depth, sediment discharge, water quality, turbidity, evapotranspiration, flash flood water depth, inundation status, ocean surface CO_2, wildfire prediction, wildfire monitoring, and wildfire recovery as shown in Table 1. These subcategorized subjects were then reclassified into four research fields: 1) biomass-vegetation, 2) soil properties, 3) hydrology and water resources, and 4) wildfire management. While some subcategorized subjects of collected studies were closely related and ambiguous to distinguish, the classification focused on the objective of this study. The research field of soil properties had the highest number of publications, followed by wildfire management, hydrology and water resources, and biomass-vegetation research fields. Publications were distributed as follows: 93 (47%) papers on soil properties, 52 (26%) papers on wildfire management, 50 (25%) papers on hydrology and water resources, and 5 (2%) papers on biomass-vegetation as shown in Figure 4.

3. Results and Discussion

3.1. Types and Frequencies of RS Data Used in Soil and Water Conservation Research

Of 200 studies collected, a total of 41 different types of RS techniques were identified (Table A2). Figure 5 depicts the number of publications that utilized each RS data, highlighting only those RS data used more than twice across the 200 studies. The following RS data were used only once each: AGRS, ALOS-2, AVIRIS-NG, GF-1, Triplesat, PALSAR-2, Terra, ZH-1, ETM+, SVC, SRTM, NLCD, Himawari-8, TDC, AMSR-E, MERIS, MERRA-2, Chinese Environmental 1A satellite, GOES-16, TM, SAR, and SPOT-4. This study analyzed types and frequencies of RS data used across different research fields in soil and water conservation. RS data were classified based on four research fields. Below is an overview of the most frequently used RS data types within each field. Table 3 summarizes the number of publications and the top three most commonly used RS data in different environmental research fields, including biomass-vegetation, soil properties, hydrology and water resources, and wildfire management. In the field of biomass-vegetation, MODIS and UAVs were the most commonly used algorithms, appearing two times. These tools are crucial for large-scale biomass estimation and vegetation monitoring. For Soil Properties, Landsat 8 was the most frequently used (32 times). Its high-resolution images and multispectral capabilities are particularly effective for evaluating soil characteristics. In hydrology and water resources, Landsat 8 was again predominant. It was used 18 times for tracking changes in water bodies, flood monitoring, and resource management. Lastly, in wildfire management, MODIS was the leading algorithm. It was used 20 times, offering vital support for real-time fire monitoring and management across damaged areas.

Landsat and MODIS are essential satellite sensors, each suitable for specific research applications due to differences in data characteristics and resolution. Landsat 8, Sentinel-2, and MODIS are frequently used in soil and water resource studies due to their technical features. Landsat, with its high spatial resolution of 30 meters, is ideal for detailed studies of soil characteristics and hydrology. It allows for precise analysis of small-scale features such as soil moisture and water flow. In contrast, MODIS with a lower resolution of 250 to 1,000 meters but a higher revisit frequency of twice daily is better suited for monitoring large-scale, rapidly changing events like wildfires. MODIS is particularly effective for wildfire management due to its ability to capture wide areas quickly and detect heat through infrared bands, while Landsat’s detailed imagery is preferred for soil and hydrological research.

These RS data also provide broad spatial and temporal coverage. MODIS collects global data daily, while Landsat 8 and Sentinel-2 offer high-resolution imagery with revisit periods of 16 days and 5 days, respectively. This allows for long-term monitoring of environmental changes and time-series analysis of hydrological events or soil changes. Moreover, these RS data offer long-term historical datasets, which is a significant advantage. For instance, the Landsat program has been accumulating data since the 1970s, making it valuable for analyzing decades of environmental changes or soil erosion patterns. Similarly, Sentinel-2 and MODIS provide consistent datasets over the years, supporting accurate trend analysis and modeling. The preference for these applications reflects their efficiency in capturing crucial information for soil and water resource conservation.

3.2. Types and Frequencies of ML Algorithm Used in Soil and Water Conservation Research

A total of 50 distinct ML algorithms were identified and their usage frequencies are illustrated in Figure 6. Algorithms that appeared more than twice are shown, while those not depicted to prevent excessive clutter, each used only once, include RTM, ANFIS, ARD, BAGGING, BDT, SA, SCA-Elman, SoLIM, SOM, SR, PSO-SVR, B-CART, CBR, DR, DELM, GAN, BAYE, LGBM, GSC, GRNN, PCR, PKR, RPART, MR-CNN, LMM, EBP, ETR, ELR, EM, EFS, EPR, DBN, DRF, DMP, LDA, MDN, MLPR, MT, Nue-SICR, SICR, FR, FCN, FNN, AdaBag, BST, M5P, YOLO, and IF. The number of publications across research fields related to RS data and the top three most commonly used algorithms along with their frequency of usage is shown in Table 4. In the biomass-vegetation field, RF and ANN were used most frequently. The Soil properties field had the highest number of publications, with RF being the most prevalent algorithm. Similarly, hydrology and water resources and wildfire management also showed a preference for RF as the leading algorithm. In cases where algorithms were used with the same frequency, they were ranked equally. The total number of algorithm usages does not necessarily match the total number of publications due to the use of multiple algorithms in some studies.

RF, ANN, and SVM were the most commonly used ML algorithms across different research fields. RF is an ensemble learning method that constructs multiple decision trees during training and aggregates their outputs for prediction. This approach enhances accuracy and mitigates overfitting by averaging results of numerous trees, which can help manage noisy and high-dimensional datasets common in environmental research. RF is frequently used in environmental research due to its effectiveness in identifying feature importance, which helps researchers determine the most influential variables affecting environmental processes such as water quality prediction [143,156] and land cover classification[80] . Its robustness and capacity to manage large datasets enhance its utility across various ecological and hydrological applications. ANNs excel at modeling complex, non-linear relationships capturing intricate data patterns. Comprised of interconnected neurons that can adjust weights through iterative training, ANNs can optimize predictions for complex environmental variables. This adaptability makes them highly effective for tasks such as flood prediction and groundwater level forecasting, where capturing non-linear interactions is essential. Their ability to generalize from extensive datasets ensures precise modeling of dynamic environmental systems. This adaptability makes ANNs particularly useful for tasks such as flood prediction or groundwater level forecasting, where capturing non-linear interactions is crucial. Their capability to generalize from extensive training datasets enables accurate modeling of dynamic environmental systems and phenomena.

3.3 Field-Specific Observations

3.3.1. Biomass-Vegetation

In the research field of biomass and vegetation, researchers frequently combine MODIS and UAV data with ML algorithms such as RF and ANN to assess vegetation conditions and estimate biomass. MODIS is effective for large-scale vegetation monitoring and tracking seasonal changes and long-term trends due to its global coverage and frequent revisit times. UAV, on the other hand, provides high spatial resolution and flexibility, capturing detailed images at plot or landscape levels. Equipped with sensors like RGB cameras, multispectral, hyperspectral, and LiDAR, UAVs can be utilized to detect specific wavelengths that indicate plant health, structure, and biomass. RF is particularly useful for managing large datasets and modeling complex relationships between spectral data and vegetation attributes such as Leaf Area Index (LAI), chlorophyll content, and biomass density.

3.3.2. Soil Properties

In the research field of soil properties, numerous studies have combined RS data from Landsat 8 and Sentinel-2 with RF. Landsat 8 and Sentinel-2 are widely utilized for mapping and monitoring soil properties at regional and global scales due to their high-resolution multispectral imagery. These satellites provide data that can be used to derive indicators related to soil properties such as organic carbon content, soil moisture, and soil texture. ML algorithms such as RF and SVM are particularly effective in this field because they can handle large and complex datasets and model nonlinear relationships between RS-derived variables and soil attributes. RF is especially useful for processing large amounts of data and analyzing complex patterns that link spectral information with soil properties, while SVM is often employed to classify and predict soil properties by maximizing the margin between different types of soil data. The integration of RS data and ML algorithms enables more accurate and efficient prediction and mapping of soil characteristics, which is essential for sustainable land management, agriculture, and environmental conservation.

3.3.3. Hydrology and Water Resources

In the research field of hydrology and water resources, studies have mainly focused on predicting river flow, groundwater levels, and water quality parameters. RS data from Landsat 8 and Sentinel-2 have been widely used due to their ability to capture detailed spatial and temporal information related to water bodies and terrain. These satellites can provide high-resolution multispectral imagery that is crucial for monitoring and assessing various hydrological variables such as surface water extent, vegetation cover, and soil moisture, which can directly influence hydrological processes. In this field, RF and SVR are preferred for modeling complex and nonlinear hydrological processes due to their robustness and accuracy. RF has been utilized to process large datasets and identify patterns in river flow and groundwater levels, while SVR has been employed to predict continuous variables such as water quality parameters, leveraging its ability to model relationships in data with limited observations. The combination of RS data from Landsat 8 and Sentinel-2 with ML algorithms such as RF and SVR can enhance the ability to accurately predict and manage water resources, which is essential for sustainable water resource management and planning.

3.3.4. Wildfire Management

In the research field of wildfire management, ML techniques are extensively used to predict and monitor wildfire occurrences. MODIS, known for its daily global coverage, is one of the most frequently utilized RS data sources in this domain. MODIS provides critical information for real-time monitoring and historical analysis of wildfires, enabling the detection of active fires, mapping of burn scars, and assessment of the extent of fire-affected areas. Its frequent revisit times are particularly useful for tracking wildfire progression and immediate impacts. To predict fire-prone areas and assess post-fire effects on soil and vegetation, MODIS data are commonly used with RF and SVM. RF is highly effective in identifying complex patterns among environmental variables, such as vegetation type, moisture content, and weather conditions, which can influence wildfire risk. SVM is also employed to classify regions based on fire vulnerability and to assess the severity of fires within ecosystems. The integration of MODIS data with ML algorithms such as RF and SVM can enhance the ability to predict wildfires, mitigate risks, and manage post-fire recovery efforts, contributing to more effective wildfire management strategies.

4. Challenges and Limitations

4.1. Data-Related Challenges

One of the fundamental challenges in the application of ML to soil and water conservation lies in the availability, quality, and consistency of RS data. RS data are often characterized by varying spatial, spectral, and temporal resolutions, which can introduce significant variability into datasets used for model training and validation. For instance, while Landsat can provide data with moderate (15 ~ 120 m) spatial resolution and a long temporal record, Sentinel-2 offers higher spatial resolution (10 ~ 60 m) but with a shorter historical dataset. The integration of these diverse data sources can be problematic, as differences in resolution, sensor characteristics, and data acquisition periods can lead to discrepancies that need to be harmonized. In addition, inconsistent or incomplete datasets are a common issue, particularly in regions with limited historical monitoring or where cloud cover frequently obstructs satellite observations. These data gaps can introduce biases into ML models, leading to inaccurate results for prediction. For example, if training data are not representative of the full range of environmental conditions, the model may fail to generalize effectively, resulting in poor performance when applied to new or unseen conditions. Moreover, preprocessing of RS data, including tasks such as georeferencing, atmospheric correction, and resampling to a common spatial and temporal grid, can be technically demanding and resource-intensive. Harmonization of data from multiple sensors requires advanced techniques, such as data fusion and cross-calibration, to ensure consistency of inputs for ML models.

4.2. Technological Limitations

The implementation of ML in the context of soil and water conservation is often constrained by availability of computational resources and inherent complexity of algorithms employed. High-dimensional datasets—characterized by a large number of variables and extensive temporal records—are common in environmental studies. Processing these datasets requires significant computational power, including high-performance computing (HPC) clusters or cloud-based solutions, which may not be readily available in all research settings. The storage of such large volumes of data also poses challenges, as traditional data storage solutions might be insufficient to handle the scale and complexity of RS data. Furthermore, the complexity of ML algorithms—particularly those involving deep learning (e.g., convolutional neural networks, recurrent neural networks)—requires not only computational resources, but also specialized expertise. Deep learning models, for instance, often involve a large number of hyperparameters and require extensive tuning to achieve optimal performances. This complexity can be a significant barrier to the adoption of ML in resource-limited settings, where access to both infrastructure and skilled personnel might be limited. Model interpretability is another a significant concern in environmental applications. Many ML models, especially those classified as “black-box” models, offer limited insights into the underlying decision-making processes, which can hinder their acceptance and use in policy-making or by stakeholders. Decision-makers often require not only accurate predictions, but also an understanding of the rationale, which can be challenging to provide with complex ML models. Finally, the scalability limits the broader applicability of ML models across diverse geographic regions with availability constraints.

4.3. Implementation Issues

The practical application of ML in conservation efforts is also limited by challenges in model interpretation and transparency. Many ML techniques, particularly deep learning models, operate as “black boxes,” making it difficult for stakeholders to understand the decision-making process. This lack of interpretability can hinder the integration of ML outcomes into policy and management strategies.

5. Future Directions and Research Opportunities

Future research on ML for soil and water conservation will benefit from advancements in big data analytics and cloud computing, which can address current limitations related to data processing and storage. The development of more interpretable ML models, such as explainable AI (XAI), is also expected to enhance the integration of ML into decision-making, making outputs more accessible to non-experts. There is also a growing need for research that focuses on the development of hybrid models that can strength various ML algorithms and RS data to enhance prediction accuracy and provide a comprehensive understanding of soil and water processes. Such approaches can improve prediction accuracy and provide a more holistic understanding of environmental processes. Additionally, research should explore the potential of integrating ground-based sensor networks with RS data, enhancing real-time monitoring and predictive capabilities. The adoption of ML in soil and water conservation has significant policy implications, particularly in the context of climate change adaptation and sustainable land management. Future research should emphasize the importance of interdisciplinary collaboration, bringing together experts in ML, environmental science, and policy to ensure that technological advancements can translate into practical conservation outcomes. Collaborative efforts should also focus on capacity-building initiatives to equip stakeholders with necessary skills and knowledge to implement ML-driven solutions effectively.

The selection of algorithms in each field is influenced by characteristics of the data, complexity of the problem, and objectives of the prediction. RF is popular in many environmental and resource management fields due to its ability to handle complex interactions among variables and manage nonlinearity effectively. SVM is useful for high-dimensional data or problems with nonlinear boundaries. Meanwhile, ANN and MLP excel at learning complex nonlinear patterns. The choice of these algorithms is made to optimize outcomes based on characteristics of the data and goals of problem-solving. The selection of ML algorithms varies depending on factors such as data characteristics, problem complexity, and interpretability. Future research should focus on optimizing algorithm selection and improvement by considering these factors. Additionally, comparing performances of various algorithms and conducting comprehensive evaluations will be crucial for proposing the most suitable methodologies for each field, offering significant insights into best practices.