Self-Explaining Attention-Based Deep Learning Models for Soil Property Regression Using Multi-Temporal Sentinel-2 Data

1. Introduction

A crucial, non-renewable natural resource, soil serves as the basis for both human civilization and terrestrial ecosystems. It serves as the foundation for plant growth, offers habitat to innumerable creatures, and is crucial in controlling the cycles of carbon, nutrients, and water that are necessary for ecosystem health and agricultural output [1]. Soil is a property created at the Earth’s surface by the combined effect of parent material, climate, water, topography, flora, and fauna, and consisting of a complex and dynamic mixture of minerals, organic matter, water, air, and living beings [2]. A variety of soil types, each with distinct physical and chemical characteristics, are produced by this special composition and the processes that shape it.

The significance of soil goes much beyond its use in agriculture. In addition to being a natural habitat for plants and animals, soil also acts as a filter and groundwater storage facility, a medium for the breakdown and recycling of organic waste, and a crucial regulator of climate and atmospheric gases [3]. Therefore, soil is essential for food security as well as for preserving environmental quality and ecosystem services as well.

The ability of the soil to promote productivity, maintain ecological functions like carbon sequestration and nutrient cycling, and support plant and microbial life is reflected in the broader notion of soil health, which is becoming more widely acknowledged [4]. Strong plant growth, biological diversity, and resource efficiency are all made possible by healthy soils. Conversely, soil degradation brought on by pollution, erosion, compaction, or salinization can drastically lower ecosystem resilience and productivity [5].

The function and health of soil are largely dependent on its physical (texture, structure, porosity, and bulk density), chemical (pH and nutrient content), and biological (microbial activity and organic matter concentration) qualities [6]. Soil fertility, water retention, aeration, root development, and the capacity to sustain a variety of biological communities are all directly impacted by these characteristics. For instance, soil structure affects aeration and root penetration, whereas soil texture controls drainage and water retention [7]. Organic matter-rich soils, such mollisols and alfisols, have high fertility and they are among the most productive for agriculture due to their ability to retain water [8]. On the other hand, degraded soils or entisols, which have little nutrient content or weak structure, may need significant management inputs to sustain productivity [9]. Modifications to these characteristics that can drastically impact crop yields and ecological resilience, such as decreased aggregation or increased compaction [10].

Thus, knowledge of and control over soil characteristics are crucial for long-term ecosystem health, sustainable agriculture, and management of this priceless asset.

Soil is essential for maintaining crops as well as for controlling hydrological processes, filtering pollutants, and storing carbon, all of which support environmental health and climate regulation [8].

Sustainable soil management techniques, including crop rotation, cover crops, conservation tillage, and organic amendments, are strongly advocated as ways to combat degradation. It has been demonstrated that these methods promote nitrogen cycling, boost organic matter, improve soil structure, and encourage beneficial microbial activity [11].

Sustainable soil stewardship is crucial in light of global issues including land degradation, climate change, and the rising need for food. Long-term agricultural productivity, environmental quality, and future generations’ access to food depend on preserving and improving soil health [12].

Research Objectives

Since neural networks are frequently "black boxes," the goal is to determine which Sentinel-2 image input features—bands and periods—have the biggest impact on soil property prediction. As a result, consumers benefit from more openness, less data is needed for future applications, and models with different sources can be used (drones, more satellites) with a restricted number of channels.

2. Background and Related Work

Recent developments in artificial intelligence (AI) have completely changed the agricultural industry and ushered in the era of Agriculture 5.0, where precision agriculture is made possible by the convergence of interconnected technologies like robotics, remote sensing, the Internet of Things (IoT), and advanced data analytics [13].

In order to maximize crop management, resource allocation, and yield prediction, these AI-driven systems analyze enormous volumes of real-time data from a variety of sources, including weather stations, satellite imaging, soil sensors, and unmanned aerial vehicles. Using deep learning models and machine learning (ML), these days, agronomists and farmers may make data-driven, well-informed decisions that boost food systems’ resilience to climate change, increase productivity, and lessen their negative effects on the environment [13].

Even though contemporary AI models are remarkably predictive, interpretability and transparency have become increasingly problematic due to their growing complexity, especially as deep learning architectures like Long Short-Term Memory (LSTM) networks have become more widely used. These alleged "black box" devices frequently don’t have the ability to give concise, intelligible justifications for their results, which may prevent them from being used in high-stakes industries like agriculture where responsibility, trust, and regulatory compliance are crucial [14]. A strong framework for explainability is required since the opacity of AI models is particularly troublesome in situations when judgments affect resource distribution, land management, or environmental stewardship [14].

The goal of Explainable Artificial Intelligence (XAI), a crucial field, is to demystify the sophisticated AI systems’ decision-making processes. The goal of XAI is to provide both global insights into model behavior and local, instance-specific justifications for individual predictions. Examples of these techniques include SHapley Additive exPlanations (SHAP), Local Interpretable Model agnostic Explanations (LIME), Permutation Importance (PI), and Partial Dependence Plots (PDP) [14,15].

These techniques help stakeholders comprehend the directionality of their effects, the relative relevance of input attributes, consider the ways in which factors interact, therefore encouraging more trust in suggestions generated by AI [15].

The General Data Protection Regulation (GDPR), which requires the right to an explanation for automated choices that impact humans, is one regulatory framework that XAI is essential to guaranteeing compliance with [14].

XAI has several different applications in agriculture. Explainability, in general, helps farmers, agronomists, and policymakers to examine, verify, and improve AI-driven insights by bridging the gap between sophisticated analytics and real-world field application [13].

Transparent models allow for cooperative communication between human expertise and machine intelligence, which empowers stakeholders to identify abnormalities, pose important queries, and customize treatments for particular agro-ecological settings [14]. XAI improves the accountability and transparency of suggestions for crop selection, fertilizer application, pest control, and irrigation scheduling in smart farming applications, ultimately promoting more effective and sustainable farming methods [13].

With a particular focus on soil parameter estimate, XAI is essential for a number of reasons. Plant development, ecosystem function, and climate regulation are all based on the characteristics of the soil, including temperature, moisture, organic carbon content, and nutrient availability [16]. To maximize input use, these characteristics must be estimated accurately in order to enhance productivity projections and reduce environmental hazards. However, advanced modeling techniques are required due to the complexity and heterogeneity of soil systems as well as the large dimensionality of sensor and remote sensing data. Researchers and practitioners can analyze model predictions using XAI techniques, determine which environmental characteristics are most influential, then evaluate if the linkages that have been learned are consistent with accepted soil science principles [16].

For example, research using XAI to predict soil temperature has shown that air temperature and radiation are the main factors, and that model transparency improves the accuracy and usefulness of prediction tools for soil management [16]. Furthermore, by exposing data biases, identifying erroneous correlations, and pointing out regions that require more data collection or domain knowledge integration, XAI helps to continuously enhance soil models [13,16]. In addition to increasing model accuracy, this iterative feedback loop guarantees that AI-driven soil parameter estimation stays rooted in physical reality and is flexible enough to accommodate changing problems in agriculture [13,16].

Building trust, complying with regulations, and utilizing the full potential of digital transformation in sustainable soil management will all depend on the integration of XAI as agriculture continues to embrace AI on a large scale [13,14].

Using plant development characteristics, Abekoon et al. [17] created an explainable machine learning model to forecast soil levels of nitrogen (N), phosphorous (P), and potassium (K) in cabbage production. To interpret the model’s predictions, the study used Local Interpretable Model-agnostic Explanations (LIME) and SHapley Additive exPlanations (SHAP), increasing the AI system’s openness and credibility.

Over the course of 85 days, the researchers gathered data from cabbage plants cultivated in the Sri Lanka’s central hills. The model’s inputs included important plant characteristics like the number of leaves, plant height, average leaf area, and number of days. Higher plant height and leaf count had a favorable effect on the prediction of soil nutrient levels, according to SHAP research, however more days and a bigger average leaf area had a negative effect. In addition to providing thorough interpretations for individual forecasts, LIME offered instance-specific explanations, validating the global insights from SHAP.

15 greenhouse-grown cabbage plants at different stages of growth were examined in order to validate the model, and the actual measurements and the anticipated nutrient levels nearly matched.

Kakhani et al. [18] used XAI to improve the interpretability of soil organic carbon (SOC) prediction models, addressing a rising requirement in digital soil mapping. Although machine learning techniques like Random Forests (RF) have demonstrated encouraging results in simulating SOC, their opaqueness sometimes restricts their applicability in research and policymaking. In order to clarify the inner workings of SOC prediction models trained on environmental factors, this study constructed SHAP.

The scientists trained RF models to predict SOC using data from two different sites in England. They then used SHAP to calculate the relative contributions of each input variable, such as soil depth, topography, remote sensing indices, and land cover. In addition to confirming the predominance of factors like elevation and soil depth, SHAP also showed subtle feature interactions that would otherwise go undetected in conventional feature importance scores.

Interestingly, the study showed that improving model explainability did not result in a decrease in prediction accuracy.

The interpretability of convolutional neural networks (CNNs) in the context of digital soil mapping (DSM) is examined in a study by Beucher et al. [19], which focuses on the classification of acid sulfate (AS) soils in Denmark’s Jutland wetlands.

Using a dataset of 5,885 labeled soil samples and 14 optimized environmental covariates derived via recursive feature elimination, the authors evaluate CNNs and Random Forest models for binary classification of possible AS soil occurrence. CNNs perform marginally better than RFs in terms of prediction accuracy (68% vs. 61–63%), but their main contribution is the effective use of SHAP to show the local and spatial effects of variables. Key indicators identified by the study include detrended DEM, depth to pre-Quaternary deposits and the distance to the Littorina Sea shoreline. Beucher et al. demonstrate how spatial variables grouped in image-like patches can supply CNNs with contextual information, enabling a more detailed representation of environmental gradients.

Beucher et al. show that CNNs can benefit from contextual information provided by spatial covariates arranged in image-like patches, allowing for a more nuanced representation of environmental gradients.

In order to increase the accuracy of crop output forecasts, Mohan et al. [20] investigate the merging of artificial intelligence and explainable AI. The authors stress that although AI techniques like random forest models, deep neural networks, and support vector machines are frequently used in agriculture, their interpretability issues still pose a obstacle to further adoption. The authors created a hybrid prediction system that uses SHAP and LIME for model interpretability and a Deep Neural Network for yield prediction. By combining crop management information, soil characteristics (pH, organic content), and meteorological parameters (rainfall, temperature, and humidity), they applied this approach to a dataset with over 10,000 occurrences spread over several growing regions.

The DNN model outperformed other machine learning models including Random Forest (MAPE = 9.83%) and SVM (MAPE = 11.27%), with a Mean Absolute Percentage Error (MAPE) of 7.56% and an R2 score of 0.92. These findings show that intricate, nonlinear relationships between environmental elements and agricultural productivity more efficiently can be captured by deep learning architectures.

Crucially, the authors employed SHAP to pinpoint and illustrate the main forces influencing the DNN’s predictions. For example, average soil nitrogen levels, rainfall during the flowering period, and the variations in temperature throughout the vegetative phase were consistently found to be the most significant characteristics in all regional models. Higher rainfall variability had a detrimental effect on yield, as demonstrated by SHAP summary plots, whereas stable nitrogen levels were associated with high production in a positive way.

Additionally, localized explanations for individual forecasts were produced using LIME, providing assistance in the interpretation of anomalies, such as surprisingly poor yields in fields with otherwise ideal circumstances.

In order to improve agricultural decision-making processes in contexts with limited data and high risk, Shams et al. [21] look into the use of XAI to improve crop recommendation systems. Despite their strong predictive capabilities, the authors contend that traditional AI models employed for crop recommendations frequently lack interpretability, which undermines farmers’ and agronomists’ confidence and practical implementation.

The study presents an integrated system that combines SHAP for interpretability with an optimized Extreme Gradient Boosting (XGBoost) model for classification. A real-world dataset from Egypt’s agricultural zones was used to train the model. This dataset included information on crop yield history, soil type, pH, rainfall, temperature, and levels of nitrogen, phosphorus, and potassium.

Performance-wise, the XGBoost model with classification accuracy of 96.2

The authors explained the model’s recommendations both locally and globally using SHAP. The top three criteria determining crop recommendations, according to the global SHAP research, were temperature range, yearly rainfall, and soil nitrogen concentration. For instance, rice recommendations were mostly linked to high rainfall and slightly acidic soil (pH 6.0), but wheat was heavily preferred under moderate nitrogen and low rainfall conditions. To illustrate how the model modified suggestions in response to subtle variations in feature values, local SHAP explanations were applied to particular case studies. These individualized explanations demonstrated how sensitive the model was to slight changes in the pH or potassium content of the soil, providing farmers with useful information about how to apply soil amendments.

Additionally, the study evaluated the interpretability of the domain-specific explanations from specialists. According to a user assessment, 85% of agronomists thought SHAP plots were useful for comprehending the recommendations for certain crops, suggesting that XAI-enhanced decision support tools have a lot of promise for real-world use.

A study on the application of XAI techniques in combination with LSTM networks to improve the precision and comprehensibility of soil temperature prediction models is presented by Geng et al. [22]. Agricultural planning, irrigation scheduling, and seed germination all depend on accurate soil temperature predictions, but standard models frequently fail to account for nonlinear and temporal correlations in environmental data.

The authors of this study created an LSTM-based model that was trained over a five-year period using a dataset that included soil and weather data from five distinct Chinese regions. Air temperature, relative humidity, solar radiation, wind speed, and soil moisture content were important input characteristics. With a coefficient of determination (R²) of 0.92, and a Root Mean Square Error (RMSE) of 1.18°C, the LSTM model showed good predictive ability, surpassing conventional models like Random Forest (RMSE = 1.65°C) and Support Vector Regression (RMSE = 2.07°C).

The authors used SHAP to examine how each input feature contributed to the model’s output in order to overcome the interpretability issue. According to SHAP values, air temperature accounted for more than half of the predictive power, making it the most important element. Soil moisture and solar radiation came in second and third, respectively. Curiously, the impact of wind speed and relative humidity differed greatly by season, underscoring how feature relevance in time series models depends on context.

To gain a better understanding of the LSTM’s basic logic, temporal SHAP value analysis was also employed to illustrate how each feature’s effect changed across various time steps. This method improved the understanding of model sensitivity to extreme weather events and allowed researchers to identify aberrant prediction patterns.

AgroXAI, a novel crop recommendation system that combines XAI with machine learning algorithms to improve decision-making in the context of Agriculture 4.0, is introduced by Turgut et al. [23]. The study tackles a prevalent issue in intelligent agricultural systems: although AI models provide accurate forecasts, their opaqueness undermines user confidence and restricts practical implementation among farmers and agronomists.

An XGBoost model trained on a curated dataset of 7,000 Turkish agricultural records is used by the AgroXAI framework. These data contain characteristics like the type of soil, climatic zone, pH, precipitation, average temperature, and concentrations of macronutrients (NPK).

The model suggests the best crops (such as wheat, corn, sunflower, and cotton) based on the agro-environmental characteristics of a particular area. With an accuracy of 97.6%, the XGBoost model outperformed other models that were examined, including Random Forests (94.8%) and Decision Trees (91.4%).

The system combines SHAP and LIME to provide recommendations both locally and internationally, increasing transparency. According to the SHAP analysis, the most important variables influencing all of the suggestions were pH levels, precipitation patterns, and soil nitrogen levels. For example, the model consistently recommended cotton in warmer, high-nitrogen, low-precipitation zones and wheat in areas with moderate nitrogen and slightly alkaline soil (pH 7–7.5). LIME was used to produce case-specific interpretations that explain the recommendations for specific crops under a specific input combination.

Using data from unmanned aerial vehicles (UAVs), Kumar et al. [24] describe a solid study aimed at enhancing corn yield prediction by combining ML models and XAI approaches. The necessity for both interpretability and high-accuracy forecasting in actual agricultural situations is what motivates the research.

Using multispectral UAV footage, the authors gathered red, green, blue, red-edge, and near-infrared) over Ontario, Canada’s cornfields in the growth seasons of 2022 and 2023. A variety of vegetation indices (NDVI, GNDVI, NDRE, etc.) were created to act as input features for modeling once ground truth yield data was captured at harvest.

Random Forest, Gradient Boosting Machines (GBM), and XGBoost were the tested models. XGBoost performed the best, outperforming RF (R2 = 0.88) and GBM (R2 = 0.87), with an R2 of 0.91 and an RMSE of 0.46 t/ha. Recursive feature elimination was used to choose the input characteristics, which included canopy temperature, solar radiation, and vegetation indices.

The authors used SHAP to analyze model predictions in order to address interpretability. The two most important factors in yield estimation, according to the worldwide SHAP summary, were canopy temperature and NDRE (Normalized Difference Red Edge Index). Higher mid-season NDRE values, for instance, were favorably associated with yield, whereas higher canopy temperatures during the reproductive stages were associated with lower production as a result of heat stress. Targeted management decisions were made possible by the researchers’ ability to identify yield-affecting conditions at the plot level using local SHAP plots. For example, in several instances, concurrently high canopy temperature caused high NDVI values to result in unexpectedly low yield forecasts; this realization was made feasible by SHAP interaction effects.

To the best of our knowledge, while several research have used attention mechanisms in image analysis or regression models for predicting soil properties, the combination of attention mechanisms and regression models for multitemporal the mapping of soil properties using Sentinel-2 has never been documented in the literature before.

3. Materials and Methods

3.1. Description of the Data

The data used in this study were collected from a total of 45,000 different locations covering the entire area of the United States during 2023. The locations were randomly generated throughout the USA, and for each, data on soil properties and satellite images were collected. For each geolocation, values are available for 55 soil properties, as well as 12 monthly Sentinel-2 images.

To preserve quality and consistency, all locations that did not have complete data on all soil properties were excluded from the analysis. At this stage, no additional filters were applied, such as selection by land cover type or altitude.

Soil property data were obtained from the California Soil Resource Laboratory [25] at UC Davis University and are based on current USDA-NCSS data obtained from official field surveys. Properties are aggregated within square cells of 800 m².

Available soil properties are classified into three main categories:

• Chemical properties: calcium carbonate, cation exchange capacity (CEC), pH value, soil organic matter, sodium adsorption ratio, electrical conductivity
• Physical properties: bulk density, sand content, silt content, clay content, saturated hydraulic conductivity, rock fragments, soil texture
• Land use related properties: depth to restrictive layer, wind erodibility index, soil depth, hydrologic group, land capability class, wind erodibility group, soil order, soil temperature regime

The complete list of all available soil properties, with those used in this study (✓) specially marked, is given in Table 1.

For a complete description of each property, see Appendix A.

Satellite Data

For each location, 12 satellite images are provided – one for each month during 2023. Sentinel-2 (Level-1C) images were used, which enable multispectral analysis of the land surface with high spatial resolution. Each satellite image covers a square area of 400 × 400 meters (which corresponds to an image resolution of 40 × 40 pixels per location). For each image, all relevant spectral channels of the Sentinel-2 mission (13 bands in total) are available.

Note:

• Channels in the visible and near-infrared spectrum (B02–B08) are key for vegetation and soil status analysis.
• SWIR channels (B11 and B12) are useful for moisture detection and soil type classification.
• Red edge channels (B05, B06, B07) are especially valuable for monitoring vegetation stress.
• B09 and B10 are most often used for atmospheric correction but may contain useful information for specific tasks.

For a complete description of each band, see Appendix B.

3.2. Data Preprocessing

Dynamic augmentation during training. During training, for each location and in each iteration, samples are generated dynamically: from the original image of 40x40 pixels, 32x32 pixels are randomly selected, and the selected pixels are randomly permuted and arranged in a new image of the same dimensions. This augmentation method is applied in every batch during training, which increases data diversity and efficiently reduces overfitting.

Target data are normalized in the range 0–1.

3.3. Model Architecture

In this research, we used two deep learning models based on convolutional neural networks, with an added attention mechanism that helps the model “focus” on the most important input data. These models are trained to predict several different soil properties at the same time, using both satellite images and field measurements.

The first model, attention-merge, uses a single shared attention layer for all target properties. This means the model finds which satellite channels and months are most important for predicting all soil properties together.

The second model, attention-split, has a separate attention layer for each soil property. In this way, the model can learn which data is most important for each property individually, which makes the results more precise and easier to interpret.

Figure 1 shows both models: on the left is the model with shared attention for all properties, and on the right is the model that uses separate attention for each property.

In both architectures, the attention layer is implemented as a Dense layer with sigmoid activation that generates weights for all input channels and months, which are then applied to the input data by element-wise multiplication.

3.4. Model Training

Training was conducted in two successive phases with the aim of thoroughly examining the importance of input data for model performance.

Phase 1: In the first phase, each of the models (attention-merge and attention-split) was trained 100 times, using the Adam optimizer with a learning rate of 0.001, batch size of 64, for 128 epochs (with 32 batches per epoch). In each of the 100 experiments (trainings), the sample set was randomly split into training, validation, and test sets in a 60:20:20 ratio. Trainings were performed on a workstation with an Intel Core i5-13600 processor, 64 GB RAM, and NVIDIA RTX 3090 GPU. TensorFlow and Keras were used for implementation and experimentation.

After completion of the first phase, the attention layers were analyzed to identify input bands and time periods (months) that have the highest significance for the model in predicting soil properties. This analysis enables the detection of input data that "stand out" as most informative.

Phase 2: In the second phase, all previously identified input bands and months (those with the highest attention coefficients) were omitted from the new training process. The models were then retrained and evaluated with the same settings, to assess whether their exclusion causes a statistically significant deterioration in training quality, i.e., the ability to regress the output (measured) parameters.

Note: The aim of the experimental design is not to achieve maximum model performance but to isolate and confirm the importance of individual input data. The number of epochs and training size were chosen to enable as much experiment replication and reliable statistical estimation of importance as possible, even at the cost of slightly compromised maximum accuracy metrics. The key value of this approach is the confirmation that the exclusion of attention-identified bands indeed leads to a reduction in the model’s ability to learn the appropriate relations, thus confirming their key role in the model.

4. Results and Discussion

Figure 2 shows the training process for the attention-merge model across 100 repeated experiments. Blue lines represent training loss, and red lines represent validation loss for each experiment.

It can be seen that the training (validation) loss has not reached its minimum and there is still plenty of room for optimization. For now, it is important to state that in the planned 128 episodes, no overfitting occurred.

Figure 3 and Figure 4 show the Normalized Mean Absolute Error (NMAE) and R² metric for all observed properties. The graphs show results for both models (attention-merge and attention-split), obtained as the mean from 100 independent experiments with 128 epochs per experiment. For comparison, results are also shown when each model was trained only once for 2048 epochs, to assess the longer-term trend of performance.

Figure 5 shows the heatmap of the attention scores obtained from the attention layer of the attention-mergee model, showing the dependence between months and satellite bands based on the average from 100 repeated experiments.

It can be noticed that bands B09 (Water vapor, 945 nm) and B10 (Short-wave infrared, cirrus, 1375 nm) stand out the most, and to some extent B12 (Short-wave infrared, 2190 nm) and B01 (Coastal aerosol, 443 nm). The dominance of these channels can be explained by their sensitivity to the physical-chemical properties of the soil and the presence of moisture. B09 (water vapor) and B10/B12 (shortwave infrared region) are particularly significant because the SWIR range provides important information on soil moisture content, texture, the presence of mineral matter, and organic components. B10 (cirrus) is extremely sensitive to water vapor and moisture, which may be directly related to agricultural potential and soil condition. B01 (coastal aerosol) can contribute to the detection of surface particles, dust, and other aerosols, which indirectly affects surface soil properties.

In Figure 6(1/2),(2/2), heatmaps are shown for the attention layer for each soil property individually for the attention-split model. In most maps, bands B09 and B10 are still the most pronounced, while B11 and B12 are somewhat less represented. For certain properties, such as electrical conductivity and sodium adsorption ratio, RGB channels also play a significant role, especially during the middle of the calendar year.

In the second phase, during the training of the attention-merge model, channels B09 and B10 were excluded from the input data, and the entire training procedure was repeated under the same conditions as in the previous experiments. Figure 7 shows comparative results for Normalized Mean Absolute Error (NMAE) values, trained with all channels and without channels B09 and B10, based on the mean from 100 repeated experiments. For each soil property, the 95% confidence interval is also shown, allowing a clearer assessment of the importance of individual channels for model performance.

Looking at the chart 7, which shows results for the attention-merge model, it is clear that for 11 of the total 17 shown soil properties, there is no overlap of the confidence intervals between the model trained with all channels and the model trained without channels B09 and B10. This further confirms the findings from the previous heatmap and indicates the importance of these channels for model accuracy. Interestingly, in the setting where one attention layer is used for all soil properties at the same time, excluding channels B09 and B10 leads to improvement only in one property (sodium adsorption ratio), while for the other properties it either worsens performance or a reliable conclusion cannot be drawn from the shown data.

5. Conclusions

In this research, we analyzed which spectral channels from Sentinel-2 satellite images and in which months contribute the most to the estimation of soil properties using deep neural networks. Two architectures were tested: a model with a shared attention layer (attention-merge) and a model with separate attention layers for each target (attention-split) variable individually. The results showed, somewhat surprisingly, that the greatest influence among the input data is from bands B09 (Water vapor, 945 nm) and B10 (Short-wave infrared, cirrus, 1375 nm). Regarding seasonal dependence, it is not possible to unequivocally highlight a certain month, but it can be observed that satellite images from the middle of the calendar year are generally more important for the prediction of soil properties.

To confirm the findings obtained from the analysis of attention layer values, attention-merge model was further trained after excluding the most significant channels (B09 and B10). The results achieved show that such exclusion significantly reduces model accuracy, confirming the key role of these channels in the prediction process.

To draw final conclusions, further research is needed, including systematic removal of other, less significant channels, to assess their contribution and enable further optimization of the amount and structure of input data. This would create the basis for even more efficient models, adapted to limited resources or application to other satellites and sensors.