We introduce HyBEAR : a comprehensive benchmark for bare soil detection in hyperspectral imagery. It is composed of the following pivotal components:

Data. We release a large-scale (1954 hyperspectral patches, with approx. 108 million pixels), high-resolution, spatially-heterogeneous dataset of hyperspectral image patches accompanied by precise ground-truth delineations of bare soil areas. The HyBEAR dataset is available at the following link: 10.5281/zenodo.17607897 .

The HyBEAR dataset, together with the procedures of (i) generating ground-truth delineations and (ii) extracting the image patches, as well as the cross-validation settings and quality metrics, is discussed in Sect. . In Sect. , we report and discuss the experimental results constituting the baseline results of the HyBEAR benchmark. Section  concludes the paper.

Traditional methods for estimating soil parameters often rely on field-point sampling, involving the localized collection of soil or crop samples for subsequent laboratory analysis . However, the inherent limitation of these in-situ methods, which typically focus on a few selected locations, hinders the ability to map the spatial distribution of key indicators . Consequently, considerable research efforts have been recently directed toward establishing correlations between data acquired through field methods (serving as the ground-truth information) and data obtained from satellite imagery  or aerial platforms such as manned aircraft and drones . Recent reviews highlight the limitations of traditional soil sampling in capturing spatial variability. Additionally, they are extremely difficult to scale to large regions, and thus developing the remote sensing techniques for high-resolution soil property mapping is the direct remedy to this challenge . Drone-borne hyperspectral imagery, for instance, enables high-resolution mapping of crucial soil nutrients . Similarly, airborne and satellite imagery can be effectively used to estimate various soil parameters at a (potentially global) scale, offering inherent spatial scalability .

Remote Sensing (RS) has emerged as a powerful tool for the identification of bare soil , which is a prerequisite for estimating soil parameters and crop monitoring, owing to its capacity to assess extensive areas repeatedly over time . In the agricultural context, the methods involving the acquisition of Multispectral Imaging (MSI)  and Hyperspectral Imaging (HSI)  Images are prevalent. The multispectral data facilitate the straightforward estimation of chlorophyll content using green vegetation indices . However, the broad bandwidth of multispectral imaging can limit the accuracy of early detection of subtle negative symptoms in PA, including nutrient deficiencies and plant diseases . In contrast, HSIs, with their high spectral resolution (narrow and contiguous bands), enable the acquisition of finer details in the spectral response of a given area. HSI-based techniques can potentially detect various anomalies earlier than MSI due to the richer spectral information contained within their narrower bands .

Following the removal of images obscured by clouds to ensure data quality , the accurate identification of bare soil areas becomes a crucial subsequent step in the RS data processing chain for agricultural applications . Therefore, isolating the spectral response directly from the soil surface is a critical step in the data processing chain for various agricultural and environmental applications. By accurately detecting and potentially masking out non-bare-soil pixels, researchers can enhance the reliability and accuracy of subsequent analyses aimed at estimating crucial soil properties, including the moisture content, nutrient levels, organic matter, and texture . Furthermore, the identification of bare soil is also essential for monitoring agricultural practices, such as tillage, tracking fallow land, and assessing soil erosion risks . Additionally, pruning non-soil areas can play a key role in on-board processing, where MSIs/HSIs are analyzed on edge devices, e.g., satellites. In this scenario, removing the parts of the image that do not contain the objects of interest (here, bare soil areas) will substantially accelerate the entire analysis process and make it more memory-efficient . Therefore, bare soil detection can be considered a “smart data compression” step, in which areas of interest are determined to guide further analysis in a data-driven manner (note that extracting soil parameters from non-soil pixels would obviously lead to noisy, inherently incorrect estimates).

Detecting bare soil areas is approached through various methodologies, categorized as (i) green vegetation index-based methods and (ii) pixel-level machine learning classification techniques. The first group of algorithms applies vegetation indices originally designed for broader MSI bands, with the Normalized Difference Vegetation Index (NDVI) being the most widely recognized . Recent studies, such as the Intuition-1 satellite mission (KP Labs, Gliwice, Poland), explore in-orbit bare-soil detection using spectral vegetation indices derived from hyperspectral imagery . By applying a threshold to the NDVI values, it is possible to delineate areas with dense green vegetation, as well as regions with medium or no vegetation . In this context, areas lacking vegetation include bare soil, water bodies, and infrastructure. The filtering of these non-bare soil objects is often achieved using various spectral indices, such as the Visible, Green-Based Built-up Index, also referred to as the Normalized Difference Built-up Index (NDBI), or Modified Normalized Difference Water Index (MNDWI) for water bodies . Other examples of green vegetation indices used to estimate crop volume are the Enhanced Vegetation Index (EVI) and the Optimized Soil Adjusted Vegetation Index (OSAVI) . Dedicated bare-soil indices, e.g., the Bare Soil Index (BSI) , have also been investigated in the literature.

The pixel-level classification techniques frequently implement Machine Learning (ML) algorithms, with Random Forests (RFs) being a popular choice , along with its variations, such as the Guided Regularized Random Forest (GRRF) . Deep learning algorithms have been widely investigated and have indeed established the state of the art in a multitude of fields. Some studies have explored deep learning models of various architectures, such as detection models, Capsule Networks (CapsNets), and semantic segmentation models , as well as the aggregation of results from different methods using ensemble techniques , for bare soil identification. It is worth emphasizing that some deep learning methods, e.g., convolutional neural networks (CNNs), exploit contextual information while inferring a pixel-level prediction – this is in contrast to the pure pixel-level algorithms. Recent advancements in this area include the application of quantum-kernel support vector machines for the detection of bare soil in hyperspectral imagery . To enhance the accuracy of bare-soil pixel detection and mitigate potential error sources, researchers have used specific spectral indices as features. For instance, a mask for distinguishing bare soil pixels in raster data has been effectively created using NDVI and the Cellulose Absorption Index (CAI) .

Building on these established methodologies, recent research continues to advance bare-soil detection. There is a growing interest in leveraging advanced deep learning architectures to improve the accuracy and robustness of these techniques. For example, introduced a novel Hybrid Attention Network (HA-Net) designed explicitly for bare soil extraction from optical RS images. Furthermore, a review provides a comprehensive overview of the satellite RS techniques employed for identifying bare soil, discussing the latest advancements, limitations, and challenges associated with various methodologies . The significance of high-quality, diverse datasets remains a central theme, and the broader RS community recognizes the critical need for large-scale annotated datasets to train and evaluate the algorithms effectively .

The literature review reveals a critical gap in the availability of standardized, publicly accessible benchmark datasets specifically tailored for bare-soil detection using hyperspectral imagery . Notably, to our knowledge, there are no datasets that would include the entire fields of bare soil and would allow the community to develop and validate the algorithms for identifying such fields that are free of vegetation. Indeed, the existing methods and collections focus on pixel-level annotations (thus supporting the building of pixel-level classification algorithms), which can be misleading for certain precision agriculture processes, especially those that relate to the whole-field procedures (e.g., fertilization). In this article, we address this research and development gap.

The HSIs were acquired by QZ Solutions, a company based in Poland and focused on making farming more productive and sustainable with new technologies

For more details concerning QZ Solutions, we refer to the official webpage of the company: https://qzsolutions.eu/ (last access: 23 January 2026).

Data were collected for two areas: the first one depicted with the (i) P1 orthophoto map covering 7637 [ha], that is 19 092 581 hyperspectral pixels, and the second one, referred to as (ii) P2, for which we provide imagery of 35 588 [ha] (88 972 010 pixels). These areas are presented in Fig. . The maps match two different locations in Poland: Lower Silesian Voivodeship (P1), where the map covers the fields in the vicinity of the village of Przeworno, and Opolskie Voivodeship (P2), for which the map reveals hundreds of hilly fields in the area south of the town of Głubczyce. Both locations are more than 60 km apart, and the images were acquired within an hour of each other. Thus, due to the dynamic position of the sun and clouds, the lightning conditions differed, and the HyBEAR dataset is heterogeneous at the spatial and image acquisition levels, hence may be used to quantify the generalization abilities of ML algorithms. Figure  displays the significant difference between patches – compare images (a) and (b) with (c) and (d) to see how they differ in terms of reflectance levels. Finally, the collected data may be used for two purposes: (i) for a regression task aimed at estimating soil component levels, which was the goal of the HYPERVIEW challenge , and (ii) for bare soil detection, which conditions the accuracy of the estimated component levels. In this paper, we focus exclusively on the latter task.

High-quality Ground Truth (GT) is paramount for the validity and utility of any benchmark dataset, particularly in remote sensing applications. Accurate GT labels are essential for training robust machine learning models and for objectively evaluating their performance. The preparation of GT for bare soil detection in HSIs poses unique challenges, including spectral overlap between bare soil and other non-vegetated surfaces (e.g., roads and buildings), as well as the potential for sparse or early-stage vegetation that may be difficult to discern.

In this benchmark, the GT was meticulously prepared through a combination of automated and manual interpretation techniques to ensure high accuracy and reliability. While preparing the manual bare-soil outlines, we considered information from vegetation indices, which are sensitive to chlorophyll and can effectively indicate the absence of mature vegetation. However, relying solely on vegetation indices has limitations. For instance, fresh or very sparse vegetation might not yield a strong enough signal in vegetation indices to be reliably excluded as bare soil. Conversely, non-vegetated areas such as roads, buildings, and artificial surfaces can exhibit low vegetation index values, potentially leading to misclassifications if only this information source is used. Additionally, RGB channels were insufficient for accurate bare soil mapping, primarily because the visual distinction between bare soil and other non-vegetated surfaces can be subtle and is dependent on flight altitude. Also, early or stressed vegetation might not be visually apparent in RGB imagery, leading to errors in GT.

In Fig. , several examples of patches are presented. We prepared two different views of the hyperspectral patches: the RGB composition of the bands, and the CIR (Color Infrared using the near-infrared band) image. As one can see, bare soil fields are not easily distinguishable using only RGB images (examples a through e). Relying solely on a CIR image can also be misleading (examples b and e). Objects present in the field of view, such as dirt roads or various shadows, can make identifying the precise borders of bare soil fields difficult, as seen in (c) and (d) of the attached figure (see the zoomed parts of the images and GT masks).

To overcome these limitations, a strategy combining the analysis of both vegetation indices and visual information was applied. Initial outlines of potential bare-soil areas were informed by vegetation index analysis, helping identify regions where vegetation was likely absent. Subsequently, these outlines were carefully reviewed and refined by human experts (with 10, 4, and 1 years of hands-on experience), who visually inspected the HSI data and potentially high-resolution RGB imagery (if available). This manual refinement step was crucial for accurately delineating bare soil areas and excluding non-vegetated areas, such as roads, buildings, and shadows, that vegetation indices might have flagged. This iterative process of leveraging automated tools and expert knowledge ensures high accuracy and consistency in the GT labels.

The careful preparation of the GT is a critical aspect that contributes to the remote sensing community. By providing a reliable and accurate reference for bare soil, HyBEAR enables researchers to develop and evaluate algorithms for HSI data analysis, ultimately advancing the field of bare soil detection.

To make the data analysis within HyBEAR more affordable and practical, the large source orthophotomaps were split into smaller image patches. The HyBEAR dataset is composed of 1954 square patches of size 250 × 250 that were extracted from two different source orthophotomaps (P1 and P2). We divided the images according to longitude into five folds (F0, F1, F2, F3, and F4) (Fig. ). Since the second map is roughly four times larger, we assigned one fold to P1 and four folds to the images extracted from P2 to evenly distribute the images among folds. The presence of NO-DATA pixels in certain patches (visible as black areas in Figs.  and ) is a technical consequence of the irregular geographical boundaries of the flight strips and the subsequent orthorectification process. These regions, which fall outside the actual flight coverage when source orthophotomaps are divided into a regular grid, are encoded with the specific value of -9999 to support their automated filtration. This encoding is a result of the data processing workflow rather than any sensor noise or equipment failure. For each of those patches, in the dataset folder HyBEAR/images/, we store two files: an HSI with 430 spectral bands (IMG_****_F*.TIFF) and the labels' file (GT_****_F*.TIFF), each with its metadata and geospatial information encoded. The name of the file is built using a consecutive file identifier (ID), and a fold number (e.g., IMG_0555_F1.TIFF). In addition to the full version of the HyBEAR dataset, we offer a lighter version of our collection, referred to as the MINI subset of HyBEAR. To create the MINI subset, we selected exactly 50 images from each fold using stratified random sampling (with a fixed random seed of 171) based on the proportion of soil pixels. This procedure ensures that the MINI collection maintains a balanced and reproducible representation of the geographical diversity found in the FULL dataset. The dataset statistics are summarized in Table .

It is worth noting that the primary reason we chose to use the TIFF file format for all images is that it allows us to include geolocation information for each patch. Also, it preserves the full image encoding precision and supports decent data compression. For HyBEAR, we selected a lossless deflate compression method that reduced the average patch size from approx. 215 MB per image to around 40–48 MB.

The resulting image collection was tested using two different environments: QGIS (versions 3.28 and 3.40) and a Python-based application (employing Rasterio, NumPy, and Scikit-Learn libraries – the configuration file of the Python environment, as well as the code we use for testing, is delivered with this dataset to support reproducibility).

The HyBEAR dataset consists of two independent aerial hyperspectral scenes, P1 and P2, which were carefully selected to enable a comprehensive and rigorous evaluation of algorithms dedicated to bare soil detection. One of the key aspects of HyBEAR is the evaluation protocol, based on the multi-fold cross-validation scenarios. To verify the generalization capability of the bare soil detection methods, as well as their robustness to the variability of the acquired data, we defined the experimental scenarios within a five-fold cross-validation framework:

Stage 1. Training an algorithm on the folds F1 to F4 (TF1:F4) → testing on the fold F0 (ΨF0).

This evaluation protocol has been designed with several key objectives in mind, relevant to the evaluation of algorithms in the context of real-world applications of bare soil detection from aerial data:

Evaluation of generalization to spatially independent areas. The hyperspectral scenes P1 (F0) and P2 (F1 : F4) represent data acquired from different geographical locations. Testing on one area after training on the other allows a direct assessment of the algorithm's ability to generalize knowledge to new, previously unseen areas. This is crucial, as in practice, models are rarely applied to exactly the same areas where they were trained.

To facilitate a fair comparison of different bare soil detection methods, the following commonly adopted classification and segmentation metrics will be used to quantitatively evaluate the models within the presented evaluation protocol – these metrics are presented in the next section.

To evaluate the performance of the emerging ML models, a series of commonly adopted classification metrics is used . This choice is motivated by the need for a comprehensive analysis of the models' ability to correctly identify bare soil pixels.

Accuracy (ACC) measures the overall percentage of correctly classified samples, and it is calculated as:1ACC=TP+TNTP+TN+FP+FN,where TP, TN, FP, and FN represent true positives (correctly classified bare soil image pixels), true negatives (correctly classified non-bare soil image pixels), false positives (the non-bare soil image pixels incorrectly classified as those containing bare soil), and false negatives (bare soil image pixels incorrectly classified as the non-bare soil image pixels). Accuracy represents a fundamental measure of classification effectiveness.

Collectively, these metrics are crucial to provide a comprehensive, objective, and standardized way to assess and compare the performance of various methods in the bare soil detection task.

The data is organized in the following directory structure:

HyBEAR/ – the main directory serving as the root for the entire dataset. It contains all the necessary files and subdirectories to organize the hyperspectral patches and their corresponding GTs for both P1 and P2.

The hyperspectral image patches (files from IMG_0000_F0.tiff to IMG_1953_F4.tiff) contain information about the light reflectance values in different spectral ranges for each pixel. The GT patches (files from GT_0000_F0.tiff to GT_1953_F4.tiff) store binary information and indicate which pixels in the hyperspectral patch belong to:

the SOIL class – encoded as (1),

The patches have fixed dimensions of 250 × 250 pixels and were generated using a grid-based method from the original hyperspectral scenes P1 and P2. The dataset (consisting of these patches) is available at 10.5281/zenodo.17607897 . The entire dataset comprises 1954 data tuples and has a total size of 96 GB (for the compressed files, with the average size of a single tuple being 48 MB).

In addition to the dataset, several machine learning models were evaluated under the introduced cross-validation regime. We trained models on the hyperspectral pixels extracted separately from each image, and we employed the standard machine learning model architectures: Logistic Regression (later denoted as LR), Linear models with L2 Regularization (L2), Adaptive Boosting (AB), Support Vector Machines with a linear kernel (SVM), Decision Trees (DT), and Random Forest (RF). All models operate on feature vectors of size 430, which are the values of all spectral bands captured within a particular pixel. The selected models enable the efficient processing of the extensive HyBEAR dataset, for which the number of pixels for an image could reach 62 500. Since the number of images in each fold varies, and as we exclude the background pixels that do not contain any data, the overall number of training samples varies accordingly. The ratio of “useful” pixels (excluding background) for each fold is reported in Table .

In the experiments, we exploit both versions of the HyBEAR benchmark, with all the images (FULL) and using only a subset of 50 images from each fold (MINI). Table  provides the aggregated metrics, together with their standard deviation. The detailed results of every fold are gathered in the appendix (Table ). Additionally, the results are depicted in Fig. . Here, in addition to the metrics computed for every fold (points on the dashed lines), we plotted the average metric levels (dotted lines). Following this visualization should allow for spotting the models that are more consistent across the folds, and therefore, would return a more robust and universal classifier.

The highest performance scores were obtained for the LR-based models for the majority of metrics, with the SVM-based models reaching nearly the same performance, both for the FULL and MINI versions of HyBEAR. In Fig. , we present three example patches with the GT labels and the corresponding predictions elaborated using the investigated machine learning models. Here, the red areas show the false-positive bare-soil detections, whereas the yellow areas correspond to the false-negative detections. The average accuracy for some models reached a decent level of 0.926–0.927 (for SVM and LR, respectively), and the F-score was almost 0.9 (0.898 for the best-performing LR model). When previewing the predictions for these models over the example test patches (Fig. ), we can observe that there is still significant room for improvement for most of the models. For some less challenging patches (i.e., IMG_0717_F1), the results are satisfactory, but for some others (i.e., IMG_0022_F0), most of the models perform poorly. The performance drop observed in Fold 0 is particularly noteworthy, as it serves as the primary benchmark for cross-location generalization within HyBEAR (Table A1 and Fig. ). While Folds 1–4 are derived from scene P2, Fold 0 is the only fold extracted from scene P1, located more than 60 km away. Despite the data being acquired on the same day, the dynamic position of the sun and clouds resulted in varying lighting conditions and reflectance levels between these geographic areas. This domain shift is the primary reason why non-linear models like Random Forest, AdaBoost, and Decision Trees – which likely overfitted the specific spectral characteristics of the larger P2 scene – showed significant degradation in sensitivity and F1-scores when applied to Fold 0. In contrast, simpler linear models (LR, SVM) proved to be more robust and universal in this cross-location scenario.

Furthermore, Fold 0 presents an inherently greater difficulty due to class imbalance, as it contains the lowest soil pixel ratio in the dataset (28.3 % for FULL and 25.6 % for MINI). Qualitative inspection of failure cases, such as IMG_0022 _F0 (Fig. ) reveals that errors are often caused by physical factors like tree shadows and dirt roads that might mimic bare soil, or crop residues that introduce spectral mixing. These results suggest that capturing more complex spatial relationships through, e.g., automated representation learning in deep learning models, may be necessary to further boost the capabilities of the detectors.

Despite the significant scale and high resolution of the HyBEAR benchmark, several limitations should be acknowledged. First, the dataset represents a regional and temporal snapshot, as it was acquired in Southern Poland on a single day in early spring. Consequently, it may not capture the full spectral variability of all global soil types or the influence of diverse seasonal and climatic conditions. Second, while the benchmark provides unique parcel-level annotations, our current baseline experiments are primarily pixel-wise. These results serve as a fundamental performance reference but do not yet fully exploit the contextual, field-level nature of the annotations for advanced parcel-based decision-making. Finally, although the evaluation protocol employs spatially disjoint scenes (P1 and P2), the benchmark primarily addresses regional cross-location generalization rather than global agroecological robustness. Future expansions of HyBEAR are intended to incorporate data from more diverse geographic zones and timeframes to address these constraints.

The HyBEAR benchmark is designed to support a wide range of practical applications in Earth observation:

Precision Agriculture. The provided bare soil masks enable targeted parcel-based fertilization planning and accurate soil property mapping, ensuring that analytical models for moisture or organic matter are applied only to relevant pixels.