Soil Classification Datasets

Please ensure to cite the paper when utilizing the dataset in a research study. Refer to the paper link or BibTeX provided below.

This repository contains comprehensive datasets for soil classification and recognition research. The Original Dataset comprises soil images sourced from various online repositories, which have been meticulously cleaned and preprocessed to ensure data quality and consistency. To enhance the dataset's size and diversity, we employed Generative Adversarial Networks (GANs), specifically the CycleGAN architecture, to generate synthetic soil images. This augmented collection is referred to as the CyAUG Dataset. Both datasets are specifically designed to advance research in soil classification and recognition using state-of-the-art deep learning methodologies.

This dataset was curated as part of the research study titled "An advanced artificial intelligence framework integrating ensembled convolutional neural networks and Vision Transformers for precise soil classification with adaptive fuzzy logic-based crop recommendations" by Farhan Sheth, Priya Mathur, Amit Kumar Gupta, and Sandeep Chaurasia, published in Engineering Applications of Artificial Intelligence.

Links

• Paper: An advanced artificial intelligence framework integrating ensembled convolutional neural networks and Vision Transformers for precise soil classification with adaptive fuzzy logic-based crop recommendations
• Code: Development Code (model and recommendation system)

Application produced by this research is available at:

• Agro Companion: Agro Companion
• Server Code: Agro Companion Server

Note: If you are using any part of this project; dataset, code, application, then please cite the work as mentioned in the Citation section below.

Dataset

Both dataset consists of images of 7 different soil types.

The Soil Classification Dataset is structured to facilitate the classification of various soil types based on images. The dataset includes images of the following soil types:

• Alluvial Soil
• Black Soil
• Laterite Soil
• Red Soil
• Yellow Soil
• Arid Soil
• Mountain Soil

The dataset is organized into folders, each named after a specific soil type, containing images of that soil type. The images vary in resolution and quality, providing a diverse set of examples for training and testing classification models.

Original Dataset Details

• Total Images: 1189 images
• Image Format: JPG/JPEG
• Image Size: Varies
• Source: Collected from various online repositories and cleaned for consistency.

CyAUG Dataset Details

• Total Images: 5097 images
• Image Format: JPG/JPEG
• Image Size: Varies
• Source: Generated using CycleGAN to augment the original dataset, enhancing its size and diversity.

Input and Output Parameters

• Input Parameters:
  • Image: The images of the soils (JPG/JPEG format).
  • Label: The labels are in the format 'soil types' (folder names).
• Output Parameter:
  • Classification: The predicted class (soil type) based on the input image.

Citation

If you are using any of the derived dataset, please cite the following paper:

@article{SHETH2025111425,
  title = {An advanced artificial intelligence framework integrating ensembled convolutional neural networks and Vision Transformers for precise soil classification with adaptive fuzzy logic-based crop recommendations},
  journal = {Engineering Applications of Artificial Intelligence},
  volume = {158},
  pages = {111425},
  year = {2025},
  issn = {0952-1976},
  doi = {https://doi.org/10.1016/j.engappai.2025.111425},
  url = {https://www.sciencedirect.com/science/article/pii/S0952197625014277},
  author = {Farhan Sheth and Priya Mathur and Amit Kumar Gupta and Sandeep Chaurasia},
  keywords = {Soil classification, Crop recommendation, Vision transformers, Convolutional neural network, Transfer learning, Fuzzy logic}
}

This dataset is a digital soil survey and generally is the most detailed level of soil geographic data developed by the National Cooperative Soil Survey. The information was prepared by digitizing maps, by compiling information onto a planimetric correct base and digitizing, or by revising digitized maps using remotely sensed and other information.

This dataset consists of georeferenced digital map data and computerized attribute data. The map data are in a soil survey area extent format and include a detailed, field verified inventory of soils and miscellaneous areas that normally occur in a repeatable pattern on the landscape and that can be cartographically shown at the scale mapped. A special soil features layer (point and line features) is optional. This layer displays the location of features too small to delineate at the mapping scale, but they are large enough and contrasting enough to significantly influence use and management. The soil map units are linked to attributes in the National Soil Information System relational database, which gives the proportionate extent of the component soils and their properties.

SSURGO depicts information about the kinds and distribution of soils on the landscape. The soil map and data used in the SSURGO product were prepared by soil scientists as part of the National Cooperative Soil Survey.

This data set provides soil maps for the United States (US) (including Alaska), Canada, Mexico, and a part of Guatemala. The map information content includes maximum soil depth and eight soil attributes including sand, silt, and clay content, gravel content, organic carbon content, pH, cation exchange capacity, and bulk density for the topsoil layer (0-30 cm) and the subsoil layer (30-100 cm). The spatial resolution is 0.25 degree. The Unified North American Soil Map (UNASM) combined information from the state-of-the-art US General Soil Map (STATSGO2) and Soil Landscape of Canada (SLCs) databases. The area not covered by these data sets was filled by using the Harmonized World Soil Database version 1.21 (HWSD1.21). The Northern Circumpolar Soil Carbon (NCSCD) database was used to provide more accurate and up-to-date soil organic carbon information for the high-latitude permafrost region and was combined with soil organic carbon content derived from the UNASM (Liu et al., 2013). The UNASM data were utilized in the North American Carbon Program (NACP) Multi-Scale Synthesis and Terrestrial Model Intercomparison Project (MsTMIP) as model input driver data (Huntzinger et al., 2013). The driver data were used by 22 terrestrial biosphere models to run baseline and sensitivity simulations. The compilation of these data was facilitated by the NACP Modeling and Synthesis Thematic Data Center (MAST-DC). MAST-DC was a component of the NACP (www.nacarbon.org) designed to support NACP by providing data products and data management services needed for modeling and synthesis activities.

Context

https://storage.googleapis.com/kagglesdsdata/datasets/1262694/2104731/GridMaps250m_Info.png?X-Goog-Algorithm=GOOG4-RSA-SHA256&X-Goog-Credential=databundle-worker-v2%40kaggle-161607.iam.gserviceaccount.com%2F20210410%2Fauto%2Fstorage%2Fgoog4_request&X-Goog-Date=20210410T121915Z&X-Goog-Expires=172799&X-Goog-SignedHeaders=host&X-Goog-Signature=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" alt="IMG">

Content

Maps of clay, silt and sand contents (g kg-1) were predicted at 0-20 cm, 20-60 cm and 60-100 cm depths intervals by random forest regression in Google Earth Engine. Gridded soil information covers a part of the Midwest Brazil, from 12° S to 20° S and from 45° W to 54° W, and is available with 250m resolution. The maps were cross-validated and had Coefficient of Determination ranging from 0.64 to 0.85 at all depth intervals.

Acknowledgements

Poppiel, Raúl Roberto; Lacerda, Marilusa Pinto Coelho; Safanelli, José Lucas; Rizzo, Rodnei; Pereira de Oliveira Junior, Manuel; Novais, Jean Jesus; Dematte, Jose Alexandre (2020), “250 m-gridded soil texture at multiple depths of Midwest Brazil”, Mendeley Data, V4, doi: 10.17632/52cfcm3xr7.4

Photo by Clay Banks on Unsplash

The dataset consists of three raster GeoTIFF files describing the following soil properties in the US: available water capacity, field capacity, and soil porosity. The input data were obtained from the gridded National Soil Survey Geographic (gNATSGO) Database and the Gridded Soil Survey Geographic (gSSURGO) Database with Soil Data Development tools provided by the Natural Resources Conservation Service. The soil characteristics derived from the databases were Available Water Capacity (AWC), Water Content (one-third bar) (WC), and Bulk Density (one-third bar) (BD) aggregated as weighted average values in the upper 1 m of soil. AWC and WC layers were converted to mm/m to express respectively available water capacity and field capacity in 1 m of soil, and BD layer was used to produce soil porosity raster assuming that the average particle density of soils is equal to 2.65 g/cm3. For each soil property, soil maps with CONUS, Alaska, and Hawaii geographic coverages were derived from s ...

These Soil Mapping Data Packages include 1. a Soil Map dataset which includes the equivalents to Soil Project Boundaries, Soil Survey Spatial View mapping polygons with attributes from the Soil Name and Layer Files, plus + A Soil Site dataset which includes soil pit site information and detailed soil pit descriptions and any associated lab analyses, and + The Soil Data Dictionary which documents the fields and allowable codes within the data. The Soil Map geodatabase contains the 'best available' data ranging from 1:20,000 scale to 1:250,000 scale with overlapping data removed. The choice of the datasets that remain is based on connectivity to the soil attributes (soil name and layer files), map scale and survey date. (Note: the BC Soil Landscapes of Canada (BCSLC) 1:1,000,000 data has not been included in the Soil_Map or SIFT, but is available from: CANSIS. (A complete soils data package with overlapping soil survey mapping and BCSLC is available on request. Note that the soil survey data with attributes can also be viewed interactively in the [Soil Information Finder Tool](The Soil Map dataset is also available for interactive map viewing or as KMZs from the Soil Information Finder Tool website.

This dataset contains high-quality images of soil surfaces categorized into three moisture levels—Wet, Moderate, and Dry—captured under natural outdoor lighting in Sirajganj, Bangladesh. Images were collected at seven time intervals (0 min, 30 min, 1 hr, 2 hr, 4 hr, 5 hr, and 7+ hr after saturation) using a Sony Xperia 1 Mark II smartphone. A total of 1,177 raw images were captured, with blurry, noisy, and low-quality photos removed during pre-processing. The dataset reflects real-world agricultural conditions and serves as a benchmark for training machine learning and deep learning models for non-invasive soil moisture classification. Subject Areas: Computer Science, Agriculture Science, AI, Computer Vision, Environmental Monitoring, Pattern Recognition Data Format: JPG images (raw and filtered) Data Collection: Captured using Sony Xperia 1 Mark II under natural outdoor lighting in multiple soil locations. Organized into three labeled categories (Wet, Moderate, Dry) based on time intervals after saturation. Can be split into training and testing sets (recommended 80:20 ratio). Usage Notes: Ideal for developing AI models in soil moisture classification, precision irrigation scheduling, and image-based environmental monitoring. Supports affordable, sensor-free soil analysis for sustainable farming practices, particularly in resource-limited settings.

This is a public compendium of global, regional, national and sub-national soil samples and/or soil profile datasets (points with Observations and Measurements of soil properties and characteristics). Datasets listed here, assuming compatible open license, are afterwards imported into the Global compilation of soil chemical and physical properties and soil classes and eventually used to create a better open soil information across countries. Please feel free to contribute entries. See GitHub repository for more detailed instructions.

To see the most up-to-date version of this compendium please visit: https://soildb.OpenLandMap.org/

To read about the import steps and quality control, refer to the Hengl et al. (2026; ESSD).

This data set provides gridded data for selected soil parameters derived from data and methods developed by the Global Soil Data Task, an international collaborative project with the objective of making accurate and appropriate data relating to soil properties accessible to the global change research community. The task was coordinated by the International Geosphere-Biosphere Programme (IGBP-DIS). The data in this data set were produced by the International Satellite Land-Surface Climatology Project, Initiative II (ISLSCP II) staff from data obtained from the Oak Ridge National Laboratory Distributed Active Archive Center (ORNL DAAC, http://daac.ornl.gov/). See the related data sets section below. Two-dimensional gridded maps of selected soil parameters, including soil texture, at a 1.0 by 1.0 degree spatial resolution and for two soil depths are provided. All data layers have been adjusted to match the ISLSCP II land/water mask. There are 36 data files with this data set.

Soil texture classes (USDA system) for 6 soil depths (0, 10, 30, 60, 100 and 200 cm) at 250 m Derived from predicted soil texture fractions using the soiltexture package in R. Processing steps are described in detail here. Antarctica is not included. To access and visualize maps outside of Earth Engine, use this page. If you discover a bug, artifact or inconsistency in the LandGIS maps or if you have a question please use the following channels: Technical issues and questions about the code General questions and comments

The Harmonized World Soil Database version 2.0 (HWSD v2.0) is a unique global soil inventory providing information on the morphological, chemical and physical properties of soils at approximately 1 km resolution. Its main objective is to serve as a basis for prospective studies on agro-ecological zoning, food security and climate change.

The Harmonized World Soil Database (HWSD) was established in 2008 by the International Institute for Applied Systems Analysis (IIASA) and FAO, and in partnership with International Soil Reference and Information Centre (ISRIC), the European Soil Bureau Network (ESBN) and the Institute for Soil Sciences Chinese Academy of Sciences (CAS). The data entry and harmonization within a Geographic Information System (GIS) was carried out at IIASA, with verification of the database undertaken by all partners. HWSD was then updated in 2013 (HWSD v1.2) and in 2023 (HWSD v2.0).

This updated version (HWSD v2.0) is built on the previous versions of HWSD with several improvements on (i) the data source that now includes several national soil databases, (ii) an enhanced number of soil attributes available for seven soil depth layers, instead of two in HWSD v1.2, and (iii) a common soil reference for all soil units (FAO1990 and the World Reference Base for Soil Resources). This contributes to a further harmonization of the database.

The GIS raster image file is linked to the soil attribute database. The HWSD v2.0 soil attribute database provides information on the soil unit composition for each of the near 30 000 soil association mapping units. The HWSD v2.0 Viewer, provided with the database, creates this link automatically and provides direct access to the soil attribute data and the soil association information.

Note: - A tutorial for accessing HWSD ver. 2.0 using R (prepared by David Rossiter, June 2023) has been added as an 'associated resource' (NOTE: Needs the SQLite version of HWSD v2 as provided below). - Soil property estimates in HWSDv2 were derived from Batjes (2016), Geoderma (https://doi.org/10.1016/j.geoderma.2016.01.034).

This map depicts soils data from the USDA NRCS SSURGO dataset. The soil type is indicated in the MUSYM field. The data was downloaded from the NRCS website.The full Kansas geospatial catalog is administered by the Kansas Data Access & Support Center (DASC) and can be found at the following URL: https://hub.kansasgis.org/

The Southern Great Plains 1997 (SGP97) Hydrology Experiment originated from an interdisciplinary investigation, "Soil Moisture Mapping at Satellite Temporal and Spatial Scales" (PI: Thomas J. Jackson, USDA Agricultural Research Service, Beltsville, MD) selected under the NASA Research Announcement 95-MTPE-03. The core of the 1997 experiment involves the deployment of the L-band Electronically Scanned Thinned Array Radiometer (ESTAR) for daily mapping of surface soil moisture. The region selected for investigation is the best instrumented site for surface soil moisture, hydrology and meteorology in the world. This includes the USDA/ARS Little Washita Watershed, the USDA/ARS facility at El Reno, Oklahoma, the ARM/CART central facility, as well as the Oklahoma Mesonet. The temporal coverage for this dataset is as follows: Begin datetime: 1995-10-01 00:00:00, End datetime: 2001-03-31 23:59:59. The Department of Energy (DOE) Atmospheric Radiation Measurement (ARM) Southern Great Plains (SGP) Soil Texture Data Set is one of the various sub-surface data sets developed for the ARM/GCIP (Global Energy and Water Cycle Experiment (GEWEX) Continental-scale International Project) 1996 Near-Surface Observation (NESOB-96) Data Set. This data set contains a summary table of the percentages of sand, silt, and clay fractions in each soil layer at each of the ARM SWATS (Soil Water and Temperature System) sites at the SGP site. Also included is the corresponding USDA texture class as determined from the "soil triangle". The soil characterizations were perfomed by Oklahoma State University. Resources in this dataset:Resource Title: GeoData catalog record. File Name: Web Page, url: https://geodata.nal.usda.gov/geonetwork/srv/eng/catalog.search#/metadata/SGP97armTexture_JJM_2015-04-23_1409

This dataset contains hydrological and geotechnical soil properties of the different soil layers.

The International Soil Reference and Information Centre-World Inventory of Soil Emission Potentials (ISRIC-WISE) international soil profile data set consists of a homogenized, global set of 1,125 soil profiles for use by global modelers. These profiles provided the basis for the Global Pedon Database (GPDB) of the International Geosphere-Biosphere Programme (IGBP) - Data and Information System (DIS). The data set consists of a selection of 665 profiles originating from the Natural Resources Conservation Service (NRCS, Lincoln), 250 profiles obtained from the Food and Agriculture Organization (FAO, Rome), and 210 profiles from the reference collection of the International Soil Reference and Information Centre (ISRIC, Wageningen). All profiles are georeferenced and classified according to the 1974 Legend of the FAO-UNESCO Soil Map (FAC-UNESCO, 1974) of the World, as well as the 1988 Revised Legend of FAO-UNESCO (FAO, 1990). The data set includes information on soil classification, site data, soil horizon data, source of data, and methods used for determining analytical data. The data files are in a comma-delimited format. Data Citation: The data set should be cited as follows: Batjes, N. H. (ed). 2000. Global Soil Profile Data (ISRIC-WISE). Available on-line from the ORNL Distributed Active Archive Center, Oak Ridge National Laboratory, Oak Ridge, Tennessee, U.S.A.

The Food and Agriculture Organization of the United Nations (FAO) and the International Institute for Applied Systems Analysis (IIASA) combined the available regional and national soil information with the data already contained within the 1:5,000,000 scale FAO-UNESCO map, into a new comprehensive Harmonized World Soil Database (HWSD_v121). This map has a resolution of about 1 km (30 arc seconds) and consists of a 30-cm topsoil layer, and a 70-cm subsoil layer. The soil variables provided in the Harmonized World Soil Database (2009) and FAO/UNESCO Soil Map of the World included soil texture (%sand, %silt, %clay), organic carbon, pH, and EC. However, from a hydrological point of view, we are in need of parameters such as bulk density, water storage capacity, and hydraulic conductivity for different soil layers. Hence, we have used various pedotransfer functions from the literature to estimate the soil parameters needed in a Soil and Water Assessment Tool (SWAT model). The associated SWAT2012.mdb and lookup table is available at 2w2e GmbH website. […]

A global data set of soil types is available at 0.5-degree latitude by 0.5-degree longitude resolution. There are 106 soil units, based on Zobler?s (1986) assessment of the FAO/UNESCO Soil Map of the World. This data set is a conversion of the Zobler 1-degree resolution version to a 0.5-degree resolution. The resolution of the data set was not actually increased. Rather, the 1-degree squares were divided into four 0.5-degree squares with the necessary adjustment of continental boundaries and islands. The computer code used to convert the original 1-degree data to 0.5-degree is provided as a companion file. A JPG image of the data is provided in this document. The Zobler data (1-degree resolution) as distributed by Webb et al. (1993) [http://www.ngdc.noaa.gov/seg/eco/cdroms/gedii_a/datasets/a12/wr.htm#top] contains two columns, one column for continent and one column for soil type. The Soil Map of the World consists of 9 maps that represent parts of the world. The texture data that Webb et al.(1993) provided allowed for the fact that a soil type in one part of the world may have different properties than the same soil in a different part of the world. This continent-specific information is retained in this 0.5-degree resolution data set, as well as the soil type information which is the second column. A code was written (one2half.c) to take the file CONTIZOB.LER distributed by Webb et al. (1993) [http://www.ngdc.noaa.gov/seg/eco/cdroms/gedii_a/datasets/a12/wr.htm#top] and simply divide the 1-degree cells into quarters. This code also reads in a land/water file (land.wave) that specifies the cells that are land at 0.5 degrees. The code checks for consistency between the newly quartered map and the land/water map to which the quartered map is to be registered. If there is a discrepancy between the two, an attempt was made to make the two consistent using the following logic. If the cell is supposed to be water, it is forced to be water. If it is supposed to be land but was resolved to water at 1 degree, the code looks at the surrounding 8 cells and picks the most frequent soil type and assigns it to the cell. If there are no surrounding land cells then it is kept as water in the hopes that on the next pass one or more of the surrounding cells might be converted from water to a soil type. The whole map is iterated 5 times. The remaining cells that should be land but couldn't be determined from surrounding cells (mostly islands that are resolved at 0.5 degree but not at 1 degree) are printed out with coordinate information. A temporary map is output with -9 indicating where data is required. This is repeated for the continent code in CONTIZOB.LER as well. A separate map of the temporary continent codes is produced with -9 indicating required data. A nearly identical code (one2half.c) does the same for the continent codes. The printout allows one to consult the printed versions of the soil map and look up the soil type with the largest coverage in the 0.5-degree cell. The program manfix.c then will go through the temporary map and prompt for input to correct both the soil codes and the continent codes for the map. This can be done manually or by preparing a file of changes (new_fix.dat) and redirecting stdin. A new complete version of the map is outputted. This is in the form of the original CONTIZOB.LER file (contizob.half) but four times larger. Original documentation and computer codes prepared by Post et al. (1996) are provided as companion files with this data set. Image of 106 global soil types available at 0.5-degree by 0.5-degree resolution. Additional documentation from Zobler?s assessment of FAO soil units is available from the NASA Center for Scientific Information.

The Multiclass Soil Image Dataset is a curated collection of soil surface images intended to support research and experimentation in soil type recognition and agricultural analysis. The dataset brings together visually distinct soil categories that commonly appear in agricultural fields and natural environments, providing a diverse visual representation of soil conditions.

This dataset contains 1,378 soil images, each corresponding to a specific soil category. The images capture variations in texture, color, granularity, and surface structure that are characteristic of different soil types. Such diversity makes the dataset valuable for studying soil variability and supporting image-based analysis in agricultural and environmental applications.

The dataset is suitable for academic research, educational purposes, and experimental evaluation related to soil analysis, precision agriculture, land management, and environmental monitoring. By offering a balanced representation of multiple soil classes, it enables comparative studies and benchmarking of image-based soil recognition approaches without embedding assumptions about specific algorithms or modeling techniques.

Alluvial Soil – Images representing fertile soils commonly found along river plains and floodplains.

Black Soil – Images depicting dark-colored soils rich in organic matter and moisture retention.

Cinder Soil – Images showing coarse, granular soils formed from volcanic or cinder-like materials.

Clay Soil – Images illustrating fine-grained soils with dense texture and compact structure.

Laterite Soil – Images representing iron-rich soils typically found in tropical and subtropical regions.

Peat Soil – Images depicting organic-rich soils formed from partially decomposed plant material.

Red Soil – Images showing soils characterized by reddish coloration due to iron content.

Yellow Soil – Images representing soils with yellowish tones commonly observed in humid regions.

One of the obstacles in applying advanced crop simulation models such as DSSAT at a grid-based platform is the lack of gridded soil input data at various resolutions. Recently, there has been many efforts in scientific communities to develop spatially continuous soil database across the globe. The most representative example is the SoilGrids 1km released by ISRIC in 2014. In addition recent AfSIS project put a lot of efforts to develop more accurate soil database in Africa at high spatial resolution. Taking advantage of those two available high resolution soil databases (SoilGrids 1km and ISRIC-AfSIS at 1km resolution), this project aims to develop a set of DSSAT compatible soil profiles on 5 arc-minute grid (which is HarvestChoice’s standard grid). Six soil properties (bulk density, organic carbon, percentage of clay and silt, soil pH and cation exchange capacity) available from the original SoilGrids 1km or ISRIC-AfSIS were directly used as DSSAT inputs. We applied a pedo-transfer function to derive some soil hydraulic properties (saturated hydraulic conductivity, soil water content at field capacity, wilting point and saturation) which are critical to simulate crop growth. For other required variables, HarvestChoice’s HC27 database are used as a reference. Final outputs are provided in *.SOL file format (DSSAT soil database) for each country at 5-min resolution. In addition, uncertainty maps for organic carbon and soil water content at wilting points at the top 15 cm soil layers were generated to provide brief idea about accuracy of the final products. The generated soil properties were evaluated by visualizing their global maps and by comparing them with IIASA-IFPRI cropland map and AfSIS-GYGA’s available water content maps.