The delineation of agricultural field boundaries has a wide range of applications, such as for crop management, precision agriculture, land use planning and crop insurance, etc. Manually digitizing agricultural fields from imagery is labor-intensive and time-consuming. This deep learning model automates the process of extracting agricultural field boundaries from satellite imagery, thereby significantly reducing the time and effort required. Its ability to adapt to varying crop types, geographical regions, and imaging conditions makes it suitable for large-scale operations.Using the modelFollow the guide to use the model. Before using this model, ensure that the supported deep learning libraries are installed. For more details, check Deep Learning Libraries Installer for ArcGIS.Fine-tuning the modelThis model can be fine-tuned using the Train Deep Learning Model tool. Follow the guide to fine-tune this model.InputSentinel-2 L2A 12-bands multispectral imagery using Bottom of Atmosphere (BOA) reflectance product in the form of a raster, mosaic or image service.OutputFeature class containing delineated agricultural fields.Applicable geographiesThe model is expected to work well in agricultural regions of USA.Model architectureThis model uses the Mask R-CNN model architecture implemented in ArcGIS API for Python.Accuracy metricsThis model has an average precision score of 0.64 for fields.Training dataThis model has been trained on an Esri proprietary agricultural field delineation dataset.LimitationsThis model works well only in areas having farmlands and may not give satisfactory results in areas near water bodies and hilly regions. The results of this pretrained model cannot be guaranteed against any other variation of the Sentinel-2 data.Sample resultsHere are a few results from the model.

Agricultural field polygons within smallholder farming systems are essential to facilitate the collection of geo-spatial data useful for farmers, managers, and policymakers. However, the limited availability of training labels poses a challenge in developing supervised methods to accurately delineate field boundaries using Earth Observation (EO) data. This data setallows researchers to test and benchmark machine learning methods to delineate agricultural field boundaries in polygon format. The large-scale data set consists of 439,001 field polygons divided into 62 tiles of approximately 5×5 km distributed across Vietnam and Cambodia, covering a range of fields and diverse landscape types. The field polygons have been meticulously digitized from satellite images, following a rigorous multi-step quality control process and topological consistency checks. Multi-temporal composites of Sentinel-2 (S2) images are provided to ensure cloud-free data. We anticipate that this large-scale data set will enable researchers to further enhance the delineation of agricultural fields in smallholder farms and support achieving the Sustainable Development Goals (SDG). Date Submitted: 2023-10-02

Satellite-based field delineation has rapidly evolved due to recent advances in machine learning for computer vision. However, the scarcity of labeled data for complex and dynamic smallholder landscapes remains a major bottleneck for operational field delineation and downstream applications, particularly in Sub-Saharan Africa. We here provide reference field boundaries collected in the scope of a research project funded by the F.R.S.-FNRS. A detailed description of the data can be found in our corresponding pre-print.

The dataset contains multiple files:

• sites.gpkg: Vector dataset containing the sampling sites covered by the dataset. We covered 513 sites of 600x600m, or 36 ha each.
• human_fields_train.gpkg: 1,518 field delineations distributed across 200 sites. Individual fields were manually digitized based on very high resolution satellite imagery in Google Earth Pro. Each polygon contains the corresponding image acquisition date. These fields were used for model training in the paper.
• human_fields_test.gpkg: 2,199 field delineations distributed across 313 sites. Individual fields were manually digitized based on very high resolution satellite imagery in Google Earth Pro. Each polygon contains the corresponding image acquisition date. These fields were used for evaluation of all experiments described in the paper.
• pseudo_fields_train.gpkg: 766 pseudo labels obtained from predictions using a pre-trained FracTAL ResUNet model. The pseudo-labels correspond to the selection using P99(SemCN). For other sets of pseudo-labels please contact us.

Brief description of methods

For site selection, we developed a stratified random sampling scheme to sample from regions with actively used cropland. To identify these regions, we used an existing map of active and fallow cropland for the growing season of September 2020 through August 2021 (Rufin et al., 2022). We aggregated the map to a 1 ha grid and calculated the proportions of active cropland. We sampled 1,000 sites from regions mapped as containing at least 50% of active cropland within a one-hectare grid cell. We defined a site extent of 600 by 600 meters, or 36 ha, in order to assure that a sufficient number of fields can be delineated, even in regions with comparatively large field size. The selected sites were screened for VHR image quality and acceptable visibility of at least five fields, resulting in 513 sample sites.

For human labels, we tasked human annotators to collect sparse labels (i.e. at least five fields) per site. We tasked the interpreters to collect only fields containing non-tree crops by systematically excluding tree crop plantations from our data. While individual trees in the field interior were included in our labels, trees overlapping with the field boundaries were avoided and the tree canopy was considered as the field boundary for completing the labels. All field delineations underwent an iterative quality assessment, where 18% of the initial field delineations and 7% of the field delineations in a second iteration were discarded.

The pseudo labels provided here were selected using the 1% most confident predictions from the pre-trained model. The confidence scores were computed as the median of all pixel-level field extent probabilities for each field instance.

For more details please read the corresponding paper:

Rufin, Wang, Lisboa, Hemmerling, Tulbure & Meyfroidt, P. (2023). Taking it further: Leveraging pseudo labels for field delineation across label-scarce smallholder regions. https://doi.org/10.48550/ARXIV.2312.08384

A vector dataset of Field Boundaries, automatically delineated from Sentinel-2 satellite imagery from May-June 2022.

Automatic field delineation refers to the process of automatically tracing the boundaries of agricultural parcels from satellite or aerial imagery. We consider an agricultural parcel as a spatially homogeneous land unit used for agricultural purposes, where a single crop is grown. The result of the FD is a set of closed vector polygons marking the extent of each agricultural parcel. Such polygons are the input to a multitude of applications, ranging from the management of agricultural resources, such as the Area Monitoring for the Common Agricultural Policy, to precision farming, to the estimation of damages to crop yield due to natural (e.g. drought, floods), and human-made disasters (e.g. war). Automatic estimation of parcels with high fidelity in a timely manner allows therefore to characterize the changes of agricultural landscapes due to anthropogenic activities, agricultural practices, and climate change consequences.

Additional links

Automatic Field Delineation Blog post

Dataset description on Fiboa

The Canadian major and minor crop field trial regions were developed following extensive stakeholder consultation and have been harmonized between the Pest Management Regulatory Agency (PMRA) and the Environmental Protection Agency of the USA. The Canadian major and minor crop field trial regions were delineated, using the geographic information system (GIS) data processing hardware and software facilities in Spatial Analysis and Geomatics Applications (SAGA), Agriculture Division, Statistics Canada. In general, the delineation process involved integration, evaluation and reference to numerous geographic data sources in a GIS to determine the best sources for the delineation. There are seven major and four minor field trial regions. Each of these regions recognizes physical characteristics, such as soils, and crops and climate, that make the region unique within the Canadian agricultural landscape. The subzones address differences within a region, generally reflected in the types of crops grown in that region. The Canadian regions, as much as possible, correspond to the U.S. regions

The Canadian major and minor crop field trial regions were developed following extensive stakeholder consultation and have been harmonized between the Pest Management Regulatory Agency (PMRA) and the Environmental Protection Agency of the USA. The Canadian major and minor crop field trial regions were delineated, using the geographic information system (GIS) data processing hardware and software facilities in Spatial Analysis and Geomatics Applications (SAGA), Agriculture Division, Statistics Canada. In general, the delineation process involved integration, evaluation and reference to numerous geographic data sources in a GIS to determine the best sources for the delineation. There are seven major and four minor field trial regions. Each of these regions recognizes physical characteristics, such as soils, and crops and climate, that make the region unique within the Canadian agricultural landscape. The subzones address differences within a region, generally reflected in the types of crops grown in that region. The Canadian regions, as much as possible, correspond to the U.S. regions

Soil is essential to human survival. We rely on it for the production of food, fibre, timber and energy crops. Together with climate, the soil determines which crops can be grown, where and how much they will yield. In addition to supporting our agricultural needs, we rely on the soil to regulate the flow of rainwater and to act as a filter for drinking water. With such a tremendously important role, it is imperative that we manage our soils for their long-term productivity, sustainability and health.

The first step in sustainable soil management is ensuring that the soil will support the land use activity. For example, only the better agricultural soils in Manitoba will support grain and vegetable production, while more marginal agricultural soils will support forage and pasture-based production. For this reason, agricultural development should only occur in areas where the soil resource will support the agricultural activity. The only way to do this is to understand the soil resource that is available. Soil survey information is the key to understanding the soil resource.

Soil survey is an inventory of the properties of the soil (such as texture, internal drainage, parent material, depth to groundwater, topography, degree of erosion, stoniness, pH and salinity) and their spatial distribution over a landscape. Soils are grouped into similar types and their boundaries are delineated on a map. Each soil type has a unique set of physical, chemical and mineralogical characteristics and has similar reactions to use and management. The information assembled in a soil survey can be used to predict or estimate the potentials and limitations of the soils’ behaviour under different uses. As such, soil surveys can be used to plan the development of new lands or to evaluate the conversion of land to new uses. Soil surveys also provide insight into the kind and intensity of land management that will be needed.

The survey scale of soils data for Manitoba ranges from 1:5,000 to 1:126,720, as identified in the 'SCALE' column.1:5,000. The survey objective at this scale is to collect high precision field scale data and it is mostly used in research plots and other highly intensive areas. It is also applicable to agricultural production and planning such as precision farming, agriculture capability, engineering, recreation, potato/irrigation suitability and productivity indices. Profile descriptions and samples are collected for all soils. At least one soil inspection exists per delineation and the minimum size delineation is 0.25 acres. The soil taxonomy is generally Phases of Soil Series. The mapping scale is 1:5,000 or 12.7 in/ mile.

This file also contains soils data that has been collected in Manitoba at a survey intensity level of the second order. This includes data collected at a scale of 1:20,000. The survey objective at this scale is to collect field scale data and it is mostly used in agricultural production and planning such as precision farming, agriculture capability, engineering, recreation, potato/irrigation suitability and productivity indices. Soil pits are generally about 200 metres apart and are dug along transects which are about 500 metres apart. This translates to about 32 inspections sites per section (640 acres). The soils in each delineation are identified by field observations and remotely sensed data. Boundaries are verified at closely spaced intervals. Profile descriptions are collected for all major named soils and 10 inspection sites/section and 2 to 3 horizons per site require lab analyses. At least one soil inspection exists in over 90% of delineations and the minimum size delineation is generally about 4 acres at 1:20,000. The soil taxonomy is generally Phases of Soil Series. The mapping scale is 1:20,000 or 3.2 inch/ mile.

This file also contains data that has been collected at the third order. This includes scales of 1:40,000 and 1:50,000. The survey objective at this scale is to collect field scale or regional data. If the topography is relatively uniform, appropriate interpretations include agriculture capability, engineering, recreation, potato/irrigation suitability and productivity indices. Soil pits are generally dug adjacent to section perimeters. This translates to about 16 inspection sites per section (640 acres). Soil boundaries are plotted by observation and remote sensed data. Profile descriptions exist for all major named soils and 2 inspection sites/section and 2 to 3 horizons per site require lab analyses. At least one soil inspection exists in 60-80% of delineations and the minimum size delineation is generally in the 10 to 20 acre range. The soil taxonomy is generally Series or Phases of Soil Series. The mapping scale is 1:40,000 or 2 inch/ mile; 1:50,000 or 1.5 inch/mile.

This file also contains soils data that has been collected at a survey intensity level of the fourth order. This includes scales of 1:63,360, 1:100,000, 1:125,000, and 1:126,720. The survey objective is to collect provincial data and to provide general soil information about land management and land use. The number of soil pits dug averaged to about 6 inspections per section (640 acres). Soil boundaries are plotted by interpretation of remotely sensed data and few inspections exist. Profile descriptions are collected for all major named soils. At least one soil inspection exists in 30-60% of delineations and the minimum size delineation is 40 acres (1:63,360), 100 acres (1:100,000), 156 acres (126,700) and 623 acres (250,000). The soil taxonomy is generally phases of Subgroup or Association.

As of 2022, soil survey field work and reports are still currently being collected in certain areas where detailed information does not exist. This file will be updated as more information becomes available. Typically, this is conducted on an rural municipality basis.

In some areas of Manitoba, more detailed and historical information exists than what is contained in this file. However, at this time, some of this information is only available in a hard copy format. This file will be updated as more of this information is transferred into a GIS format.

This file has an organizational framework similar to the original SoilAID digital files and a portion of this geographic extent was originally available on the Manitoba Land Initiative (MLI) website.

Domains and coded values have also been integrated into the geodatabase files. This allows the user to view attribute information in either an abbreviated or a more descriptive manner. Choosing to display the description of the coded values allows the user to view the expanded information associated with the attribute value (reducing the need to constantly refer to the descriptions within the metadata). To change these settings in ArcCatalog, go to Customize --> ArcCatalog Options --> Tables tab --> check or uncheck 'Display coded value domain and subtype descriptions'. To change these settings in ArcMap, go to Customize --> ArcMapOptions --> Tables tab --> check or uncheck 'Display coded value domain and subtype descriptions'. This setting can also be changed by opening the attribute table, then Table Options (top left) --> Appearance --> check or uncheck 'Display coded value domain and subtype descriptions'. The file also contains field aliases, which can also be turned on or off under Table Options.

The file - "Manitoba Municipal Boundaries" - from Manitoba Community Planning Services was used as one of the base administrative references for the soil polygon layer.

Also used as references were the hydrological features mapped in the 1:20,000 and 1:50,000 NTS topographical layers (National Topographic System of Canada). Typically this would relate to larger hydrological features such as those designated as perennial lakes and perennial rivers.

This same capability is available in ArcGIS Pro.

For more info:

https://www.gov.mb.ca/agriculture/soil/soil-survey/importance-of-soil-survey-mb.html#

This dataset contains Manitoba Agriculture soil survey data at various scales ranging from highly detailed to broader reconnaissance level information. Soil is essential to human survival. We rely on it for the production of food, fibre, timber and energy crops. Together with climate, the soil determines which crops can be grown, where and how much they will yield. In addition to supporting our agricultural needs, we rely on the soil to regulate the flow of rainwater and to act as a filter for drinking water. With such a tremendously important role, it is imperative that we manage our soils for their long-term productivity, sustainability and health. The first step in sustainable soil management is ensuring that the soil will support the land use activity. For example, only the better agricultural soils in Manitoba will support grain and vegetable production, while more marginal agricultural soils will support forage and pasture-based production. For this reason, agricultural development should only occur in areas where the soil resource will support the agricultural activity. The only way to do this is to understand the soil resource that is available. Soil survey information is the key to understanding the soil resource. Soil survey is an inventory of the properties of the soil (such as texture, internal drainage, parent material, depth to groundwater, topography, degree of erosion, stoniness, pH and salinity) and their spatial distribution over a landscape. Soils are grouped into similar types and their boundaries are delineated on a map. Each soil type has a unique set of physical, chemical and mineralogical characteristics and has similar reactions to use and management. The information assembled in a soil survey can be used to predict or estimate the potentials and limitations of the soils’ behaviour under different uses. As such, soil surveys can be used to plan the development of new lands or to evaluate the conversion of land to new uses. Soil surveys also provide insight into the kind and intensity of land management that will be needed. The survey scale of soils data for Manitoba ranges from 1:5,000 to 1:126,720, as identified in the 'SCALE' column.1:5,000. The survey objective at this scale is to collect high precision field scale data and it is mostly used in research plots and other highly intensive areas. It is also applicable to agricultural production and planning such as precision farming, agriculture capability, engineering, recreation, potato/irrigation suitability and productivity indices. Profile descriptions and samples are collected for all soils. At least one soil inspection exists per delineation and the minimum size delineation is 0.25 acres. The soil taxonomy is generally Phases of Soil Series. The mapping scale is 1:5,000 or 12.7 in/ mile. This file also contains soils data that has been collected in Manitoba at a survey intensity level of the second order. This includes data collected at a scale of 1:20,000. The survey objective at this scale is to collect field scale data and it is mostly used in agricultural production and planning such as precision farming, agriculture capability, engineering, recreation, potato/irrigation suitability and productivity indices. Soil pits are generally about 200 metres apart and are dug along transects which are about 500 metres apart. This translates to about 32 inspections sites per section (640 acres). The soils in each delineation are identified by field observations and remotely sensed data. Boundaries are verified at closely spaced intervals. Profile descriptions are collected for all major named soils and 10 inspection sites/section and 2 to 3 horizons per site require lab analyses. At least one soil inspection exists in over 90% of delineations and the minimum size delineation is generally about 4 acres at 1:20,000. The soil taxonomy is generally Phases of Soil Series. The mapping scale is 1:20,000 or 3.2 inch/ mile. This file also contains data that has been collected at the third order. This includes scales of 1:40,000 and 1:50,000. The survey objective at this scale is to collect field scale or regional data. If the topography is relatively uniform, appropriate interpretations include agriculture capability, engineering, recreation, potato/irrigation suitability and productivity indices. Soil pits are generally dug adjacent to section perimeters. This translates to about 16 inspection sites per section (640 acres). Soil boundaries are plotted by observation and remote sensed data. Profile descriptions exist for all major named soils and 2 inspection sites/section and 2 to 3 horizons per site require lab analyses. At least one soil inspection exists in 60-80% of delineations and the minimum size delineation is generally in the 10 to 20 acre range. The soil taxonomy is generally Series or Phases of Soil Series. The mapping scale is 1:40,000 or 2 inch/ mile; 1:50,000 or 1.5 inch/mile. This file also contains soils data that has been collected at a survey intensity level of the fourth order. This includes scales of 1:63,360, 1:100,000, 1:125,000, and 1:126,720. The survey objective is to collect provincial data and to provide general soil information about land management and land use. The number of soil pits dug averaged to about 6 inspections per section (640 acres). Soil boundaries are plotted by interpretation of remotely sensed data and few inspections exist. Profile descriptions are collected for all major named soils. At least one soil inspection exists in 30-60% of delineations and the minimum size delineation is 40 acres (1:63,360), 100 acres (1:100,000), 156 acres (126,700) and 623 acres (250,000). The soil taxonomy is generally phases of Subgroup or Association. As of 2022, soil survey field work and reports are still currently being collected in certain areas where detailed information does not exist. This file will be updated as more information becomes available. Typically, this is conducted on an rural municipality basis. In some areas of Manitoba, more detailed and historical information exists than what is contained in this file. However, at this time, some of this information is only available in a hard copy format. This file will be updated as more of this information is transferred into a GIS format. This file has an organizational framework similar to the original SoilAID digital files and a portion of this geographic extent was originally available on the Manitoba Land Initiative (MLI) website. Domains and coded values have also been integrated into the geodatabase files. This allows the user to view attribute information in either an abbreviated or a more descriptive manner. Choosing to display the description of the coded values allows the user to view the expanded information associated with the attribute value (reducing the need to constantly refer to the descriptions within the metadata). To change these settings in ArcCatalog, go to Customize --> ArcCatalog Options --> Tables tab --> check or uncheck 'Display coded value domain and subtype descriptions'. To change these settings in ArcMap, go to Customize --> ArcMapOptions --> Tables tab --> check or uncheck 'Display coded value domain and subtype descriptions'. This setting can also be changed by opening the attribute table, then Table Options (top left) --> Appearance --> check or uncheck 'Display coded value domain and subtype descriptions'. The file also contains field aliases, which can also be turned on or off under Table Options. The file - "Manitoba Municipal Boundaries" - from Manitoba Community Planning Services was used as one of the base administrative references for the soil polygon layer. Also used as references were the hydrological features mapped in the 1:20,000 and 1:50,000 NTS topographical layers (National Topographic System of Canada). Typically this would relate to larger hydrological features such as those designated as perennial lakes and perennial rivers. This same capability is available in ArcGIS Pro. For more info: https://www.gov.mb.ca/agriculture/soil/soil-survey/importance-of-soil-survey-mb.html#

These LULC maps were created through automatic digital classification of RapidEye imagery acquired during the cropping season of 2012 and early dry season of 2013. Three monthly time-steps (June and October, 2012; February 2013) were analyzed.Reference (or field) data on which the classification was based were acquired through a field campaign that lasted from May to October 2012. Standard image pre-processing techniques such as geometric and radiometric correction were conducted on the data prior to classification. The Random Forest classification algorithm was used for classification. Two levels of classification were conducted: (1) a level 1 classification which includes four broad LULC classes and (2) a level 2 classification which comprises of nine LULC classes. The poor temporal coverage of the RapidEye imagery made the accurate delineation of certain crop classes (e.g. groundnuts) very challenging. Nonetheless, an overall accuracy of 70% was obtained