Source: MISSISSIPPI STATE UNIV submitted to NRP
S3DTOOL: A NOVEL SATELLITE-BASED SOIL SAMPLING DESIGN TOOL FOR SITE-SPECIFIC MANAGEMENT USING HISTORICAL CROP SPECTRAL VARIABILITY
Sponsoring Institution
National Institute of Food and Agriculture
Project Status
ACTIVE
Funding Source
Reporting Frequency
Annual
Accession No.
1029896
Grant No.
2023-67019-39169
Cumulative Award Amt.
$299,940.00
Proposal No.
2022-09307
Multistate No.
(N/A)
Project Start Date
Jun 1, 2023
Project End Date
May 31, 2026
Grant Year
2023
Program Code
[A1401]- Foundational Program: Soil Health
Recipient Organization
MISSISSIPPI STATE UNIV
(N/A)
MISSISSIPPI STATE,MS 39762
Performing Department
(N/A)
Non Technical Summary
Soil sampling is a key practice for gathering field-specific data that supports decisions in soil health assessment and management. Successful stratified sampling design relies on the amount and quality of the preliminary data to determine the management zones, but this site-specific information is difficult and costly to acquire, and spatially limited over time, especially for small and medium-sized farmers. There is a critical need for an innovative, science-driven, and easy-to-use decision tool to support soil sampling plans. This research project will develop and validate a new operational large-area Satellite-based Soil Sampling Design Tool (S3DTool) using historical crop variability analysis and auxiliary data at field-scale. The proposed research will include four tasks: (1) generate multi-temporal composites of spectral vegetation index from 10-m Sentinel-2 MSI for sub-field crop variability analysis, (2) extract crop fields from 1-m NAIP imagery using automated deep learning algorithm, (3) develop a S3DTool algorithm to efficiently derive site-specific management zones, and (4) demonstrate and validate the S3DTool in Mississippi State University Farms. Project deliverables are the annual 10-m crop peak growth composite database for sub-field variability analysis; high-resolution quality-controlled field boundaries database; a validated soil sampling design tool that provides operational zone management delineation; and distribution of collected soil analysis results in different soil conditions. These proposal outcomes will contribute for soil health and agricultural sustainability by promoting a robust, validated, timely and significant tool that can enhance soil assessment and management decisions and increase the economic feasibility of soil testing in heterogeneous areas.
Animal Health Component
50%
Research Effort Categories
Basic
0%
Applied
50%
Developmental
50%
Classification

Knowledge Area (KA)Subject of Investigation (SOI)Field of Science (FOS)Percent
10201992080100%
Goals / Objectives
The major goal of this project is todevelop and validate a new satellite-based soil sampling design tool, called S3DTool, using historical crop variability analysis from contemporaneous satellite observations and auxiliary data. The prototype tool will be developed and validated in the state of Mississippi. The specificobjectives of this project are:Objective1: Generate multi-temporal composites of spectral vegetation index from 10-m Sentinel-2 MSI for sub-field crop variability analysisObjective2: Extract crop fields from 1m NAIP imagery using automated deep learning algorithmObjective3: Develop a satellite-based soil sampling design tool (S3DTool) to efficiently derive site-specific management zonesObjective4: Demonstrate and validate the S3DTool in Mississippi State University Farms
Project Methods
There are four main objectives and each one has a specific method.Objective 1:Generate multi-temporal composites of spectral vegetation index from 10-m Sentinel-2 MSI for sub-field crop variability analysisThe annual Sentinel-2 EVI composite will be generated between the start and end dates of peak growth in each year (2015 - 2022). First, we will construct a cloud-free time series data of EVI for each pixel. Second, we will fit a harmonic model in the EVI time series, and then derive the start of peak (SOP) and end of peak (EOP) dates from fitted harmonic values, considering only values >75% EVI amplitude.We will generate a cloud-free median EVI composite between SOP and EOP at each gridded pixel location to create a set of temporal information-embedded metric image.Annual EVI composites will be generated for entire Mississippi between 2015 and 2022.Objective 2:Extract crop fields from 1-m NAIP imagery using automated deep learning algorithmWe will develop a deep learning instance segmentation for field extraction from annual 1-m NAIP imagery.The U-Net-ID model will be trained with input variables as annual cloud-free NAIP false-color bands and output layer as field boundaries that will be manually generated by the research team.We will perform a quality assessment of classified field boundary results by visual inspection, and if needed, more samples will be included to re-train the model in landscape conditions of low model confidence.After a few iterations, the minimum error is expected in the final result, and the field boundary maps will be integratedfor sub-field analysis.Objective 3:Develop a satellite-based soil sampling design tool (S3DT) to efficiently derive field-specific management zonesWe will develop a S3DTool at field level by integrating the multi-temporal peak growth EVI composites and auxiliary soil and topographic attributes for management zone delimitation and random sampling point per zone.First, Principal Component Analysis (PCA) will be applied to annual EVI composites, soil, and topographic variables, and the first principal component, which typically has >90% of data variance, will be used to characterize the spatial variability of aggregated variables.Second, the first principal component will be the input in K-means clustering algorithm for management zone delineation.The resulting clusters can be used to identify management zones (MZs), i.e. within-field areas with similar characteristics (Moral et al., 2010). Optimal number of management zones will be defined using Silhouette analysis with different number of clusters.The Silhouette method estimates the clustering performance by measuring the similarity of the data within its cluster by comparing it to the data of other clusters.Once we defined the management zones, random sampling point is generated within each zone. The S3DTool intends to be effective and simple, and a minimum set of input parameters will be proposed: (i) minimum distance between sampling points (default: 50 m), (ii) minimum distance from field edge (default: 10 m), (iii) minimum area of each management zone (default: 2.5 acres), and (iv) number of sampling point per zone (default: 1). Once we have all verified, sampling locations are projected in the field area and exported in shapefile and excel formats.Objective 4:Demonstrate and validate the S3DTool in Mississippi State University FarmsWe will demonstrate and validate the new S3DTool across Mississippi State University research farms. A very dense soil sampling survey of soil texture and chemical properties (0 to 15 cm depth) will be sampled across a grid with a regular distance of 50 m for different soil and crop types, typically, corn, soybean, and cotton. Careful sampling will be performed to maximize the representation of soil properties, and each sampling location will be geo-referenced using a GPS device with a locational error <2 m. We will follow the recommended soil health indicators in NRCS-2018-0006 technical note, and soil testing analysis will include wet aggregate stability, soil organic C content, pH, moisture content, texture, bulk density, nitrogen, and phosphorus concentration (Co-I Wijewardane expertise). The ground-truth soil result variables will be interpolated with kriging method and the clustering analysis will be performed with k-means. The confusion matrix and accuracy metrics of S3DTool and ground-truth management zones will be estimated, and ANOVA to test the difference among the management zones to inform the ability of the proposed tool in representing soil variability.

Progress 06/01/24 to 05/31/25

Outputs
Target Audience:Target audiences are agricultural and technology students, farmers, soil scientists, and agronomists. Changes/Problems: Nothing Reported What opportunities for training and professional development has the project provided?The project has offered a great opportunity for training and professional development of the postdoctoral research fellow responsible for implementing the software and analysis needed to carry out the proposed study. Several tasks that support the development of abilities related to programming, data analysis, satellite image processing, and scientific writing have been carried out, contributing to the improvement of the technical skills and overall research capabilities of the postdoctoral fellow. In addition, field activities have complemented the experience of the postdoctoral research fellow and other students in agricultural studies. How have the results been disseminated to communities of interest?We currently have an open-access scientific paper published in Computers and Electronics in Agriculture (https://www.sciencedirect.com/science/article/pii/S0168169925001929), contributing to disseminating the results of the project. In addition, as part of our efforts to develop efficient tools, we developed two open-source Python libraries and made them available to the scientific community. The library (https://github.com/lbferreira/ezgrid) offers an easy alternative to create sampling points within a field based on different strategies. What do you plan to do during the next reporting period to accomplish the goals?To further advance our work, we plan to carry out the mapping of soil attributes with collected soil samples, and compare the management zones generated with in situ samples vs the proposed S3DTool. We willimprove the current implementation of our framework dedicated to management zones generation. New agricultural fields will be selected based on historical satellite images in order to sample fields with high spatial variability. Our framework for management zone creation will be further tested to enhance its capability to generate meaningful zones (and locations for soil sampling) based on historical satellite observationswhile keeping computational efficiency.

Impacts
What was accomplished under these goals? We have successfully developed the data processing pipeline that supports the generation of multi-temporal composites of vegetation indices (Objective 1) and have completed the development of a scalable field boundaries extraction framework (Objective 2). Our implementations offer efficient solutions for evaluating sub-field crop variability and extracting agricultural field boundaries under different agricultural scenarios. A field boundaries dataset covering the whole state of Mississippi was generated, supporting our analysis and providing a rich data layer for further investigations. As part of objective 3, we have also made extensive progress in the development of a framework to generate management zones based on multi-temporal composites of a vegetation index and terrain data (elevation and slope). The tool has multiple requirements to create soil sampling zones, such as number of samples and minimum area.Lastly, in 2024, we performed a soil sampling in a dense grid (134 samples)in two agricultural fields in a Mississippi State University farm, and we finished a second soil sampling campaign (2025) with 94 samples in one corn field, which will extend our soil database and support the validation of the proposed S3DTool.

Publications

  • Type: Peer Reviewed Journal Articles Status: Published Year Published: 2025 Citation: Ferreira, L. B., Martins, V. S., Aires, U. R., Wijewardane, N., Zhang, X., & Samiappan, S. (2025). FieldSeg: A scalable agricultural field extraction framework based on the Segment Anything Model and 10-m Sentinel-2 imagery. Computers and Electronics in Agriculture, 232, 110086.


Progress 06/01/23 to 05/31/24

Outputs
Target Audience:Target audiences are farmers, soil scientists, and agronomists. Changes/Problems: Nothing Reported What opportunities for training and professional development has the project provided?The project has offered a great opportunity for the training of the postdoctoral research fellow responsible for implementing the software and analysis needed to carry out the proposed studies. Several tasks that stimulate the development of abilities related to programming, data analysis, satellite image processing, and scientific writing have been carried out, contributing to the improvement of the technical skills and overall research capabilities of the postdoctoral fellow. How have the results been disseminated to communities of interest?As part of our efforts to develop a new supporting tool, we developed a Python library called fastnanquantile, and made available for the scientific community. We released it as an open-source project (https://github.com/lbferreira/fastnanquantile), which has more than 1,000 downloads. Our current results also include a scientific paper in final revision that will be submitted to the journal Remote Sensing of Environment, where, if accepted, it will be widely accessible to the scientific audience. What do you plan to do during the next reporting period to accomplish the goals?Based on the data/tools that we have developed (Sentinel-2 multi-temporal composites, crop field boundaries, and soil attribute maps), our next steps include the development of S3Dtool to create soil management zones and validate these results with soil health samples. These zones will help us suggest possible soil sampling locations to minimize the number of sampling points while preserving the ability to represent the actual spatial variability of the soil properties. We plan to explore different clustering techniques (K-means, DBSCAN, Gaussian Mixture Models, etc.), auxiliary datasets (topographic, soil maps, etc.), and different strategies to perform our analysis and experiments. The proposed method will be validated based on field data collected in a high spatial density.

Impacts
What was accomplished under these goals? We have made extensive progress in the development of satellite data processing pipelines for objectives 1 and 2. Our automated framework used advanced data processing libraries and parallel processing with Python coding to calculate temporal metrics from Sentinel-2 image time series over croplands in Mississippi. We also developed a crop field extraction framework based on a recently released foundational model called Segment Anything Model (SAM). Our framework was validated in different agricultural regions of the world, offering promising results, and currently, we are generating crop field boundaries covering the state of Mississippi. As part of our efforts to develop and validate a soil sampling tool (S3Dtool) in objectives 3 and 4, we selected two crop fields on Mississippi State University farms and carried out a soil sampling in a dense grid (25 m, >100 sub-surface locations). These soil samples were sent for laboratory analysis and will help us to evaluate how the soil maps generated based on the S3Dtool compare with the soil maps produced with the dense grid sampling (reference).

Publications