• HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
Wednesday, November 18, 2026
BIOENGINEER.ORG
No Result
View All Result
  • Login
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
No Result
View All Result
Bioengineer.org
No Result
View All Result
Home NEWS Science News Technology

Multitask Deep Learning Model Simultaneously Classifies Soil Nitrogen and Phosphorus Levels

by
November 18, 2026
in Agriculture, Technology
Reading Time: 5 mins read
0
Multitask Deep Learning Model Simultaneously Classifies Soil Nitrogen and Phosphorus Levels

Multitask Deep Learning Model Simultaneously Classifies Soil Nitrogen and Phosphorus Levels

Share on FacebookShare on TwitterShare on LinkedinShare on RedditShare on Telegram

A recent study published in the journal Smart Agricultural Technology has demonstrated that multitask deep learning models can simultaneously classify soil nitrate-nitrogen (N) and Olsen phosphorus (P) concentrations using hyperspectral imaging. By training a single architecture to perform both prediction tasks, researchers achieved classification performance comparable to that of separate, single-task models. This integrated approach offers potential improvements in the efficiency of precision nutrient management by reducing the overall training time and computational overhead required for soil analysis. The research, conducted across multiple agricultural sites in North Dakota, represents a step toward the development of scalable, automated systems for assessing soil nutrient variability at the scale at which fertilizer decisions are typically made.

Precision nutrient management relies on the ability to identify spatial variability within fields to adjust fertilizer application according to specific crop needs. Uniform application often results in over-application in some areas and under-application in others, leading to increased input costs, potential environmental nutrient losses, and suboptimal yields. Conventional laboratory analysis for soil nutrients, while accurate, is often prohibitively laborious and expensive for the level of sampling density required for site-specific management. Sensing technologies such as hyperspectral imaging have emerged as promising alternatives, as they can capture extensive spectral information about soil composition. In this study, the researchers focused on leveraging these spectral features to predict the nutrient status and variability within agricultural fields, aiming to move beyond single-nutrient estimation models that require high computational resources for training and deployment.

The study utilized a dataset of 284 soil samples collected from two distinct agricultural locations in North Dakota: a controlled research plot in Casselton and a commercial farm in Wolford. These sites were chosen to capture a diverse range of management regimes, soil characteristics, and cropping histories, including sugar beet, soybean, corn, and oats. Soil samples were collected at two specific time points during the growing season—pre-planting and post-harvest—to account for seasonal variations in nutrient accumulation and depletion. Samples were taken from the top 12 inches of soil, processed through air-drying and sieving to 2 mm, and subjected to standard laboratory analysis at AGVISE Laboratories. Laboratory measurements for nitrate-nitrogen were performed using the cadmium reduction method, while phosphorus levels were determined via the Olsen method. These concentrations were then grouped into five standardized interpretive classes ranging from very low to very high, providing categorical labels for the training and evaluation of the hyperspectral classification models.

Hyperspectral data acquisition was conducted in a laboratory environment using two high-resolution push-broom sensors: a SPECIM FX10 camera capturing reflectance in the visible and near-infrared (VNIR) region (400–1000 nm) and a SPECIM FX17 camera capturing short-wave infrared (SWIR) data (900–1700 nm). Imaging was performed in a controlled darkroom with regulated halogen illumination to minimize noise and scatter effects from external light. Calibration was systematically maintained throughout the process using dark reference images and a Spectralon white reference tile, which provided a standard for adjusting reflectance data. The resulting hyperspectral cubes were unfolded into two-dimensional pixel-based matrices for analysis, with non-informative pixels containing zero intensity systematically removed. This rigorous acquisition protocol aimed to isolate the soil-specific spectral reflectance from background container interference.

The modeling framework investigated six deep learning architectures, split between standard multitask models and multi-gate mixture-of-experts (MMoE) architectures. The multitask models featured a shared backbone—CNN, ResNet-18, or DenseNet—designed to learn common spectral features, which were then channeled into two parallel task-specific classification heads. The MMoE models extended these backbones by incorporating expert sub-networks and task-specific gating networks to learn both shared and unique feature representations. These models were trained to produce classification probabilities for nitrogen and phosphorus simultaneously, utilizing the softmax activation function to assign each soil sample to one of the five AGVISE classes. The study utilized a stratified 5-fold cross-validation procedure, with three repetitions, to ensure robust evaluation of model performance on the collected samples.

Experimental comparisons revealed that the six models achieved similar performance across both the VNIR and SWIR spectral regions. The mean macro-averaged F1 score, used as the primary metric to account for class imbalance, ranged between 0.429 and 0.462 across all model architectures and spectral inputs. Statistical analysis using the corrected repeated cross-validation t-test with Holm adjustment indicated no significant differences between the models, nor did combining VNIR and SWIR data offer any improvement over the individual spectral ranges. Consequently, the researchers identified the multitask CNN architecture paired with 400–1000 nm spectral data as the final model, favored for its computational efficiency, requiring the fewest trainable parameters and the shortest training time per fold.

Investigations into spectral preprocessing and feature selection further clarified the requirements for successful soil nutrient classification under controlled conditions. Techniques including Standard Normal Variate normalization and Savitzky-Golay first and second-order derivatives, which are often used to correct for scatter and noise in field-acquired data, did not yield consistent improvements in accuracy for this laboratory-based study. In fact, some transformations, such as the second-order derivative, were associated with decreased performance. Similarly, feature selection strategies—specifically mutual information and random forest importance—failed to enhance classification compared to using the full spectrum. These findings suggest that for data acquired under rigorous laboratory standards, the raw reflectance spectra already provide sufficient information for the models to learn the necessary associations, while preprocessing or dimensionality reduction may introduce unnecessary computational burden.

The analysis of spectral band importance, conducted via integrated gradients, provided insight into the physical basis of these classifications. Both nitrogen and phosphorus models relied on overlapping spectral regions: 490–510 nm in the VNIR range, and 1350–1530 nm in the SWIR range. These regions are associated with broad soil properties such as organic matter, iron oxides, soil moisture, and mineral hydroxyls, rather than the nutrient-specific absorption features of nitrate or phosphate ions, which are typically negligible in these spectral ranges and concentrations. The reliance on identical spectral bands for both nutrients and the similar performance of single-task models compared to the multitask network suggest that these models predict nutrient status indirectly through covariation with other soil constituents. This shared spectral relationship underscores the potential effectiveness of standard multitask learning, where models leverage shared feature representations to account for correlated soil properties.

The evaluation of the final multitask CNN on a separate set of confirmation folds revealed a sample-wise accuracy of 0.53 for nitrogen and 0.50 for phosphorus, with macro-F1 scores of 0.41 and 0.50, respectively. These values were highly consistent with the initial cross-validation estimates. Most classification errors occurred between adjacent classes, as evidenced by a high within-one-class accuracy of 0.85 for nitrogen and 0.92 for phosphorus, and quadratic weighted kappa values of 0.76 and 0.75. However, testing the model’s transferability using leave-one-field-out validation—where the entire Wolford commercial farm site was withheld from the training set—revealed significant limitations. Accuracy on the unseen site decreased to 0.37 for nitrogen and 0.15 for phosphorus, indicating that the relationships learned at the Casselton research site did not fully transfer to the Wolford location. The researchers attribute this to site-specific differences in soil characteristics, management practices, and potential confounding variables between the sites that cannot be resolved with the current dataset.

The study highlights both the promise and the current limitations of applying multitask deep learning to soil spectroscopy. While the framework successfully demonstrates the feasibility of simultaneous nitrogen and phosphorus classification without loss of accuracy, the models remain confined to the specific sites represented in the training data. The moderate classification accuracies reported are influenced by the challenging five-class agronomic threshold scheme and the indirect nature of the nutrient-spectral relationship. Future research directions suggested by the authors include the integration of additional soil properties—such as potassium, pH, and organic matter—to build more comprehensive multitask models. Expanding the dataset to include greater geographic and environmental diversity will be essential for assessing model transferability, while exploring regression-based frameworks may provide more continuous, accurate estimates than those offered by categorical classification. Advancements in transformer-based or attention-driven architectures, alongside techniques for edge-computing optimization, represent potential pathways for moving this technology from laboratory analysis toward eventual field-level deployment.

Subject of Research: Hyperspectral classification of soil nitrogen and phosphorus using multitask deep learning.

Article Title: Simultaneous hyperspectral classification of soil nitrogen and phosphorus using multitask deep learning across VNIR and SWIR spectral regions

Article References: Vullaganti, N., Zhang, X., Aderholdt, W., Overby, P., & Sun, X. (2026). Simultaneous hyperspectral classification of soil nitrogen and phosphorus using multitask deep learning across VNIR and SWIR spectral regions. Smart Agricultural Technology, 15, Article 102744. https://doi.org/10.1016/j.atech.2026.102744

Image Credits: AI Generated

DOI: 10.1016/j.atech.2026.102744

Keywords: soil science, hyperspectral imaging, multitask deep learning, precision agriculture, nitrogen classification, phosphorus classification, VNIR, SWIR, CNN, ResNet, DenseNet, soil fertility, remote sensing, nutrient management, agricultural technology

News Source: Blake Davidson. (November 18, 2026). Multitask Deep Learning Model Simultaneously Classifies Soil Nitrogen and Phosphorus Levels. Scienmag.

Tags: Agricultural technologyCNNDenseNethyperspectral imagingmultitask deep learningnitrogen classificationNutrient Managementphosphorus classificationprecision agricultureremote sensingResNetsoil fertilitysoil scienceSWIRVNIR
Share12Tweet7Share2ShareShareShare1

Related Posts

Rethinking Inertial Terms in Halpern-Type Iterations for Strong Convergence

Rethinking Inertial Terms in Halpern-Type Iterations for Strong Convergence

November 18, 2026
Reliability-Aware Query Modulation Improves Accuracy in Transformer-Based Multi-Object Tracking

Reliability-Aware Query Modulation Improves Accuracy in Transformer-Based Multi-Object Tracking

November 18, 2026

Ginsenoside Re Protects Dopaminergic Neurons Through Gut-Brain Axis Modulation

November 18, 2026

GATOC-ERC Framework Combines Conversational Graphs and Language Models for Emotion Recognition

November 18, 2026

POPULAR NEWS

  • Elite Luge Athletes Experience Seasonal Shifts in Body Composition and Biomarkers

    Elite Luge Athletes Experience Seasonal Shifts in Body Composition and Biomarkers

    29 shares
    Share 12 Tweet 7
  • Olive Oil Supplementation Remodels Plasma Lipids and Metabolites in Aged Tauopathy Mice

    29 shares
    Share 12 Tweet 7
  • Mitophagy Inhibitor Ac220 Enhances Magnolol Antitumor Action in Ewing Sarcoma

    29 shares
    Share 12 Tweet 7
  • Systematic Review Finds Long-Term Data Scarce for Chronic Neuropathic Pain Treatments

    29 shares
    Share 12 Tweet 7

About

We bring you the latest biotechnology news from best research centers and universities around the world. Check our website.

Follow us

Recent News

Study Reveals Significant Geographic Disparities in Iranian Hospital Workforce Distribution

Childhood Cardiometabolic Trajectories Persist into Old Age and Predict Adult Disease

Spinal Stenosis Patients in China Balance Treatment Fears Against Cultural Trust

Subscribe to Blog via Email

Enter your email address to subscribe to this blog and receive notifications of new posts by email.

Join 88 other subscribers
  • Contact Us

Bioengineer.org © Copyright 2023 All Rights Reserved.

Welcome Back!

Login to your account below

Forgotten Password?

Retrieve your password

Please enter your username or email address to reset your password.

Log In
No Result
View All Result
  • Homepages
    • Home Page 1
    • Home Page 2
  • News
  • National
  • Business
  • Health
  • Lifestyle
  • Science

Bioengineer.org © Copyright 2023 All Rights Reserved.