The authors applied semi‑supervised deep‑learning to super‑resolution images of modern and fossil grass pollen, training convolutional neural networks to extract abstract morphological features. These features were used to quantify past grass community diversity and C3:C4 ratios in a 25,000‑year lake‑sediment record, revealing a marked diversity loss during the last glacial and a gradual decline of C4 grasses in the Holocene.

The authors introduce AdaPoinTr, a geometry-aware transformer that predicts the alpha‑shape of coniferous tree crowns from incomplete terrestrial or mobile laser‑scanning point clouds, focusing on crown reconstruction rather than full tree completion. Trained on synthetically generated partial crowns, the model consistently improves crown shape similarity and reduces height estimation bias across three diverse forest datasets, providing a cost‑effective solution for enhanced 3D forest structural monitoring.

The authors used a bottom‑up thermodynamic modelling framework to investigate how plants decode calcium signals, starting from Ca2+ binding to EF‑hand proteins and extending to higher‑order decoding modules. They identified six universal Ca2+-decoding modules that can explain variations in calcium sensitivity among kinases and provide a theoretical basis for interpreting calcium signal amplitude and frequency in plant cells.

The study assessed how well common deep learning models (ResNet, EfficientNet, Inception, MobileNet) generalize across different tomato pest and disease image datasets. While models performed well on the dataset they were trained on, they suffered substantial accuracy drops when applied to other datasets, indicating that architectural changes alone cannot overcome dataset variability. The results highlight the necessity for more diverse, representative training data to improve real-world deployment of PPD diagnostic tools.

The study demonstrates that hyperspectral imaging can non‑destructively differentiate active nitrogen‑fixing root nodules from non‑fixing nodules and root tissue based on distinct spectral signatures. By integrating deep‑learning models, the authors created an automated nodule counting pipeline that works across multiple legume species and growth conditions, eliminating labor‑intensive manual counting and reliably detecting nodules within dense root systems.

The study introduces the Botanical Spectrum Analyzer (BSA), a GUI that incorporates a modified U‑Net deep neural network for accurate segmentation of plant images from RGB and hyperspectral (VNIR and SWIR) data. BSA was tested on wheat, barley, and Arabidopsis datasets, achieving >99% accuracy and F1‑scores above 98%, and markedly outperformed commercial tools on root segmentation tasks.

The study introduces a deep‑learning based image analysis pipeline that scores blackleg disease severity from stem cross‑section images of canola species, achieving greater consistency than median expert raters while preserving comparable heritability of susceptibility traits. This standardized scoring method aims to improve selection of resistant varieties in breeding programs.

The study presents a deep‑learning pipeline that uses state‑of‑the‑art convolutional neural networks to automatically estimate the establishment of perennial groundcovers in agricultural research plots from smartphone images. By employing region‑of‑interest markers and deploying the models on AWS SageMaker with a lightweight Django web interface, the approach provides fast, objective, and reproducible assessments that can be adopted by researchers and growers across the Midwest.

The study used an interpretable convolutional neural network to predict ABA responsiveness from proximal promoter sequences in Arabidopsis thaliana, revealing both known ABF-binding motifs and novel regulatory elements. Model performance was boosted by advanced data augmentation, and predicted regulatory regions were experimentally validated using reporter lines, confirming the inferred cis‑regulatory code.

The study presents GenoRetriever, an interpretable deep learning framework trained on STRIPE-seq data from soybean and other crops, that predicts transcription start site locations and usage by identifying 27 core promoter motifs. Validation using in silico motif insertions, saturation mutagenesis, and CRISPR‑Cas9 promoter editing demonstrates high predictive accuracy and reveals domestication‑driven motif usage shifts and lineage‑specific effects. The tool is provided via a web server for promoter analysis and design, offering a new resource for plant functional genomics and crop improvement.