Research

Image-level weed severity grades provide supervision for spatial segmentation. A Sqrt-Error-Weighted ordinal objective learns severity-dependent responses; class activation maps are converted into pseudo-masks for segmentation training.

The workflow produces broadleaf and grass-weed masks from grade annotations, reducing reliance on manually drawn pixel labels. Generalization across field conditions remains an evaluation focus.

Image variance decreases with increasing flight altitude, providing a scale-sensitive cue for aerial-image restoration. Variance-Average-Spatial Attention (VASA) incorporates this cue into the iterative reconstruction of the EVADM diffusion model.

On the published real-image benchmarks, SRFI improved by 27% at ×2 and 6% at ×4 relative to the baselines. Cross-dataset and downstream evaluations support restoration of fine crop structure for agricultural perception.

Cascaded waterbody and algal-bloom segmentation separates affected water regions from shoreline and terrestrial backgrounds. Direct geo-location maps image detections into geographic coordinates.

The output is a spatial distribution map of algal-bloom regions, supporting rapid lake surveys and targeted follow-up inspection.

Ghost Bilateral Network (GBiNet) combines efficient ghost-convolution features with bilateral segmentation to identify rice regions and four phenological stages in UAV images.

GBiNet_r2 achieved 91.50% test mIoU on PaddySeg, compared with 90.95% for BiSeNetV2, with lower computational cost. The resulting trait masks support field-scale phenology mapping.

Camera parameters and central projection map image pixels directly to ground coordinates. Incremental sparse sampling removes redundant observations from overlapping UAV images before trait-map interpolation.

The published evaluation reported a mean deviation of 21 cm across 20 ground-control points, below 1% of relative flight height. This provides a direct route from UAV image sequences to georeferenced crop-trait maps.

Ghost convolution extracts local symptom features with inexpensive linear operations; transformer self-attention adds global context for grape-leaf disease and pest classification against complex field backgrounds.

On GLDP12k, the ImageNet-initialized GeT model achieved 98.14% classification accuracy with 1.16 million parameters. The compact architecture supports efficient field-image diagnosis.

A cascaded I2C architecture links modular weighing nodes to a shared data-acquisition system. Low-power Arduino Pro Mini controllers and plug-in PCBs provide a scalable hardware interface.

The implemented prototype acquires weight measurements across 24 channels, providing a modular basis for parallel weighing experiments.

HSV segmentation isolates citrus fruit under controlled-background imaging. Bounding-box pixel dimensions are converted into diameter estimates using the camera imaging geometry.

The workflow provides per-fruit outlines and size estimates for automated measurement; the original calibration plot documents the pixel-to-diameter relationship.

Laser-induced breakdown spectroscopy resolves soil-metal signatures from plasma emission spectra. Argon-assisted excitation and a light-convergence chamber are evaluated for signal enhancement, with spectral analysis identifying informative wavelength features.

The batch-testing platform design combines sample handling and optical acquisition for laboratory soil analysis. The linked slides and device patent document the spectroscopy and signal-collection approach.

Coordinated motor and valve control combines seed metering with liquid-fertilizer application in one mechanism.

The assembled prototype synchronizes seed delivery and spraying, demonstrating integrated electrical control of two field operations.

A curved-groove bearing mechanism converts rotary input into reciprocating linear motion through constrained rolling contact.

The mechanical design and motion animation demonstrate an alternative rotary-to-linear transmission concept for reciprocating machinery.

A labeled UAV-image dataset covers five weed types in seedling-stage wheat fields. Real-time convolutional models translate aerial observations into weed-category predictions.

The workflow connects aerial image acquisition, annotation, and recognition for field weed surveys. The poster and related UAV spraying-device patent preserve the original project materials.

Polynomial curve fitting represents electric-vehicle growth, while a multi-objective location model evaluates candidate charging-station configurations.

The MATLAB implementation produces spatial deployment scenarios for demand-oriented charging infrastructure planning.