Metadata-Guided Appearance-Shift Consistency for Weed Detection Under Environmental Variation

Environmental variation changes the background appearance against which weeds must be detected. We propose metadata-guided appearance-shift consistency (MASC), a training framework that uses environmental metadata and soil-colour statistics to vary background appearance. Original and transformed images share ground-truth detection supervision, with an additional loss encouraging stable predictions across the two views. The YOLO11n inference network remains unchanged. Evaluation used 8034 CropAndWeed images with acquisition-group-disjoint training, development and test partitions, followed by direct transfer to 3SeasonWeedDet10 without target-domain training or tuning. At a common confidence threshold of 0.25, external weed recall increased from 45.05 ± 4.04% to 49.97 ± 4.51%. Weed average precision at intersection over union 0.5 (AP50) increased from 41.62 ± 2.60% to 44.21 ± 4.52% (mean ± sample standard deviation). Recall, F1 and AP50 improved in all four seed pairs on the complete external release. The mean recall gain remained 4.31 percentage points after excluding all exact-duplicate groups. On the principal source partition, recall was 59.17 ± 1.40% versus 58.81 ± 2.47%, whereas mean all-class mAP50 increased from 66.05% to 67.88%. An alternative partition favoured MASC by 3.92 recall points for one seed. Development-selected oversampling with shifted-view supervision increased source recall by 2.62 points in a single-seed comparison. MASC improved weed detection on the evaluated external dataset while retaining baseline inference complexity, supporting its effectiveness for cross-dataset transfer.

Authors

Institutions

Publication Details

Journal
Electronics
Published
2026-11-16
DOI
https://doi.org/10.3390/electronics15225325
Primary Topic
Smart Agriculture and AI
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Metadata-Guided Appearance-Shift Consistency for Weed Detection Under Environmental Variation

Qiang Li, Yingjia Hu, Shengyuan Chi, Jia Yu
1 links
Electronics
Smart Agriculture and AI
article

Metadata-Guided Appearance-Shift Consistency for Weed Detection Under Environmental Variation

Qiang Li, Yingjia Hu, Shengyuan Chi, Jia Yu
article en

Abstract

Environmental variation changes the background appearance against which weeds must be detected. We propose metadata-guided appearance-shift consistency (MASC), a training framework that uses environmental metadata and soil-colour statistics to vary background appearance. Original and transformed images share ground-truth detection supervision, with an additional loss encouraging stable predictions across the two views. The YOLO11n inference network remains unchanged. Evaluation used 8034 CropAndWeed images with acquisition-group-disjoint training, development and test partitions, followed by direct transfer to 3SeasonWeedDet10 without target-domain training or tuning. At a common confidence threshold of 0.25, external weed recall increased from 45.05 ± 4.04% to 49.97 ± 4.51%. Weed average precision at intersection over union 0.5 (AP50) increased from 41.62 ± 2.60% to 44.21 ± 4.52% (mean ± sample standard deviation). Recall, F1 and AP50 improved in all four seed pairs on the complete external release. The mean recall gain remained 4.31 percentage points after excluding all exact-duplicate groups. On the principal source partition, recall was 59.17 ± 1.40% versus 58.81 ± 2.47%, whereas mean all-class mAP50 increased from 66.05% to 67.88%. An alternative partition favoured MASC by 3.92 recall points for one seed. Development-selected oversampling with shifted-view supervision increased source recall by 2.62 points in a single-seed comparison. MASC improved weed detection on the evaluated external dataset while retaining baseline inference complexity, supporting its effectiveness for cross-dataset transfer.

ElectronicsVol. 15(22)
Harbin Engineering University (CN), Harbin Institute of Information Technology (CN)
Zero hunger
Smart Agriculture and AI
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon