Source-linked AI summary

On the Transferability of Agricultural Weed Detection Under Cross-Field Distribution Shift

Nikhilesh Prabhakar, Pranuthi Tenali, Wilfredo Abudeye Fernandez, Shekhar Borah, Athresh Karanam, Erik Blasch, Prabha Sundaravadivel, Sriraam Natarajan

arXiv:2608.21254v1cs.CVcs.LG

TL;DR

Cross-field and cross-crop weed detection lacks evidence that source-trained models generalize to new deployment domains. The paper introduces a cotton UAV dataset and compares unsupervised DAOD with domain-adjacent pretraining plus few-shot fine-tuning across target-label budgets. Few-shot fine-tuning with k=25 labeled target examples outperforms DAOD in the reported cross-crop comparison, suggesting modest target supervision may be more productive than algorithmic sophistication in this setting.

  • Problem

    Existing weed-detection studies provide limited evidence about whether models trained on one crop and field generalize to new fields or crop types, despite costly target-field annotation.

  • Method

    The study introduces a cotton UAV dataset, pairs it with soybean data, and compares unsupervised DAOD against domain-adjacent source pretraining followed by few-shot target fine-tuning across label budgets.

  • Results

    k=25 labeled target examples enabled few-shot fine-tuning to outperform DAOD in the reported cross-crop comparison.

  • Takeaways & Limitations

    Source domain selection combined with modest target supervision may be more productive than algorithmic sophistication for cross-crop UAV weed detection.

  • Takeaways & Limitations

    ALDI was applied out of the box, so more extensive hyperparameter or augmentation search may improve DAOD results.

Abstract

from arXiv · show

Accurate agricultural weed detection in real-world field conditions is essential for precision agriculture, enabling targeted intervention and reducing yield loss. Recent work has reported strong detection performance from UAV-based imagery across a range of crops, yet existing approaches evaluate within a single crop and field, leaving practitioners with little evidence that a model trained on one crop will generalize to a new field or crop type. In this work, we characterize where cross-dataset weed-localization performance degrades and which modeling choices recover it, reducing the need to relabel every new deployment field. We introduce a newly collected and annotated UAV image dataset for agricultural weed detection in cotton fields and use it alongside an existing soybean dataset collected under a similar protocol. Using these datasets, we evaluate the performance of several strategies for transferring a detector trained on one crop to another, comparing unsupervised domain adaptive object detection (DAOD) against pretraining on a domain-adjacent source dataset followed by few-shot fine-tuning on the target dataset. Our analysis spans target-domain label budgets from zero to the full target dataset, characterizing the trade-off between adaptation strategy and annotation effort. We find that few-shot fine-tuning with as few as 25 labeled target examples outperforms unsupervised DAOD in our cross-crop comparison, suggesting that source domain selection combined with modest target supervision is more productive than algorithmic sophistication in adaptation.

1 Introduction

UAV weed detectors often generalize poorly beyond the narrow fields and crops represented in their training data, motivating cross-crop transfer evaluation. This study compares unsupervised DAOD with domain-adjacent pretraining and few-shot target fine-tuning across label budgets, finding the latter effective with limited supervision.

  • Motivation: UAV weed detection supports field-scale monitoring, but conventional labeled-data collection is costly and datasets often cover narrow locations, growth stages, or species.Expert bounding-box annotation requires knowledge of crop and weed species.
  • Study scope: The study addresses underexplored UAV agricultural imagery adaptation by evaluating cross-crop transfer between newly collected cotton data and a previously published soybean dataset.The datasets were collected under similar protocols.
  • Study scope: The experiments compare unsupervised DAOD with domain-adjacent source pretraining followed by fine-tuning on small numbers of labeled target examples.The analysis varies the number of available target-domain labels.
  • Key finding: k=25 labeled target examples suffice for domain-adjacent fine-tuning to outperform DAOD on the same target in the reported cross-crop comparison.The setup spans budgets from k=0 to the full target dataset.
  • Key finding: The analysis suggests source selection plus few-shot fine-tuning may be more productive than algorithmic sophistication when modest target supervision is practical.This conclusion is scoped to cross-crop transfer in UAV-based weed detection.

2 Background and Related Work

UAV imagery enables field-scale weed detection, but overhead visual similarity and narrow dataset variation limit transfer across fields, seasons, and crops. Domain adaptation is promising yet remains limited in UAV-based weed detection research.

  • UAV weed detection: UAVs provide high-resolution, field-scale imagery at low operational cost, and deep learning has become predominant for processing it.Architectures have been explored across crops including soybean and cotton.
  • Generalization challenge: Overhead visual similarity between weeds and crops makes interclass separation and model generalization difficult.This challenge affects transferability across fields, seasons, and crop types.
  • Generalization challenge: Weed datasets often have limited variability because they are collected from a single location, growth stage, or weed species.Such narrow coverage hinders generalization to new field conditions and different crops.
  • Related work: Domain adaptation is promising but underexplored for weed detection, with prior work focused mainly on ground-level imagery while UAV applications remain limited.The paper studies domain-adaptive object detection on UAV-acquired weed datasets.

3 Datasets

The evaluation combines four weed-imagery datasets spanning crops, weed species, altitude, and sensors. It adds a newly collected, manually annotated cotton UAV dataset alongside a soybean dataset collected under a similar protocol.

  • Dataset collection: The evaluation uses four datasets varying in crop, weed species, altitude, and sensor.The supplied passage identifies soybean and cotton datasets plus two public datasets.
  • Soybean dataset: The soybean dataset contains 637 high-resolution RGB UAV images captured at 25 ft and tiled into 1628 640 × 640 images.Annotations include bounding boxes for pigweed.
  • Cotton dataset: The new cotton dataset contains 590 manually annotated UAV images collected over a ∼2.3 km survey at 25 ft and tiled into 5706 640 × 640 images.The UAV survey used an IF1200A flying at 13–15 mph with ground resolution ≈2.13 cm/pixel.

4 Methodology

The methodology frames adaptation as transferring a source-trained detector to a target field with limited labels, comparing DAOD mechanisms against supervised few-shot fine-tuning. DAOD uses a shared ALDI recipe, while fine-tuning varies labeled-target budgets and random subsets.

  • Adaptation setup: The task adapts a pretrained source detector to a target domain using n target examples, of which k are labeled.The formulation targets satisfactory performance while reducing target-field supervision requirements.
  • DAOD: The study compares adversarial feature alignment with self-distillation as the two dominant DAOD families.Alignment uses domain classifiers and gradient reversal; self-distillation uses an EMA teacher and masked-image consistency.
  • DAOD: ALDI standardizes DAOD experiments with a common Faster R-CNN backbone, training recipe, augmentation, EMA teacher, and source-only checkpoint.The shared recipe isolates adaptation objectives as the principal methodological difference.
  • Few-shot fine-tuning: Few-shot fine-tuning uses K ∈{5, 10, 25} labeled target images, with M=5 independent random subsets for each budget.It uses standard supervised training without adaptation losses or unlabeled target data.

5 Experiments

The experiments compare direct transfer, unsupervised domain adaptation, and few-shot fine-tuning across cross-field weed-detection settings. Performance varies substantially by source-target pair, with limited target labels helping on learnable targets but not uniformly.

  • Experimental design: The study evaluates whether cross-field transfer works, whether unsupervised adaptation improves it, and whether few-shot learning is more effective when adaptation falls short.All detectors use Faster R-CNN with a ResNet-50-FPN backbone under the ALDI framework and fair-baseline recipe.
  • Q1: Cross-field transfer: Most zero-shot transfers collapse to single-digit mAP@50, except Cotton→Soybean at 57.8 and Cotton→CWDet12 at 31.8.The exceptions share weed species or detection scale with the source.
  • Q2: Unsupervised domain adaptation: DAOD improvements over source-only training are highly variable and often remain far below in-domain performance.The reported comparison uses target-test mAP@50, with checkpoints selected on the target validation split.
  • Q3: Few-shot fine-tuning: On soybean, k=5 reaches 66.7±4.3 and k=25 reaches 80.0±3.2, exceeding the best DAOD result and approaching the 76.8 in-domain score.Full fine-tuning reaches 84.9±0.5, indicating added value from source pretraining.
  • Q3: Few-shot fine-tuning: Few-shot fine-tuning does not effectively transfer to cotton with only a handful of labels, behaving similarly to DAOD.Thus, the benefit of modest target supervision depends on target learnability.

6 Conclusion

The paper studies cross-crop UAV weed-detection transfer using a new cotton dataset and compares DAOD with few-shot fine-tuning across target-label budgets. In this setting, DAOD is inconsistent, while fine-tuning with as few as k=25 labeled target examples is more reliable.

  • 6 Conclusion: The study introduces a cotton UAV weed-detection dataset and compares unsupervised DAOD with few-shot fine-tuning from a domain-adjacent source.The evaluation spans target label budgets from zero to the full target dataset.
  • 6 Conclusion: Applying ALDI out of the box yields inconsistent improvements over source-only training, whereas k=25 labeled target examples provides more reliable adaptation in this setting.The paper frames this as an initial investigation of incorporating limited target-domain information within DDDAS.
  • 6 Conclusion: Future work includes closing the DDDAS loop with detection feedback and extending analysis to additional crops and domain discrepancies.The authors also propose using these observations to develop new DAOD methods.
Loading 2608.21254v1…