AgriField-40K & AgriMAE

Abstract

Field-based agricultural computer vision is important for precision agriculture, yet it largely depends on expensive annotations and costly adaptation of large pretrained models. We introduce AgriField-40K, a field-centric dataset curated from 17 public resources and covering diverse crops, weeds, and field conditions. Building on this, we present AgriMAE, a parameter-efficient continual pretraining baseline that adapts a masked autoencoder pretrained on natural images by training only lightweight adapters. We further explore semantic feature reconstruction as an alternative pretraining objective and evaluate transfer across multiple tasks. AgriMAE consistently improves downstream performance and can match or even outperform full fine-tuning while using up to 9× fewer trainable parameters, showing that AgriField-40K is a practical resource for continual pretraining in agricultural vision.

~40K

field images

public source datasets

9×

fewer trainable params vs. full FT

downstream tasks evaluated

Overview

Vision models pretrained on natural images transfer poorly to agricultural fields, where repeated plant structures, occlusion, and shifting illumination, growth stage, and acquisition platform create a real domain gap. Fully adapting a large backbone to close this gap is expensive and risks overfitting on limited field data. We tackle this with two contributions: AgriField-40K, a unified field-centric pretraining corpus, and AgriMAE, which closes the gap by training only lightweight adapters and keeping the pretrained backbone frozen throughout continual pretraining and downstream adaptation alike.

Pretrain

Start from an MAE pretrained on natural images (ImageNet).

→

Continually pretrain

Train only adapters on AgriField-40K, pixel or feature reconstruction, backbone frozen.

→

Adapt downstream

Fine-tune adapters + task heads for classification, segmentation, detection.

Contributions

01
AgriField-40K. The first field-centric agricultural dataset for representation learning and continual pretraining, built from 17 public datasets spanning diverse crops, weeds, and field conditions.
02
AgriMAE. A simple, parameter-efficient baseline that adapts a standard MAE to field imagery by freezing the pretrained backbone and training only lightweight adapters.
03
Reconstruction objectives. A study comparing pixel-level reconstruction with semantic feature reconstruction (via a frozen DINOv3 teacher) for agricultural continual pretraining.
04
Downstream validation. Continual pretraining with AgriMAE consistently improves classification, segmentation, and detection performance in parameter-efficient settings.

AgriField-40K Dataset

To support the systematic study of efficient continual pretraining, we introduce AgriField-40K, a field-centric agricultural dataset for representation learning, built from 17 public resources into a unified corpus of ~40,000 curated RGB images. Unlike leaf-level or species-ID datasets, it focuses on real field conditions (i.e., crop/weed mixtures, dense canopies, soil backgrounds, and multiple growth stages) captured with handheld cameras, robots, drones, and shrouded platforms.

Grid of sample images from AgriField-40K showing diverse crops, weeds, and field conditions — Sample images from AgriField-40K showing diverse crops, weeds, and field conditions.

We discard the original annotations and keep only RGB images, apply fixed-interval sampling to remove near-duplicate frames from video/robot sources, and manually filter low-quality images. Each image is resized and center-cropped to 512×512 using Lanczos interpolation. The dataset is split 80/20 for training and validation, and is released under CC BY-SA 4.0.

Dataset	Year	License	Retained	Domain	Task
MuST-C	2026	CC BY 4.0	7,242	Sugar Beet, Soybean, Potato, Maize, Wheat	–
PhenoBench	2026	CC BY-SA 4.0	9,606	Sugar Beet & 6 Weeds	Segmentation
LUCASVision	2023	CC BY 4.0	11,195	12 Crops	Classification
VegAnn	2022	CC BY 1.0	1,607	26+ Crops	Segmentation
WE3DS	2023	CC BY 4.0	1,553	7 Crops & 10 Weeds	Segmentation
… 12 more sources
AgriField-40K	2026	CC BY-SA 4.0	39,963	Field-Centric	Pretraining

Full breakdown of all 17 sources, licenses, and original vs. retained sizes is in the paper (Table 1).

Method

Large-scale pretraining gives vision models strong general-purpose features, but these models are typically pretrained on natural images like ImageNet. Agricultural field imagery looks quite different, with repeated plant structures, fine-grained crop/weed distinctions, occlusion, and large swings in illumination, growth stage, and acquisition platform, creating a domain gap that limits transfer to real field tasks. The natural fix is continual pretraining on in-domain data, but fully updating a large model can be costly, making parameter-efficient adaptation especially valuable when labelled data and compute are limited.

AgriMAE architecture with semantic feature reconstruction — Overview of AgriMAE with semantic feature reconstruction. An ImageNet-pretrained MAE encoder–decoder is adapted with lightweight adapter modules on AgriField-40K to predict patch-level features from a frozen feature extractor, with the loss computed over both masked and visible patches.

To keep the study controlled, we build on masked image modelling and establish a strong parameter-efficient baseline for continual pretraining. Our approach, AgriMAE, starts with an ImageNet pretrained MAE and continually pretrains it on AgriField-40K to adapt it to the agricultural field domain, while keeping the original encoder and decoder frozen. Only small adapters inserted into each transformer block are trained, both during this continual pretraining stage and the later downstream supervised fine tuning stage. This keeps adaptation cheap: a single consumer GPU is enough, and the majority of the model's parameters remain frozen.

Standard masked image modelling reconstructs pixels, which is simple and effective but provides limited semantic supervision. As an alternative, we reconstruct dense features from a strong, frozen feature extractor computed across the whole image rather than just the masked patches. This pushes the adapters to recover semantic structure rather than low-level pixel detail, which matters for field imagery where crop, weed, and background are often spread across the whole scene.

t-SNE visualization of feature representations from MAE and AgriMAE models with different pretraining objectives — t-SNE visualization of feature representations from different pretraining methods, extracted from the final layer using images from the LUCASVision subset of the AgriField-40K validation split.

Results

We provide downstream validation showing that our continual pretraining strategy consistently improves performance in parameter-efficient settings across classification, segmentation, and detection. AgriMAE can match or even outperform full fine-tuning while using substantially fewer trainable parameters. We hope AgriField-40K will support future research on efficient continual pretraining and practical agricultural vision applications.