Self-supervised training for high-resolution close-range multispectral remote sensing imagery

Jul 16, 2026

Upload by Aurelia

Self-supervised learning for close-range remote sensing can reduce annotation effort, yet its value for high-resolution multispectral UAV imagery has remained limited by data availability. This study assesses SSL pretraining for precision agriculture using cm-scale multispectral drone imagery captured across multiple sensors, years, and regions. Transformer encoders are pretrained with MoCo-v3 and Masked Autoencoders on a harmonized msuav500K plus Finland multi-year dataset. A Swin Transformer with MoCo-v3 delivers the strongest crop-weed segmentation results and shows cross-sensor and cross-region generalization while a public dataset supports future work.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages16

Reading Time40 min

Abstract

Introduction

Motivation and problem setting

Pretraining datasets and harmonization

Downstream evaluation with WeedMap

Public dataset release

Spectral-band strategy and hypotheses

Self-supervised training for high-resolution close-range multispectral remote sensing imagery Leon-Friedrich Thomas1,2, Mikael Änäkkälä1, and Antti Lajunen1 1 University of Helsinki, Department of Agricultural Sciences, P.O. Box 28, Fi-00014 Helsinki, Finland 2 University of Helsinki, Department of Forest Sciences, P.O. Box 27, Fi-00014 Helsinki, Finland Corresponding author: Leon-Friedrich Thomas (e-mail: [leon-friedrich.thomas@helsinki.fi](leon-friedrich.thomas@helsinki.fi)). “This project received funding from the Maatalouskoneiden Tutkimussäätiö (Agricultural Machinery Research Foundation) . Open access funded by Helsinki University Library.” ABSTRACT Although self-supervised learning (SSL) offers a promising way to reduce annotation effort inclose-range remote sensing, its effectiveness for high-resolution multispectral unmanned aerial vehicle (UAV) imagery remains underexplored due to limited data. This study evaluated SSL pretraining for precision agriculture using cm-scale multispectral drone imagery collected across multiple sensors, years, and regions. Transformer-based encoders were pretrained with Momentum Contrast v3 (MoCo-v3) and Masked Autoencoders on a harmonized dataset combining msuav500K with newly collected multi-year UAV imagery from agricultural fields in Finland. Pretraining used four spectral bands (Green, Red, Red-Edge, Near-Infrared) for cross-sensor compatibility. The models were evaluated on crop-weed semantic segmentation using the WeedMap dataset with 5–100 \% training data. The following two subsets served as downstream tasks: Task A (Germany, RedEdge-M), where all pretrained models were compared under partial and full fine-tuning, and Task B (Switzerland, Sequoia), where the best encoder from Task A was assessed. Our Swin Transformer pretrained with MoCo-v3 achieved the strongest performance on both tasks, surpassing the Swin Transformer model of Doornbos et al. pretrained on a pre-release of msuav500K. Our pretrained Swin Transformer further demonstrated cross-sensor and cross-region generalization. We additionally provide a public multi-year multispectral UAV dataset from Finland to support future research. INDEX TERMS Deep learning, Masked-Autoencoder, Momentum Contrast, Precision Agriculture, Unmanned aerial vehicle, msuav500k I. INTRODUCTION In recent years, deep learning has received significant attention in the field of remote sensing and has become increasingly integrated into precision agriculture research. This trend leverages the advancements in computer-vision models, particularly in tasks such as weed detection. While fully supervised deep learning and transfer learning are the predominant techniques for training and deploying deeplearning models, recent progress in computer vision have led to development of numerous foundation models tailored to remote-sensing applications [1] . These models are typically trained in a self-supervised manner on sensorspecific data or across a select range of sensor datasets [1] . In this study, we pretrained deep-learning models using high-resolution multispectral unmanned aerial vehicle (UAV) imagery acquired across multiple years, sites, and sensor types, suitable for precision agriculture. The models were pretrained in a self-supervised manner using a Momentum Contrast (MoCo-v3) or Masked-Autoencoder (MAE) framework on a large-scale multispectral dataset msuav500K, which was extended and harmonized with a self-collected multi-year multispectral UAV dataset of agricultural fields in Finland [2-4] . This combined datasetspans multiple countries, environments, and a variety of multispectral sensors used in precision agriculture and forestry. As no Nordic countries were previously represented in the msuav500K dataset due to lack of publicly available data, the combined and extended dataset which is utilized in this research is referred to as msuav500k+N [2] . To evaluate the pretrained models , we utilized a downstream task of cro

FAQs

What problem does this study address in self-supervised learning for remote sensing?

It evaluates how self-supervised learning can improve performance for high-resolution multispectral close-range UAV imagery, where effectiveness is constrained by limited data availability.

Which self-supervised methods and dataset are used for pretraining?

Transformer-based encoders are pretrained using MoCo-v3 and Masked Autoencoders on a harmonized dataset combining msuav500K with newly collected multi-year UAV imagery from Finland.

How are the pretrained models evaluated?

They are evaluated on crop-weed semantic segmentation using the WeedMap dataset, with Task A using Germany/RedEdge-M partial or full fine-tuning and Task B using Switzerland/Sequoia on the best encoder from Task A.

Jul 16, 2026

Upload by Aurelia