With the increasing uncertainty in agricultural production due to recent climate change, the importance of rapid and accurate agricultural change detection using high-resolution satellite imagery is growing. However, agricultural change detection face...
With the increasing uncertainty in agricultural production due to recent climate change, the importance of rapid and accurate agricultural change detection using high-resolution satellite imagery is growing. However, agricultural change detection faces critical technical limitations: seasonal ’pseudo-changes’ driven by crop growth cycles often appear visually stronger than actual land cover changes, causing structural false alarms. Furthermore, constructing high-quality labels on a large scale requires enormous costs, leading to a severe label scarcity problem.
To address these challenges, this paper proposes PhenologyMatch, a computer vision-based, phenology-aware semi-supervised change detection framework that elevates phenological priors to the decision-making layer of the neural network. First, to maximize spatiotemporal representation learning, the proposed method introduces a Spatiotemporal Cross-modal Fusion architecture. This combines a Video Swin Transformer-based RGB branch for extracting high-frequency textures with a 3D-CNN-based NDVI branch for modeling low-frequency phenological dynamics. Through a Cross-modal Attention mechanism, the phenological state adaptively modulates the texture response, inducing the learning of representations robust to seasonal appearance drift.
Additionally, to enhance the stability of semi-supervised learning, a novel Phenology-aware Pseudo-label Gating (Pheno-τ) mechanism is proposed. This mechanism dynamically adjusts the pseudo-label selection threshold by utilizing NDVI statistics as a proxy variable for phenological phases, thereby fundamentally suppressing the influx of seasonal noise and improving the precision of the decision boundary through soft consistency constraints on threshold-boundary samples.
Evaluated under a rigorous protocol combining patch-level mutually exclusive splitting and Season Hold-out evaluation on a large-scale dataset built from Sentinel-2 time-series data of the Matopiba region in Brazil, PhenologyMatch achieved superior cross-season generalization performance and discriminative accuracy compared to representative RGB-only video backbones and semi-supervised baseline settings. This study is expected to enhance the reliability of agricultural monitoring systems utilizing artificial intelligence and computer vision technologies, contributing to the provision of precise land cover change information for food security.