Semi-Supervised Learning for Regression Problems in Earth ObservationV

MCML Authors

Xiaoxiang Zhu

Prof. Dr.

Core PI

Data Science in Earth Observation

Abstract

Semi-supervised learning has been widely applied to classification problems, but remains under exploited in regression tasks. This is due to the lack of consistent quality measure for pseudo-label in regression. It is particularly evident for semi-supervised frameworks that require pseudo-label such as FixMatch. To bridge this gap, we propose a novel algorithm by investigating aleatoric and epistemic uncertainties using Bayesian neural network, and establish an effective and adaptive pseudo-label selection criterion to handle heteroscedastic data. Requesting the network to estimate uncertainty and minimize consistency loss in semi-supervised learning seems contradicting, as the network would need to retain and ignore the noise simultaneously. We introduced a separate projection head before the loss function in the semi-supervised branch, in order to preserve noise information. The proposed algorithm is a general framework for pseudo-label filtering in regression problems. Our results demonstrate that leveraging predicted uncertainty for pseudo-label filtering significantly improves performance, validating the effectiveness of this concept. The code will be made available after acceptance.

inproceedings WZ25

---

MIGARS 2025

IEEE International Conference on Machine Intelligence for GeoAnalytics and Remote Sensing. Bucharest, Romania, Sep 02-04, 2025.

---

Authors

Y. Wang • X. Zhu

---

Links

DOI

---

Research Area

 C3 | Physics and Geo Sciences