Machine Learning and Deep Learning Approaches in Thermal Remote Sensing: A Systematic Review (2018–2026)

Abstract

This systematic review, conducted in accordance with the PRISMA 2020 guidelines, maps the applications of machine learning (ML) and deep learning (DL) within thermal remote sensing (RS) from January 2018 to March 2026. Despite significant growth in this field driven by high-resolution satellite missions and open-source frameworks, no comprehensive PRISMA-compliant synthesis has previously focused exclusively on this domain during the post-2018 period. Through a structured search of Scopus and Google Scholar, we identified 193 peer-reviewed studies meeting inclusion criteria, of which 93 provided full-text access for in-depth methodological appraisal and 100 were analyzed at the metadata level. Our analysis reveals a concentrated architectural landscape, with convolutional neural networks (CNNs), long short-term memory networks (LSTM/BiLSTM), and support vector machines/regression (SVR/SVM) predominating. Application-wise, the literature skews heavily toward sea surface temperature (SST) forecasting and land surface temperature (LST) retrieval and downscaling, while critical areas such as wildfire detection, evapotranspiration estimation, and permafrost monitoring remain notably understudied. A key concern is the severe deficit in open science practices: fewer than 5% of the full-text accessible studies reported code availability, raising substantial reproducibility challenges. While the field demonstrates technical maturity, it remains architecturally conservative, presenting clear opportunities for exploring emerging approaches like physics-informed neural networks and transformer-based models. The findings underscore an urgent need for the community to address reproducibility gaps and to expand ML/DL applications into underserved domains, thereby broadening the scientific impact and operational utility of thermal RS in the coming years.