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

Background: The intersection of machine learning (ML) and deep learning (DL) with thermal remote sensing (TRS) has undergone a transformative expansion since 2018, driven by the proliferation of high-resolution satellite missions and open-source deep learning frameworks. Despite this rapid growth, to the best of our knowledge, no comprehensive PRISMA-compliant systematic review has synthesised ML/DL applications specifically within the thermal RS domain across the post-2018 period.Objectives: This review maps the complete landscape of ML/DL applications in thermal RS from January 2018 to March 2026 with five primary objectives: (i) quantify publication trends; (ii) classify the taxonomy of ML/DL architectures; (iii) map application domain coverage; (iv) appraise methodological quality and open science practices; and (v) identify research gaps and future directions.Methods: A systematic electronic search was conducted across Scopus and Google Scholar. Following PRISMA 2020 guidelines (Page et al., 2021), records underwent a structured multi-stage screening process implemented in Python. This consisted of five main screening stages after initial deduplication, followed by a final full-text eligibility assessment for open-access records retrieved via the Unpaywall API. Data extraction employed a structured template covering bibliographic metadata, sensor platforms, ML/DL architecture, application domain, performance metrics, and open science practices.Results: A total of 193 peer-reviewed studies met the inclusion criteria, of which 93 were available as open-access full texts and 100 were accessible through title, abstract, and structured metadata only due to institutional access restrictions. CNNs (43.7%), LSTM/BiLSTM (33.0%), and SVR/SVM (29.1%) were the dominant architectures across the 93 open-access full-text studies from which comprehensive architecture data were extracted. Application domains concentrated on SST forecasting, LST retrieval, LST downscaling, and gap-filling, leaving wildfire detection, evapotranspiration estimation, and permafrost monitoring relatively underrepresented compared to core domains. Code availability was reported in fewer than 5% of included studies.Conclusions: This review reveals a maturing but architecturally conservative field with transformative opportunities in physics-informed neural networks, transformer-based models, and underserved application domains. The persistent open science deficit represents a structural reproducibility challenge that warrants urgent community attention.