Development of an Uncertainty Workflow to Support Landsat TIRS Split Window-Derived Surface Temperature Products

TL;DR

The paper addresses unreliable Landsat Level-2 surface temperature uncertainty maps driven by cloud-mask artifacts in the current single-channel approach. It introduces a Split Window–based ST retrieval paired with a TPW-enabled uncertainty workflow, leveraging an XGBoost model to estimate per-pixel TPW from Landsat 9 TIRS-2 data using MODIS TPW as ground truth and MODTRAN/TIGR simulations to link TPW to SW algorithmic error. The key contributions are (i) a TPW–driven uncertainty framework for Landsat 8/9 ST, (ii) a machine learning TPW estimator achieving MAE $0.54$ cm and $R^2=0.89$, and (iii) an adapted SW coefficient set for L9, plus a demonstration that TPW-informed uncertainty reduces false positives and artifacts in uncertainty maps. This work enhances operational Landsat ST products by providing more reliable per-pixel uncertainty, potentially benefiting water resources, agriculture, and disaster monitoring through improved quality assurance of thermal observations.

Abstract

Current Landsat Level 2 surface temperature products are derived using a single-channel (SC) methodology to estimate per-pixel surface temperature (ST) maps from Level~1 radiance data. A known issue with the Level 2 uncertainty, however, is its susceptibility to overestimation of uncertainty due to its dependence on Landsat's cloud mask, which is prone to false-positives. Beginning with Collection 3, the split window (SW) approach will serve as the surface temperature algorithm for the level-2 product, reflecting its adaptability across conditions which necessitates the development of a dedicated uncertainty workflow. We introduce an improved uncertainty workflow, based on a physical parameter called total precipitable water (TPW), that more adequately estimates the uncertainty associated with surface temperature estimates. We leveraged a SW algorithm for estimating surface temperature to drive the uncertainty methodology discussed here. First, considering Landsat is not equipped with the optical channels necessary for deriving TPW, we create an XGBoost-based machine learning pipeline that relates TIRS bands 10 & 11 image radiance to TPW using the MODIS product as reference. The resulting modeling approach achieves a mean absolute error in estimating TPW of 0.54 [cm] and a coefficient of determination (R2) as high as 0.89. Secondly, we propose an improved (SW-based) uncertainty workflow that also uses standard error propagation but incorporates uncertainty as a function of TPW. Our work fills the gap in the operational surface temperature algorithms and their corresponding uncertainty workflow tailored for Landsat 8 and 9, and machine learning based models for predicting atmospheric water vapor using thermal infrared sensor bands on board Landsat 8 and 9.

Figures (16)

• Figure 1: (left) Illustration of the TIRS-class focal plane assembly, (right) relative spectral response functions of TIRS 1 and TIRS 2.
• Figure 2: (a) RGB color representation of a sample Landsat 8 scene with no visible clouds. (b) Corresponding Level 2 Surface Temperature (ST) product generated using the SC algorithm. (c) Associated DTC map, highlighting the presence of false positives. (d) SC ST QA band representing uncertainty in the ST retrieval. We can see that despite the absence of clouds in the scene, false positive cloud detections negatively impact the uncertainty map. Scene Product ID:LC08_L1TP_029030_20220721_20220801_02_T1; Cloud Cover = $0.08\%$.
• Figure 3: We leverage spectral emissivity dataset MODIS UCSB along with various atmospheric profiles from TIGR dataset forward-modeled using MODTRAN to train the SW algorithm and derive $b_k$ coefficients for L8 and L9.
• Figure 4: MODIS UCSB emissivity database. We use 113 emissivity spectra from various materials to train the SW algorithm.
• Figure 5: The scatter plot of algorithmic error vs. TPW. We can see that the higher TPW corresponds to higher algorithmic error.