Transformer‐Based Digital Soil Mapping Reveals Land‐Use Legacy Effects on (Sub)Tropical Soil Carbon
Year: 2026
Type: article
Abstract: Tropical and subtropical soils store approximately one‐third of global soil organic carbon (SOC) despite covering less than one‐quarter of terrestrial surfaces, yet accurate spatiotemporal prediction methods remain inadequate. We present the first application of transformer architectures to multi‐temporal Landsat observations for SOC prediction, demonstrating that temporal attention mechanisms can extract management and phenological signals that static modeling approaches cannot capture. Using 556 soil observations across the Hawaiian Islands and 36‐month sequences of Landsat imagery integrated with climate and terrain data, the model predicted SOC with R 2 = 0.57 (RPIQ = 1.88) in leave‐one‐island‐out cross‐validation and R 2 = 0.60 (RPIQ = 2.00) in 10‐fold cross‐validation. We used a space‐for‐time substitution, applying the learned SOC‐environment relationships to time‐varying covariates to infer temporal SOC patterns across ∼360 km 2 of agricultural landscapes on Maui. The model inferred widespread SOC decline from 2017 to 2024, with former sugarcane lands declining the most (−0.96 ± 0.49 percentage points in 0–20 cm soil depth). Climate significantly modulated inferred SOC decline in abandoned pineapple lands: drier zones lost twice the carbon of wetter zones, an effect absent where pineapple was converted to managed rangeland. Our transformer‐based approach provides a scalable, transferable framework for inferring carbon dynamics in data‐sparse agricultural landscapes with complex environmental gradients and land use histories.
Source: Earth s Future
DOI: 10.1029/2026ef008563
Language: English
Subfield: Soil Science
Domain: Life Sciences
Open Access status: gold
Estimated APC paid: $3,180
Awards: G3120126014, G6232481075
OpenAlex ID: W7212371375
