iEBT Transformer jointly retrieves crop health metrics from satellite data

A new AI architecture from researchers at Michigan State University and Banaras Hindu University tackles a classic remote sensing challenge: inferring multiple crop and soil properties simultaneously from satellite data. The model, called the Iterative Energy-Based Transformer (iEBT), combines Sentinel-1 C-band radar and Sentinel-2 multispectral imagery to predict surface soil moisture (SM), leaf area index (LAI), and plant height (PH) in wheat fields. Unlike typical feedforward networks, iEBT treats retrieval as an optimization problem—it starts with an initial guess and iteratively updates the state vector using gradient descent on a learned energy function. This energy-based approach allows the model to resolve ambiguities that arise when soil moisture and canopy density change concurrently.

Tested on 700 quality-controlled ground samples from Varanasi, India, iEBT achieved a four-seed mean R² of 0.854 across all three targets, with individual R² values of 0.841 for SM, 0.905 for LAI, and 0.821 for PH. Ablation studies confirmed that Sentinel-1 drives SM retrieval, Sentinel-2 dominates LAI, and PH requires both. A standout feature is the model’s terminal energy value, which acts as an uncalibrated quality score—filtering out the top 10% highest-energy samples significantly reduces root-mean-square errors. While cross-season validation revealed domain shift challenges from local management practices, the compatibility-guided fusion and built-in diagnostic path make iEBT a promising step toward reliable, field-scale crop monitoring from space.