INTEGRATING SPECTRAL AND TEXTURE INDICES WITH MACHINE LEARNING FOR RICE LEAF AREA INDEX (LAI) ESTIMATION IN SUPHAN BURI, THAILAND

Kritchayan INTARAT , Paramaporn NETSAWANG , Supatthra NARAWATTHANA , Chayapol PROMAOH , Saruda CHUENKAMOL

DOI: 10.21163/GT_2025.201.15

ABSTRACT: This study presents unmanned aerial vehicle (UAV)-based remote sensing to estimate the leaf area index (LAI) in rice fields, across multiple growth stages, by comparing univariate and multivariate methodologies. Univariate models focus on evaluating the performance of individual vegetation indices (VIs), color indices (CIs), and normalized difference texture indices (NDTIs). In contrast, multivariate approaches employ feature selection via recursive feature elimination (RFE) with ensemble machine learning (ML) models: random forest regressor (RFR), gradient boosting regressor (GBR), and eXtreme gradient boosting regressor (XGBR). Dimensionality reduction techniques such as principal component regression (PCR) and partial least squares regression (PLSR) are also employed. Results demonstrate that when associated with stage-specific complexities, NDTI(REsem, Rcor), using a univariate model (XGBR) achieves the highest R² value of 0.91 and records the lowest RMSE at 0.38, in the panicle initiation stage. Across all stages, NDTI(Rmean, Bdis) is seen to deliver a high R² value of 0.81 while reaching a higher value of 1.20 RMSE than in the specific stages. When associated with multivariate models, CIs indicate the highest accuracy associated with RFE. In this case, five features, out of ten selected, attained 0.88 R² and 0.94 RMSE. These findings highlight the potential of integrating UAV-based multivariate approaches in precision agriculture for improved, stage-specific LAI estimation.

Keywords: Vegetation indices; Color indices; Texture indices; Leaf Area Index; UAV; Machine learning

Full article here