<![CDATA[The transformation of tropical landscapes due to agricultural expansion constitutes a significant global environmental challenge. Current land-cover classification methods, however, provide limited differentiation among agricultural management systems. This study develops an agriculture-focused land-cover classification workflow that fuses Landsat 9 optical imagery and PALSAR-2 L-band SAR across a ≈2,500 km² study area in Jambi Province, Sumatra, Indonesia, to enhance discrimination of crop systems and improve spatial coherence via object-based enhancement. A 22-class land-cover taxonomy was supported by 14,029 strategically collected training points. Feature engineering produced 29 predictor variables, including conventional vegetation indices, agricultural-specific metrics, water indicators, and SAR-derived structural features. Models were evaluated on an independent test dataset comprising 4,209 samples. An agriculture-weighted Random Forest classifier with strategic class weighting was implemented and followed by Simple Linear Iterative Clustering (SLIC) object-based enhancement to suppress speckle and enforce spatial contiguity. The classification achieved an overall accuracy of 53.7%, with exceptional performance for estate crop systems (F1 = 94%) and reliable forest discrimination. SLIC reduced salt-and-pepper noise by 99.5% and substantially improved spatial coherence metrics, transforming fragmented pixel-based outputs into operationally viable products. Despite these gains, discriminating smallholder mosaics remains challenging and likely requires additional temporal or higher-resolution inputs.]]>
Copyrights © 2025
