Abstract
Cassava, a globally significant crop providing food security for nearly 1 billion people as a dietary staple and driving industrial growth through biofuel and starch production, remains challenging to map accurately due to spectral confusion with similar crops and complex intercropping systems in fragmented smallholder landscapes. This study presents the first comprehensive regional-scale, high-resolution (10 m) cassava mapping framework for Guangxi Zhuang Autonomous Region (236,700 km2), China’s largest cassava-producing region. We developed a novel multi-sensor phenology-based approach integrating two methodological advances: 1) the Cassava-Sugarcane Difference Index (CSDI), achieving superior crop discrimination (Jeffries-Matusita distance: 1.91 vs. 1.25–1.61 for conventional indices), and 2) an automated detection system that identifies and quantifies cassava intercropping patterns (with maize, peanuts, and watermelon) at regional scale using dual-season phenological analysis. By harmonizing Landsat 8/9 and Sentinel-2 time series, we maintained >70% monthly cloud-free coverage despite persistent subtropical cloud cover. Validation demonstrated robust performance across all years: 91.52% overall accuracy for 2023 (990 field samples) and 85.60%–89.22% for 2019–2022 (∼400 samples per year from retrospective high-resolution imagery), with strong correlation to agricultural statistics (R2 = 0.71–0.80). The results reveal that 80.9% of cassava follows monoculture practices while 19.1% employs intercropping systems (maize: 14.4%, peanut: 3.4%, and watermelon: 1.3%), with 90% of plantations occurring in fragmented plots <2.32 ha. The documented 12% cultivation decline in cultivation area (2019–2023) from 1.25 × 105 to 1.10 × 105 ha provides critical evidence for agricultural policy evaluation. This framework provides a transferable methodological approach for cassava monitoring in complex tropical agricultural systems, with direct applications for food security assessment and agricultural policy development across regions facing similar spectral confusion and cloud cover challenges, though local calibration enhances performance.