This study mapped 2024 rubber distribution and age structure across Northeast Thailand using Random Forest in Google Earth Engine, combining Sentinel-1, Sentinel-2, vegetation indices, and a GEDI-trained canopy-height model. Three-class rubber mapping reached 82.63% overall accuracy and about 0.84 F1; three-group age classification reached 83.39%. Plantations older than seven years dominated and were classified more reliably than young stands. The output is a regional monitoring baseline, not an error-free parcel-level decision map.
Key findings
- Rubber, other-crop, and non-crop classification reached 82.63% overall accuracy and 0.8481 macro F1. Precision was lowest for other crops at 0.7513, reflecting heterogeneous spectral signatures.
- The canopy-height model based on 6,880 GEDI footprints produced RMSE 3.00 m and MAE 1.93 m; rubber-age classification reached 83.39% overall accuracy.
- Recall for plantations under five years was 0.6590, with 969 of 1,481 points correct. Recall for plantations over seven years was 0.9113, with 2,182 of 2,400 correct, showing strongly age-dependent reliability.
- Mapped concentration was highest in Bueng Kan, Loei, and Nong Khai, and provincial area estimates correlated with Land Development Department data at about R²=0.95. This is an in-region comparison, not external validation in a new geography.
Why this matters globally
Fusing radar, multitemporal optical imagery, and spaceborne LiDAR can improve monitoring of tropical commodities under persistent cloud and support area, yield, biomass, and carbon assessment. Policy use should include uncertainty maps and decisions appropriate to 10–20 m resolution. Transfer to other regions or countries requires new field validation.
Thai researcher contribution
Eight Thailand-affiliated authors connect Mahasarakham, Khon Kaen, Prince of Songkla, and the Asian Institute of Technology. Jaturong Som-Ard leads the Earth Observation Technologies for Land and Agricultural Development unit at Mahasarakham University. Mahasarakham University funded the work under grant 6911009. The publisher page does not expose individual CRediT roles, so unsupported task-level attribution is avoided.
Limitations to consider
Accuracy of roughly 82–83% still permits errors material to parcel-level use, especially for other crops, non-crops, and plantations under five years. Sentinel’s 10–20 m pixels mix soil, weeds, and canopy, while GEDI footprints are discontinuous and may have 15–30 m geolocation error. High-dimensional fusion raises overfitting risk, and young-plantation samples did not capture full variability. Validation covered one year in Northeast Thailand rather than other years or regions, and canopy height was not directly modelled against plantation age.