How Satellites and AI Reveal Crop Rotations in Thailand?

Knowing where and what is grown is key for sustainable agriculture and future food security. In Thailand, mapping crops is tough: fields are tiny, fragmented, and the sky is often cloudy. To tackle this, researchers merged optical and radar satellite data with machine learning algorithms like Random Forest, Support Vector Machine, and Gradient Tree Boosting.

They gathered over 8,000 field reference points across three planting seasons in 2023, in the provinces of Chaiyaphum and Suphan Buri. Using this data, they produced detailed maps of 12‑13 crop types, including rice, sugarcane, and cassava. Random Forest performed best, achieving over 85% overall accuracy and more than 90% for rice and sugarcane.

The maps reveal rice, sugarcane and cassava as the dominant crops, giving farmers and policymakers timely insights for smarter crop management. This study shows how Earth observation from space, combined with AI, can transform agriculture, making it more resilient and ready for challenges like climate change.

The Random Forest (RF) model is an ensemble of decision trees that operates via bagging: each tree is trained on a random subset of data and variables, reducing variance and improving robustness against overfitting. In this study, RF benefited from the combination of optical (visible and infrared spectra) and radar (SAR backscatter) data, which provide complementary information: optical spectra capture plant photosynthetic response, while radar penetrates cloud cover and detects vegetation structure. Multitemporal fusion allowed tracking the full phenological cycle of crops, distinguishing germination, vegetative growth, and maturation phases.

Performance metrics include Overall Accuracy (OA) and the Kappa coefficient, with OA >85% and Kappa >0.80, indicating strong agreement between predicted classes and field observations. Moreover, ROC curves for the main classes (rice, sugarcane) show AUC >0.95, signifying excellent discriminative ability. Variable importance analysis highlighted that near‑infrared (NIR) bands and HH polarization of SAR were the most informative for distinguishing rice from sugarcane, due to differences in leaf structure and canopy density. This integrated approach can be scaled to other tropical regions where cloud cover limits exclusive use of optical data, paving the way for more reliable and cost‑effective agricultural monitoring systems.

Future prospects include adopting convolutional neural networks (CNN) for spatial pixel analysis and integrating real‑time weather data to predict crop yields with higher precision. Additionally, establishing an open, annually updated reference point repository would enhance model generalizability and foster international collaboration.