Abstract

Mongolia’s vast territory, dispersed agricultural production, and limited field capacity support the integration of remote sensing and machine learning into agricultural census validation. This study synthesizes national pilots combining satellite imagery, drone data, field spectrometry, and machine-learning techniques to validate crop and livestock statistics for the 2030 agricultural census round.
For crop validation, cadastral and historical records were georeferenced by field location and crop type and integrated with Sentinel-2 imagery, multispectral drone data, and field spectrometer measurements. A Random Forest classifier, trained on 75% of observations and tested on 25%, achieved 76.9% overall accuracy for wheat, rapeseed, and potato. Where crop-specific reference data were limited, Spectral Angle Mapper achieved 92.9% accuracy for cabbage, beetroot, and oats, although performance depended on timely measurements aligned with crop phenological stages.
For livestock validation, the National Statistics Office of Mongolia, PARIS21, and Masae Analytics piloted high-resolution imagery in Arkhangai province in 2024. A model using 30–50 cm satellite imagery detected large livestock—cattle, horses, and camels—with 81.63 percent accuracy. A related model identified traditional dwellings with 98.11 percent accuracy, supporting the identification of herder household locations and future livestock estimation. Small livestock were better suited to drone-based observation. During the 2024 year-end and 2025 mid-year livestock census operations, 5,630 videos were collected and converted into images at five-second intervals. Sheep and goats were annotated in 116,162 images, providing a training dataset for machine-learning models and a prototype web application that counts animals from pasture video. Further development focuses on enclosure detection, improved counting accuracy, large livestock, and summer pasture conditions.
The findings show that satellite imagery can support wide-area monitoring and targeted field verification, whereas drone images provide detailed reference data and independent checks of census responses. Operational constraints include the cost of very-high-resolution imagery and UAV deployment, cloud contamination, limited historical training data, seasonal sensitivity, equipment shortages, labour-intensive annotation, and insufficient image-processing expertise. A hybrid framework combining census observations, satellite coverage, targeted UAV missions, and machine-learning validation can improve accuracy, transparency, and field efficiency while complementing enumerator-based data collection.