Divide and conquer: using parcels to understanding agricultural activities
Conference
65th ISI World Statistics Congress
Format: CPS Paper - WSC 2025
Keywords: agricultural census, agriculture, deep-learning, machine learning, remote sensing, segmentation
Session: CPS 82 - Agricultural Statistics — Survey Methods
Monday 6 October 5:10 p.m. - 6:10 p.m. (Europe/Amsterdam)
Abstract
The emergence of Deep Learning techniques combined with Big Data al-
lows for the extraction of high-level semantic features directly from massive
volumes of data. While this paradigm has revolutionized Computer Vision
using natural RGB images, specialized domains such as Remote Sensing of-
ten lack massive labeled datasets. Consequently, training neural networks
in these fields requires methods designed to go beyond standard supervised
learning. In this context, the Brazilian Agricultural Census presents both
a significant challenge and a unique opportunity, covering 8.5 million km2
and over 5 million establishments. This work details a strategy using Remote
Sensing and Computer Vision for automated agricultural plot delineation. By
utilizing Census outcomes to guide crop mapping and yield estimation, while
simultaneously using automatic field boundary delineation as input for Census
fieldwork, this framework enables the collection of high-quality ground truth
data. The proposed methodology is capable of transforming and improving
the production of official statistics across Brazil’s diverse biomes. To demon-
strate the robustness of the approach, we provide a detailed analysis of results
from two representative municipalities.