Divide and conquer: using parcels to understanding agricultural activities
Conference
65th ISI World Statistics Congress 2025
Format: CPS Abstract - WSC 2025
Keywords: agricultural census, agriculture, deep-learning, machine learning, remote sensing, segmentation
Abstract
The Brazilian Institute of Geography and Statistics proposes using automated agriculture parcel recognition methods as a crucial step in large-scale monitoring since manual monitoring and parcel delineation are unfeasible due to the huge amount of resources needed. In this context, Machine Learning can be a useful tool for aiding in the detection, segmentation, and classification of the tiles.
Automated agricultural parcel identification delineates boundaries of fields with agricultural activities, classifying at the same time the interior of the parcel as having any agricultural activity identified or not. We are using Sentinel-2 satellite imagery with all available spectral bands. For the sake of this work, an agricultural parcel is a continuous and spatially homogeneous unit used for agricultural purposes developing one or more activities. The parcel identification can improve and aid in improving the characterization of subsistence and family farming which can be very important for policy makers in Brazil.
Neural Networks (NNs) are learning models particularly well-suited for this type of task, due to their inherent capabilities of processing large amounts of data. NNs, however, can face generalization hurdles when porting the model to out-of-distribution (OOD) remote sensing data, mainly in applications wherein domain shift must be compensated for. This is especially true for agricultural applications, where crop species, growth rates, biomes, and climate vary across different regions.
In this work, we leverage the representation learning capabilities of NNs to improve large-scale crop recognition in Brazil using a data-driven approach. Since we do not have data for Brazil yet, we are training and testing on public datasets available for a set of European countries to enable our method of producing features capable of representing the underlying data. This approach brings several undesired issues due to inherent distinctions between Brazil and European countries related to climate, territory aspects, and biomes. To mitigate the differences in data, we are producing a dataset for the Brazilian territory to consider the diversity of climate, topography, and biomes to fine-tune the pretrained model.
We employed a multitask approach, using convolutional NN architectures (e.g. FCNs, Encoder-Decoders, Mask R-CNNs, etc) to perform regression, semantic, and instance segmentation enabling the identification and delineation of the continuous agricultural regions, leveraging the spatial similarity with the temporal and spectral aspects of the remote sensing data.
We split our dataset into three distinct subsets: train, validation, and test, and the proposed methodology can produce crop parcel predictions with over 76% of mean Intersection over Union (mIOU) on the test subset. Figure 1 shows an example of produced parcels compared with the ground truth.