Data-Centric AI for Tropical Agricultural Mapping: Challenges, Strategies and Scalable Solutions

Mateus Pinto da Silva, Sabrina P. L. P. Correa, Hugo N. Oliveira, Ian M. Nunes, Jefersson A. dos Santos · 2025

Read Data-Centric AI for Tropical Agricultural Mapping: Challenges, Strategies and Scalable Solutions by Mateus Pinto da Silva, Sabrina P. L. P. Correa, Hugo N. Oliveira, Ian M. Nunes, Jefersson A. dos Santos (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2510-16207.txt.

Introduction

Mateus Pinto da Silva and four co-authors argue that mapping tropical agriculture from satellites is mainly a data problem, not a model problem. Persistent clouds, shifting crop calendars, scarce labeled examples, and high labeling costs all undermine the usual model-centric approach, which keeps tuning architectures while the training data stay weak.

Surveying the field from a data-centric standpoint, the authors assess twenty-five strategies, among them confident learning, core-set selection, augmentation, and active learning, rating each for readiness in large mapping pipelines. They finish with a practical pipeline built from the nine most mature and straightforward methods. The tutorial targets teams who need robust, scalable crop maps under real tropical conditions rather than laboratory benchmarks.

Catalog record

CREATORMateus Pinto da Silva, Sabrina P. L. P. Correa, Hugo N. Oliveira, Ian M. Nunes, Jefersson A. dos Santos
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE24 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2510-16207

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Data-Centric AI for Tropical Agricultural Mapping: Challenges, Strategies and Scalable Solutions Thanks: The authors would like to thank Serrapilheira Institute (grant #R-2011-37776), Institute of Artificial and Computational Intelligence (IDATA), Brazil National Council for Scientific and Technological Development (CNPq) -grant #141203/2025-0, National Laboratory of Scientific Computation (LNCC), part of the Ministry of Science, Technology, Innovations and Communications (MCTIC) for providing HPC resources of the SDumont supercomputer, The Brazilian Institute of Geography and Statistics (IBGE), linked to Ministry of Planning and Budget (MPO), and Brazilian Federal Agency for Support and Evaluation of Graduate Education (CAPES) for their financial support for this research. Mateus Pinto da Silva, Sabrina P. L. P. Correa, Hugo N. Oliveira, Ian M. Nunes, Jefersson A. dos Santos Abstract Mapping agriculture in tropical areas through remote sensing presents unique challenges, including the lack of high-quality annotated data, the elevated costs of labeling, data variability, and regional generalisation. This paper advocates a Data-Centric Artificial Intelligence (DCAI) perspective and pipeline, emphasizing data quality and curation as key drivers for model robustness and scalability. It reviews and prioritizes techniques such as confident learning, core-set selection, data augmentation, and active learning. The paper highlights the readiness and suitability of 25 distinct strategies in large-scale agricultural mapping pipelines. The tropical context is of high interest, since high cloudiness, diverse crop calendars, and limited datasets limit traditional model-centric approaches. This tutorial outlines practical solutions as a data-centric approach for curating and training AI models better suited to the dynamic realities of tropical agriculture. Finally, we propose a practical pipeline using the 9 most mature and straightforward methods that can be applied to a large-scale tropical agricultural mapping project. I Introduction Data-Centric Artificial Intelligence (DCAI) is an approach to developing AI systems that prioritizes improving the quality of data rather than focusing solely on enhancing models or algorithms [ 1 ] . Traditionally, the focus of Machine Learning (ML) development relies on creating and enhancing increasingly sophisticated models (model-centric). As emphasized in [ 2 ] , modern models are already powerful enough and, in many cases, the bottleneck lies in the quality of the data. Some studies have explored the use of Remote Sensing (RS) imagery from a data-centric perspective [ 3 ] . However, there is still a noticeable lack of research specifically addressing DCAI approaches for agricultural applications [ 4 ] . The existing literature largely theorizes about how data should be managed, often proposing approaches that are impractical or prohibitively expensive to implement. In light of this, we provide a review of DCAI methodologies that are suitable for large-scale agricultural mapping applications, with a particular focus on their applicability in tropical regions. We highlight the critical role of tropical countries (such as Brazil, Indonesia, and Nigeria) in global agriculture [ 5 ] , and argue that the scarcity of datasets from these regions [ 6 ] can be effectively addressed through data-centric AI strategies. Given the scarcity of datasets for tropical regions, we emphasize methods that can be employed for segmentation, noting that parcel delineation—typically addressed through semantic or instance segmentation—is a fundamental step in building large-scale agricultural crop mapping pipelines and datasets [ 6 ] . These methods, however, are not limited to this task and can be readily extended to other applications, such as land use/land cover (LULC) classification and crop mapping. In this sense, this study aims to present a pipeline for large-scale tropical agricultural mapping using DCAI. We first discuss about mapping in tropical agriculture (Sec. II ), then we present how tradicional assessment is conducted (Sec. III ). After that, we present the proposed pipeline (Fig. 1 ) in Sec. IV and concluding in Sec. V . [ 7 ] [ 8 , 9 ] [ 10 , 11 ] [ 12 ] [ 13 , 14 , 15 , 16 ] [ 17 ] [ 18 ] [ 19 , 20 , 21 , 22 , 23 ] [ 24 ] [ 25 , 13 , 14 , 15 ] [ 26 ] [ 27 , 28 , 29 ] [ 30 , 31 , 32 ] [ 33 ] [ 19 , 10 , 20 , 34 , 35 , 6 ] [ 36 ] [ 37 ] [ 38 ] [ 39 ] [ 40 ] [ 41 ] Fig. 1: Data-Centric AI pipeline, adapted from Roscher et al. [ 3 ] . DCAI approaches and their respective readiness in each step of the ML cycle are presented and contrasted with the traditional supervised ML pipeline. Letters A through D are further discussed in Sections IV-A , IV-B , IV-C , IV-D , respectively. We chose not to include and cite methods with very low levels of readiness due to space constraints. II Mapping Tropical Agriculture Agricultural mapping using Deep Learning (DL) techniques applied to remote sensing images has gained prominence in large-scale projects [ 6 ] . One of the main bottlenecks in the process is obtaining high-quality annotations for supervised training. Although the delineation of plots by experts provides a solid basis, imprecise boundaries can introduce noise into the data in semantic segmentation tasks. In addition, uneven frequency of classes may bias the model, requiring rectifying strategies such as resampling and weighted metrics [ 42 ] , specially in classification tasks such as Crop Mapping. Another critical challenge is the generalization of these models, which frequently exhibit a decline in performance when applied to real-world scenarios involving new regions with distinct agroecological characteristics. This problem is exacerbated by the fact that most of the DL agricultural mapping literature has been developed for temperate regionswith strong seasonality, well-defined agricultural calendars, and often annual crops in predictable cycles [ 6 ] , which do not occur for tropical areas. In tropical regions, the agricultural calendar is more flexible and crops are more varied in terms of both species and management practices [ 43 ] . This temporal and spatial variability makes it difficult to apply models based on regular time series, widely used in temperate regions [ 44 , 45 ] . In addition, high cloudiness for much of the year, especially in the tropical rainforests and Southeast Asia [ 46 ] , limits the availability of consistent optical imagery. This underscores the need to leverage multi-source data (such as combining radar and optical images) and to develop models that are robust to data gaps. Synthetic Aperture Radar (SAR) guarantees consistent observations even in adverse weather conditions. Along these lines, Cue et al. [ 47 ] demonstrated that convolutional networks with multitemporal SAR data, combined with prior knowledge about crop dynamics in Brazil, offered gains of up to 8.7% in F1-score for different crops in Brazil. III Traditional Data Quality Assessment Data quality can be observed as the degree in which a product is adequate to its respective specifications. In specific for geospatial data, there is the ISO 19157:2013 [ 48 ] , which defines elements for spatial data quality 1 1 1 The elements for spatial data quality are: thematic accuracy, completeness, logical consistency, temporal accuracy, positional accuracy and usability. . There are two main types of geospatial data: raster images and vector data. Traditional approaches can be generalized as data-agnostic in the sense

Read the complete text