Embedding -based Crop Type Classification in the Groundnut Basin of Senegal

Madeline C. Lisaius, Srinivasan Keshav, Andrew Blake, Clement Atzberger · 2026

Read Embedding -based Crop Type Classification in the Groundnut Basin of Senegal by Madeline C. Lisaius, Srinivasan Keshav, Andrew Blake, Clement Atzberger (2026) 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-2601-16900.txt.

Introduction

Madeline Lisaius, Srinivasan Keshav, Andrew Blake, and Clement Atzberger map crops in Senegal from space shots. Crop maps back food plans and aid for small farms. Yet most space-based maps fail on small farm plots. Plots there are tiny, mixed, and shift oft. Their 2026 preprint works in the Groundnut Basin of Senegal. It builds maps from learned shot codes, called embeddings.

Codes pack each shot's traits in a compact form. The team tests how well codes sort crop types. Small farm zones need maps that fit their scale. Sound maps can guide seed, water, and aid plans. They also track how rains shift farm health.

Catalog record

CREATORMadeline C. Lisaius, Srinivasan Keshav, Andrew Blake, Clement Atzberger
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE44 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2601-16900

Browse the full library · About the Tower of Babel

Full text

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

Embedding -based Crop Type Classification in the Groundnut Basin of Senegal Madeline C. Lisaius, Srinivasan Keshav, Andrew Blake, Clement Atzberger Abstract Crop type maps from satellite remote sensing are important tools for food security, local livelihood support and climate change mitigation in smallholder regions of the world, but most satellite-based methods are not well suited to smallholder conditions. To address this gap, we establish a four-part criteria for a useful embedding-based approach consisting of 1) performance, 2) plausibility, 3) transferability and 4) accessibility and evaluate geospatial foundation model (FM) embeddings -based approaches using TESSERA and AlphaEarth against current baseline methods for a region in the groundnut basin of Senegal. We find that the TESSERA -based approach to land cover and crop type mapping fulfills the selection criteria best, and in one temporal transfer example shows 28% higher accuracy compared to the next best method. These results indicate that TESSERA embeddings are an effective approach for crop type classification and mapping tasks in Senegal. 1 Introduction Reliable, wall-to-wall mapping of smallholder crops landscapes is an important tool for food security, climate adaptation, and sustainable land management globally. Approximately three-quarters of the world’s poor live in rural areas and rely primarily on small-scale agriculture [ 1 ] . These farms are critical not only for rural livelihoods but also for local and regional food security; globally, farms under 2 hectares (ha) contribute an estimated 30–34% of food supply, devote a greater proportion of their production to food, and maintain greater crop diversity than larger farms [ 2 ] . Knowing what crops are grown where allows for informed decision making at regional, national and global scales that can mean survival for vulnerable people. In west Africa, much of the agricultural landscape is dominated by smallholder systems, where small plots support the livelihoods of most rural households, yet these agricultural systems remain understudied relative to large farms in the Global North, making crop mapping more difficult [ 3 ] . Satellite remote sensing provides an opportunity to bridge research and mapping gaps by enabling consistent, large-scale, and low-cost monitoring of agricultural landscapes. Freely available sensor data such as from Sentinel-1 and Sentinel-2 offer regularly repeating observations that can be leveraged for crop type mapping [ 4 ] . However, most existing methods have been developed in contexts of industrial agriculture, where fields are large, monocropped, and well-documented, making them poorly suited for smallholder regions [ 5 ] . In West Africa, the challenges of frequent cloud cover, limited ground samples, small and irregular field sizes, and intercropping further complicate the creation of accurate, wall-to-wall crop type maps [ 6 , 7 ] . Recent advances in geospatial foundation models (FMs) offer a promising new approach. Embeddings derived from large-scale pre-trained models have shown strong performance on agricultural classification tasks, often matching or surpassing traditional machine learning approaches in accuracy and F1 score [ 8 , 9 , 10 ] . Yet, these methods have not been systematically evaluated under true smallholder conditions such as diverse management and intercropping which is common in West Africa. To address this gap, we evaluate the potential of embedding-based approaches for wall-to-wall crop cover classification in the groundnut basin of Senegal, a region representative of many smallholder agricultural systems in west Africa. We compare multiple crop type classification approaches across several years and assess their performance and practical utility for real-world agricultural mapping applications. 2 Recent work Some recent literature exists for wall-to-wall crop type mapping of agriculture in West Africa: Ibrahim et al. [ 6 ] employ Sentinel-1, Sentinel-2 and SkySat imagery for land and crop cover mapping in a smallholder, intercropped region of Nigeria where average field size ranges between 0.3 ha and 0.5 ha. They first create a wall-to-wall classification map of land cover using spectral temporal metrics (STMs) from both Sentinel-1 and Sentinel-2, informed by cropping season. They then use the the land cover map to mask for agriculture regions and classify their five agricultural crops of interest using Random Forest on the STMs. They use high resolution images from SkySat to verify. This work requires extensive crop-specific feature engineering to prepare data for use by RF, therefore requiring customization to adapt to new regions. Azzari et al. [ 11 ] use harmonic regressions to classify maize from non-maize in Malawi and Ethiopia using Sentinel-1 and Sentinel-2 images. Field sizes are not reported but study areas are noted as smallholder regions. Pixel-wise harmonic regressions along with rasterized weather data and topographic data are used as inputs to Random Forest classifier to distinguish maize and non-maize. This method relies on a high quality cropland mask to be applied beforehand, and relies on generation of the harmonic regression features for each year and region, and addition of other data such as topography and climate data. Other researchers using a similar harmonics approach also use it for few-class classification, which is unsuitable for mapping large, diverse crop regions and minority crops (Luan Pott 2022). Gumma et al. [ 12 ] classify nine classes of crops in Senegal for 2020 using Sentinel-2 data, across all field sizes. They compare Random Forest, CART, SVM, and spectral matching technique (SMT). They prepare NDVI for the time frame of interest, group the NDVI time series with k-means clustering, and use labeled data to create “ideal spectra" that are leveraged to classify the clusters by similarity to the ideal. They begin by classifying 10 land cover classes and 9 crop cover classes for all of Senegal. They find the spectral matching technique has the highest accuracy of the methods evaluated but requires extensive feature preparation. Rustowicz et al. [ 13 ] approach crop type classification and segmentation in Ghana, South Sudan and Germany. They compare a 3D U-Net and the 2D convolutional neural network (CNN) with convolutional long short term memory model (CLSTM) that incorporates both CNNs and recurrent neural networks (RNNs) for semantic segmentation of multi-temporal, multi-spatial satellite images from Sentinel-2. They find that their proposed approach has the highest accuracy of tested methods in Germany. However for the smallholder regions they find that Random Forest often matches the accuracy of their approach. While this method is focused primarily on segmentation, it does consider the smallholder context. It is challenged by high cloud coverage as well as limited and imbalanced labels. 3 Methods 3.1 Criteria To identify the best method for crop type mapping in a smallholder region like the groundnut basin of Senegal, we identify four qualities of a suitable wall-to-wall mapping approach: 1. Performance: A good approach must offer high accuracy and F1 scores in crop type classification. As high quality datasets are difficult to acquire and maintain in smallholder regions, we acknowledge that these metrics may be lower than those in regions with high data quality and availability. 2. Plausibility: A good approach will show mapping results that are plausible for the region of study. Maps across years should show consistency, and additionally, changes to the extent of cropland should match trends in cropland statistics. 3. Transferability: As many organizations concerned with smallholder agriculture in the Global South do not have the capacity or resources to collect ground truth labels every year, it’s important that a modeling approach

Read the complete text