[Paper Review] Embedding -based Crop Type Classification in the Groundnut Basin of Senegal

Madeline Lisaius, Srinivasan Keshav+2 more|arXiv (Cornell University)|Jan 23, 2026

Remote Sensing in Agriculture1 citations

TL;DR

This paper evaluates embedding-based land cover and crop type classification using TESSERA and AlphaEarth in Senegal’s groundnut basin, showing TESSERA achieves strong performance and transferability with lower compute than baselines.

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.

Motivation & Objective

• Define four criteria (performance, plausibility, transferability, accessibility) for embedding-based crop mapping in smallholder regions.
• Compare geospatial foundation model embeddings (TESSERA and AlphaEarth) to baseline methods for land cover and crop type mapping in Senegal.
• Assess cross-year transferability of models and the practicality (compute and feature engineering) for real-world mapping.
• Provide wall-to-wall crop and land cover maps and analyze temporal changes in cropland in the study region.

Proposed method

• Prepare labeled polygon data from JECAM for Fatick and Niakhar in 2018, 2019, and 2021; harmonize labels across years.
• Generate inputs including cloud-masked Sentinel-1/2 time series with vegetation indices, spectral-temporal metrics (STMs), and embeddings (TESSERA 128-d, AlphaEarth 64-d).
• Train multiple classifiers (RF, XGBoost, SVM, LR, MLP) per input type and build ensemble from the top two heads per input type.
• Evaluate using accuracy, macro F1, and weighted F1 on withheld test sets; assess plausibility via wall-to-wall maps; test transferability by cross-year training/testing; analyze CPU efficiency and feature engineering needs.

Figure 1: Visualized are a) the location of the region of interest within Senegal and b) Sentinel-2 satellite imagery of the region of interest. This region is dominated by dry shrub land that has been converted into agricultural land and features a river delta in the south.

Experimental results

Research questions

• RQ1 Do embedding-based approaches (TESSERA and AlphaEarth) outperform traditional baselines for wall-to-wall crop type classification in a smallholder West African region?
• RQ2 How do embeddings perform in cross-year transfer learning for land cover and crop type mapping in Senegal?
• RQ3 Are embedding-based methods more plausible, transferable, and accessible (in terms of compute and feature engineering) than traditional approaches in this context?
• RQ4 What are the spatial-temporal patterns of cropland and intercropping in Fatick/Niakhar across 2018, 2019, and 2021 as revealed by the best-performing methods?

Key findings

• TESSERA-ensemble and AlphaEarth-ensemble achieve high land cover accuracy across years, with TESSERA showing lower run-to-run variability than AlphaEarth.
• Cross-year transfer learning for land cover is feasible with embedding-based methods; TESSERA, AlphaEarth, and STM approaches show overlapping performance in some transfer scenarios.
• For crop type classification, embedding-based methods (especially TESSERA) yield higher mean performance than AlphaEarth across the three years, with notable year-to-year degradation linked to label quality differences.
• Embeddings reduce compute and feature engineering needs; TESSERA and AlphaEarth are more efficient than raw and STM inputs, aiding accessibility.

Figure 2: Pictured are three example label polygons, not at the same scale, showing a) bare soil, b) shrub land, and c) build-up surface. There is meaningful overlap in characteristics of ground cover amongst the three, particularly with build-up surface.