PLANTEUR: A Crop-Type Segmentation Benchmark for Tropical Regions

Piper Wolters; Gabriel Tseng; Favyen Bastani; Joseph Redmon; Yawen Zhang; Hadrien Sablon; Ando Shah

cs.CVNov 3, 2026

PLANTEUR: A Crop-Type Segmentation Benchmark for Tropical Regions

Authors: Piper Wolters, Gabriel Tseng, Favyen Bastani, Joseph Redmon, Yawen Zhang, Hadrien Sablon, Ando Shah

Organizations: Allen Institute for AI

Abstract

We introduce PLANTEUR, a benchmark for crop-type mapping from satellite images in the French overseas territories. Building on the PASTIS design, PLANTEUR provides georeferenced and aligned Sentinel-2, Sentinel-1, and Landsat-8 sequences with parcel-derived labels spanning six tropical regions and an 8-class label space. We define both a standard test split and a leave-one-island-out protocol to measure cross-territory generalization, and evaluate fully supervised models, linear probes, and fine-tuned foundation models. The benchmark code and data are open source and available on GitHub and Hugging Face.

Figures & tables

Figure 1: PLANTEUR pixels illustrated on maps of the six tropical regions included in the dataset. Note that the whitespace represents pixels that do not include labels, whether due to being non-agricultural (e.g., mountains) or labels not existing in the data source.

Figure 2: PLANTEUR class distribution, in terms of number of windows containing that class, number of parcels for that class, and number of pixels labeled as that class. The background class represents non-agriculture, and the void class consists of underrepresented tropical classes.

Figure 3: Example parcels for each scored crop class in PLANTEUR. Each chip is a Sentinel-2 RGB composite from a single acquisition date, with the ground-truth mask showing just that class’s pixels.

Island	test windows	test pixels	Background %	Trivial mIoU
Réunion	172	2,805,239	71.02%	8.88
Guadeloupe	69	1,130,352	85.99%	10.75
Martinique	58	950,272	84.20%	10.52
Guyane	80	1,301,548	78.83%	9.85
Mayotte	64	1,043,953	99.45%	12.43

Table 1: Per-island composition of the 443-window test split, broken down by number of windows and pixels. Note that Saint-Martin has just six windows in the train set and does not contribute to the test set.

Family		Model	mIoU	acc.
Linear probe	Released embeddings	AEF	$27.68$	$89.50$
		Tessera v1.1	$32.78$	89.84
		Tessera v1	33.13	$89.77$
	Frozen encoders	AnySat	$28.79$	$86.59$
		Galileo	$34.71$	$89.91$
		Presto	$16.25$	$80.15$

Table 2: Results on the PLANTEUR 443-window test split. Best within each block in bold, best overall highlighted.

Figure 4: MACs vs. mIoU on PLANTEUR standard test set, separated into two plots for linear probe and supervised methods. The reported MACs represent compute cost of forward pass, so released embedding products wouldn’t bear this cost. The dotted line marks AEF, a closed-weight embedding product for which encoder MACs are unavailable.

Model	S2	S1	S1+S2	S2+LS	S1+S2+LS
OlmoEarth v1.3 (64d)	35.17	24.64	34.95	34.90	35.01
OlmoEarth v1.3 (128d)	36.88	24.45	37.04	36.68	36.02
OlmoEarth v1.3 (768d)	38.94	26.91	39.39	38.50	39.03
OlmoEarth v1.2 FT	40.82	32.41	41.62	40.42	41.01

Table 3: OlmoEarth varying inputs on the 443-window PLANTEUR test split. Best input combination per row in bold.

Figure 5: Temporal sequence-length ablation for OlmoEarth v1.3 linear probes and OlmoEarth v1.2 fine-tuning: mIoU-8 as a function of the number of monthly mosaics provided as input. Performance improves with a larger sequence length, though the rate of increase is largest between a sequence length of one and three for all methods. OlmoEarth v1.3 probes in this ablation use a fixed learning rate of 0.1 rather than a validation-selected sweep.

Model	$w=4$	$w=8$	$w=16$	$w=32$	$w=64$
OlmoEarth v1.3 (64d)	33.21	34.51	35.17	36.82	37.56
OlmoEarth v1.3 (128d)	34.69	37.59	36.88	38.20	37.10
OlmoEarth v1.3 (768d)	35.92	36.16	38.94	40.20	39.36
OlmoEarth v1.2 FT	35.82	38.67	39.81	40.65	40.82

Table 4: Spatial-context ablation on the 443-window PLANTEUR test split. Best window size per row in bold.

Model	baseline	balanced
OlmoEarth v1.3 (64d)	35.17	36.74
OlmoEarth v1.3 (128d)	36.88	39.13
OlmoEarth v1.3 (768d)	38.94	40.91
Tessera v2	34.20	37.42
AEF	27.68	28.71
U-TAE	35.82	33.98

Table 5: Baseline vs. balanced, where the balanced setup reduces the number of Background samples to equal the total number of crop samples, evaluated on PLANTEUR 443-window test split. Best within each block in bold, best overall highlighted. The difference in performance between columns is not large, suggesting that the task difficulty is beyond dealing with class imbalance.

Model	Réunion	Guadeloupe	Martinique	Guyane	Mayotte	mean
OlmoEarth v1.3 (64d)	22.70	30.38	29.93	14.12	12.49	21.92
OlmoEarth v1.3 (128d)	26.43	31.16	35.16	15.50	12.42	24.13
OlmoEarth v1.3 (768d)	13.84	33.90	32.69	13.95	12.48	21.37
Tessera v2	11.50	28.79	29.57	12.26	12.43	18.91
AEF	9.88	23.08	20.77	12.75	12.42	15.78
U-TAE	22.34	28.43	30.25	14.52	12.52	21.61

Table 6: PLANTEUR leave-one-island-out (LOIO) results. Mean is across the 5 folds. Best within each block in bold, best overall highlighted.

Model	$X=10$	$X=25$	$X=100$	$X=1000$
OlmoEarth v1.3 (64d)	20.57	21.06	25.23	27.67
OlmoEarth v1.3 (128d)	21.80	21.50	26.96	28.88
OlmoEarth v1.3 (768d)	20.75	20.96	27.57	29.66
Tessera v2	18.61	18.28	23.51	27.23
AEF	16.16	17.63	21.93	24.25
U-TAE	5.84	8.14	14.49	22.59

Table 7: Few-shot on PLANTEUR labels only: train on

X

labeled PLANTEUR pixels per class, evaluate on the 443-window test split. Best within each block in bold, best overall highlighted.

Figure 6: PLANTEUR few-shot transfer: mIoU-8 against $X$ labeled PLANTEUR pixels per class, on top of PASTIS.

Model	Regime	baseline	inv	invsqrt	focal $\gamma=2$	invsqrt+focal
U-TAE	supervised	35.82	28.82	36.85	36.51	36.68
TSViT	supervised	36.33	27.55	35.76	36.45	34.48
OlmoEarth v1.2	fine-tune	40.82	33.16	41.10	40.21	39.34
OlmoEarth v1.3 (768d)	linear probe	38.94	34.93	41.43	38.61	40.49

Table 8: Class-weighting and focal-loss ablations on PLANTEUR (443-window test split), mIoU-8. Best per row in bold.

Model	Arm	$X=10$	$X=25$	$X=100$	$X=1000$
OlmoEarth v1.3 (64d)	all PASTIS	24.64	26.61	27.73	27.77
OlmoEarth v1.3 (64d)	balanced	24.51	23.31	26.39	27.78
OlmoEarth v1.3 (128d)	all PASTIS	27.57	26.98	28.58	29.29
OlmoEarth v1.3 (128d)	balanced	27.26	24.98	27.74	29.18
OlmoEarth v1.3 (768d)	all PASTIS	27.54	26.04	29.30	28.92
OlmoEarth v1.3 (768d)	balanced	24.55	23.31	29.00	30.04

Table 9: Unseen-region transfer: PASTIS plus

X

labeled PLANTEUR pixels per class. all PASTIS keeps the full PASTIS training set; balanced draws

X

pixels of every class from both sources. Best within each block in bold, best overall highlighted.

Model	mIoU-8	trop-4	acc.
OlmoEarth v1.3 (64d)	28.28	22.53	89.05
OlmoEarth v1.3 (128d)	31.93	30.86	87.53
OlmoEarth v1.3 (768d)	35.82	33.99	90.73
OlmoEarth v1.2 FT	37.83	35.65	90.13

Table 10: Unseen transfer: encoders fine-tuned on PASTIS, then linear-probed on the PLANTEUR 443-window test split. The table shows mIoU-8 and mIoU over just the four tropical classes absent from PASTIS (Sugarcane, Banana, Pineapple, Tropical tuber; trop-4). Each cell is the best of an 8-point learning-rate sweep selected on validation.

Model	Inputs	Réunion	Guadeloupe	Martinique	Guyane	Mayotte	mean
OlmoEarth v1.3 (64d)	S2	22.70	30.38	29.93	14.12	12.49	21.92
OlmoEarth v1.3 (64d)	S1	15.08	21.39	22.31	11.86	12.41	16.61
OlmoEarth v1.3 (64d)	S1+S2	20.72	30.10	29.45	14.00	12.43	21.34
OlmoEarth v1.3 (64d)	S2+LS	22.98	30.03	28.92	13.46	12.61	21.60
OlmoEarth v1.3 (64d)	S1+S2+LS	21.27	29.99	29.10	14.03	12.41	21.36
OlmoEarth v1.3 (128d)	S2	26.43	31.16	35.16	15.50	12.42	24.13

Table 11: Leave-one-island-out (LOIO) input ablations for OlmoEarth. Best mIoU-8 within each read-out block in bold, best overall highlighted, except for the Mayotte column which is not meaningful.

Probe	$w$	Réunion	Guadeloupe	Martinique	Guyane	Mayotte	mean
OlmoEarth v1.3 (64d)	4	20.57	29.29	26.78	11.70	12.40	20.15
OlmoEarth v1.3 (64d)	8	20.42	29.12	29.65	13.48	12.52	21.04
OlmoEarth v1.3 (64d)	16	22.70	30.38	29.93	14.12	12.49	21.92
OlmoEarth v1.3 (128d)	4	20.25	29.48	28.10	13.43	12.42	20.74
OlmoEarth v1.3 (128d)	8	19.63	30.83	32.30	13.35	12.53	21.73
OlmoEarth v1.3 (128d)	16	26.43	31.16	35.16	15.50	12.42	24.13

Table 12: Spatial-context ablation under LOIO protocol. Mean mIoU-8 is across the 5 folds. Best mean per section in bold.

Model	S2	S1	S1+S2	S2+LS	S1+S2+LS
OlmoEarth v1.3 (64d)	90.55	84.64	90.21	90.47	90.26
OlmoEarth v1.3 (128d)	90.46	85.76	90.47	90.30	90.61
OlmoEarth v1.3 (768d)	90.12	86.66	90.34	89.97	90.20
OlmoEarth v1.2 FT	91.80	89.49	91.63	91.50	91.40

Table 13: OlmoEarth varying inputs on the 443-window PLANTEUR test split, overall pixel accuracy (%). Same runs as Table 3 . Best input combination per row in bold.

Model	$w=4$	$w=8$	$w=16$	$w=32$	$w=64$
OlmoEarth v1.3 (64d)	89.85	90.24	90.55	90.16	89.55
OlmoEarth v1.3 (128d)	90.28	90.36	90.46	90.13	90.06
OlmoEarth v1.3 (768d)	88.89	91.01	90.12	91.18	91.09
OlmoEarth v1.2 FT	89.52	90.32	90.67	91.39	91.80

Table 14: Spatial-context ablation on the 443-window PLANTEUR test split, overall pixel accuracy (%). Same runs as Table 4 . Best window size per row in bold.

Model	baseline	balanced
OlmoEarth v1.3 (64d)	90.55	89.02
OlmoEarth v1.3 (128d)	90.46	90.18
OlmoEarth v1.3 (768d)	90.12	90.20
Tessera v2	90.26	89.69
AEF	89.50	87.74
U-TAE	90.62	89.86

Table 15: Baseline vs. balanced, overall pixel accuracy (%), where balanced reduces the number of Background samples to equal the total number of crop samples, evaluated on the PLANTEUR 443-window test split. Same runs as Table 5 . Best within each block in bold, best overall highlighted.

Model	$X=10$	$X=25$	$X=100$	$X=1000$
OlmoEarth v1.3 (64d)	69.16	58.42	68.51	70.59
OlmoEarth v1.3 (128d)	69.99	56.91	69.60	70.82
OlmoEarth v1.3 (768d)	57.76	51.71	66.04	68.33
Tessera v2	59.06	47.76	62.58	71.89
AEF	49.09	49.14	64.53	72.56
U-TAE	26.70	35.54	50.95	60.12

Table 16: Few-shot on PLANTEUR labels only, overall pixel accuracy (%): train on

X

labeled PLANTEUR pixels per class with no PASTIS data, evaluate on the 443-window test split. Same runs as Table 7 . Best within each block in bold, best overall highlighted.

Model	Réunion	Guadeloupe	Martinique	Guyane	Mayotte	mean
OlmoEarth v1.3 (64d)	83.84	87.92	87.85	83.31	99.35	88.45
OlmoEarth v1.3 (128d)	84.98	87.28	88.56	84.48	99.39	88.94
OlmoEarth v1.3 (768d)	75.03	88.73	89.12	82.75	99.25	86.98
Tessera v2	75.18	87.00	88.23	82.97	99.26	86.53
AEF	74.29	85.89	85.28	80.20	99.37	85.01
U-TAE	83.66	87.51	88.35	83.71	97.78	88.20

Table 17: PLANTEUR leave-one-island-out (LOIO), Sentinel-2 inputs: pixel accuracy (%). Mean is across the 5 folds. Best within each block in bold, best overall highlighted.

Figure 7: Qualitative PLANTEUR examples, with each row showing a Sentinel-2 mosaic, corresponding ground truth labels, and predicted mask from linear probing AEF, Tessera v2, and OlmoEarth v1.3.

Figure 8: Confusion matrices for methods trained on the imbalanced and balanced train set and evaluated on the 443-window PLANTEUR test split. Rows are ground-truth classes, columns predictions, row-normalized; darker cells carry more mass. Off-diagonal mass sits almost entirely in the Background column.

Agricultural Remote Sensing

Agricultural remote sensing uses satellite and airborne observations to map crops and fields and monitor yield, soil, and crop health.

See all 90 papers

Benchmarking Geospatial Foundation Models for Agriculture Applications
Jun 29, 2026Zhuocheng Shang, Sanmay Das, Ahmed EldawyRemote Sensing Image Segmentation Agricultural Remote Sensing
University of California, Riverside · Riverside, USA
PRUE: A Practical Recipe for Field Boundary Segmentation at Scale
Mar 28, 2026Gedeon Muhawenayo, Caleb Robinson, Subash Khanal +10Image Segmentation Remote Sensing Image Segmentation
Arizona State University · Microsoft AI for Good · Washington University in St. Louis +3
SwissCrop25: A National Multi-Year Benchmark for Operational Crop Mapping
Aug 10, 2026Thomas Lauber, Mehmet Ozgur Turkoglu, Sélène Ledain +1Agricultural Remote Sensing Land Cover Classification
Earth Observation of Agroecosystems Team, Agroscope, Switzerland

Jun 29, 2026J/K move · Enter open · S save

Benchmarking Geospatial Foundation Models for Agriculture Applications

Zhuocheng Shang, Sanmay Das, Ahmed Eldawy

University of California, Riverside · Riverside, USA

Remote Sensing Image Segmentation Agricultural Remote Sensing Geospatial Foundation Models Domain Generalization Remote Sensing Change Detection Remote Sensing Benchmarks

Open paper AskPDF arXiv

Geospatial foundation models pretrained on satellite imagery promise broad generalization across remote sensing tasks and regions, but their geographic transferability has not been systematically tested, especially in agriculture applications. This paper presents a controlled benchmark that evaluates three models, Prithvi, SpectralGPT, and SatMAE, on multi-temporal crop segmentation and change detection across four U.S. states, Iowa, North Carolina, California, and Minnesota. By assigning each train, validation, and test split to a separate region, we measure how well each model transfers to land it has not seen. All three degrade sharply under regional distribution shift, predicting only the most common crops while missing rare ones. We further find that fitting these models to a shared input format affects each one differently, which complicates direct architectural comparison. These results expose key limitations of current geospatial foundation models for agriculture and point to region aware evaluation as a necessary standard.

Abstract

Agricultural Remote Sensing

Explore similar work

Benchmarking Geospatial Foundation Models for Agriculture Applications

PRUE: A Practical Recipe for Field Boundary Segmentation at Scale

SwissCrop25: A National Multi-Year Benchmark for Operational Crop Mapping