Remote Sensing · Time-Series · Deep Learning

Identify this season's crops from last season's signal.

AgriSpec AI is an automated machine-learning pipeline that classifies crop types for the current agricultural season by learning the multi-temporal spectral signatures — NDVI, EVI, NDWI, SWIR and thermal bands — captured by Sentinel-2 and Landsat-8 across the previous cropped season.

Explore the methodology View the pipeline

10–20 m

Sentinel-2

23 / season

Temporal steps

17 EU crops

Crop classes

92.4 %

Accuracy (OA)

01 — The problem

Field-level crop identity is the missing layer of every agri-model.

Slow cadastral surveys

Official declarations arrive 6–18 months after harvest — too late for yield forecasting, subsidy control or market intelligence.

Coarse global products

Generic land-cover maps (e.g. ESA WorldCover) mix fallow, pasture and 10+ crops under a single class.

Single-date ambiguity

A snapshot in July cannot separate maize from sorghum, or winter wheat from barley — phenology is required.

Climate-driven shift

Sowing windows and rotation patterns are shifting year on year; static look-up tables decay quickly.

02 — Core hypothesis

Last season's spectral rhythm is a transferable feature.

Each crop carries a near-identical "phenological fingerprint" across years — a recurring curve of green-up, peak, senescence and harvest. A model trained on a full season of Sentinel-2 composites generalises to the next season's observations without any in-season labels.

Spectral signature of a parcel

x = [NDVI(t₁), EVI(t₂), NDWI(t₃), …, B11(t₂₃)] ∈ ℝ²³ˣ⁸

Target — current season

y_t = crop class ∈ {1, …, 17}

Source — previous season

X_{t-1}, y_{t-1} (labelled LPIS / ground truth)

Transfer assumption

p(x_t | y_t) ≈ p(x_{t-1} | y_{t-1}) for stable climate years

03 — Methodology

A six-stage automated pipeline

From raw Sentinel-2 L1C tiles to a per-parcel crop label — fully unsupervised at inference time, fully reproducible.

Data acquisition

Sentinel-2 L2A (10 m) and Landsat-8/9 (30 m) are streamed from a STAC catalogue. Only scenes with < 20 % cloud cover over the AOI of interest are kept.

Atmospheric & cloud correction

Sen2Cor → BOA reflectances, SCL mask applied. Temporal interpolation with Whittaker smoother reconstructs a continuous 5-day composite stack per parcel.

Spectral index engineering

NDVI, EVI, NDWI, GCVI, LAI-proxy and tasselled-cap wetness/greenness/brightness are derived band-wise across 23 time steps.

Temporal feature extraction

A 1D-CNN + bidirectional GRU encoder compresses the 23 × 8 cube into a 128-d latent vector per parcel — the transferable signature.

Cross-season transfer

Domain-adversarial training (Gradient Reversal Layer) on previous-season labels aligns the latent space so current-season embeddings cluster identically.

Classification & QA

Soft-attention classifier outputs class probabilities + a per-parcel confidence flag. Parcels below 0.65 are routed to a human-in-the-loop queue.

04 — Model architecture

Temporal 1D-CNN + Bi-GRU + Domain-Adversarial head

The encoder is intentionally small — 380 k parameters — so it can be retrained per AOI on commodity GPUs in under 40 minutes. The adversarial head is what makes the model transfer across seasons.

› Input 23 × 8 — composite time series per parcel
› 3 × Conv1D (64-128-128) + BatchNorm + ReLU
› Bi-GRU (128, 2 layers, dropout 0.3)
› Soft-attention pooling → 128-d latent
› GRL + domain classifier (λ = 0.5)
› Class head: Dense(64) → Softmax(17)

05 — Empirical results

Outperforming single-season baselines by 11.8 pts of F1.

92.4 %

Overall accuracy

Across 17 crop classes, 2024 evaluation set

0.87

Macro F1-score

+0.118 vs. single-season baseline

0.65

Min recall class

Sugar beet — hardest minor crop

42 min

Train time / AOI

On a single NVIDIA T4 (15k parcels)

Per-class F1 — top 10 crops

Lower-bound 0.65, upper-bound 0.97.

Winter wheat 0.97

Maize 0.95

Barley 0.93

Rapeseed 0.92

Sunflower 0.89

Sugar beet 0.87

Soya 0.86

Potato 0.84

Sorghum 0.79

Rye 0.76

Confusion hotspots

Top confusable class pairs (rows → predicted).

Winter wheat ↔ Winter barley

4.1 % — close canopy phenology in April
Soya ↔ Sunflower

3.6 % — broadleaf summer crops
Maize ↔ Sorghum

2.9 % — same C4 growth curve
Rye ↔ Triticale

2.4 % — rare classes, low support
Sugar beet ↔ Potato

1.8 % — root crops, similar NDWI peak

06 — Training data

A continental multi-season mosaic.

The model is trained on 1.2 M labelled parcels spanning 4 cropped seasons (2020-2023) across 9 European countries. Labels come from LPIS / GSAA declarations harmonised to the LUCAS taxonomy.

1.2 M

Labelled parcels

Countries

Cropped seasons

Crop classes

Time-steps / season

Spectral features / step

07 — Honest limits

What this method cannot do — yet.

Small parcels

Performance drops below 0.5 ha — mixed pixels dominate the spectral signal.

Climate shock years

Drought or flood years weaken the cross-season transfer; per-AOI fine-tuning is then required.

Crop rotations

Parcel identity is assumed stable; mergers and splits need a re-segmentation step upstream.

Build your own crop-type atlas.

Open notebooks, pre-trained encoders and a 100 k-parcel demo dataset are available for replication and extension.

Get the code Read the paper