--- title: "Sentinel-1 crop monitoring in Python: SAR time series for wheat and oilseed rape" description: "Sentinel-1 crop monitoring in Python: VV, VH and VH/VV ratio time series per field, speckle averaging, wheat vs oilseed rape, and cloud gaps in NDVI." url: https://www.technicalgis.com/tutorials/sentinel-1-crop-monitoring/ author: TechnicalGIS Team publisher: TechnicalGIS (https://www.technicalgis.com/about/) published: 2026-10-17 updated: 2026-10-17 language: en-GB --- # Sentinel-1 crop monitoring in Python: SAR time series for wheat and oilseed rape > Sentinel-1 crop monitoring in Python: VV, VH and VH/VV ratio time series per field, speckle averaging, wheat vs oilseed rape, and cloud gaps in NDVI. - Level: Intermediate - Tested with: Python 3.12 - Source code: [Download scripts](https://www.technicalgis.com/code/sentinel-1-crop-monitoring/) ## Key takeaways 1. Sentinel-1 radar imaged all 247 test fields in South Cambridgeshire on all 33 dates of one orbit, every 12 days. Sentinel-2 got a clear view of most of the fields on only 17 of 155 dates, and on none in March and April 2024. 2. Average linear backscatter over each field before converting to decibels. Single pixels scatter by 1.6 dB, but a field mean of about 900 pixels is good to about 0.06 dB. 3. Crops have distinct radar signatures. Oilseed rape's VH rose to about −10 dB in May–June while wheat's fell to −18 dB, and wheat's VH/VV ratio climbed from −10.3 dB in January to −3.1 dB in mid-May as the crop grew. 4. A nearest-centroid classifier using only the radar time series matched the Crop Map of England label for 87% of 219 fields, and for all 21 oilseed rape fields. ## Before you start - Comfortable with NumPy, pandas and GeoPandas; Sentinel-1 SAR flood mapping in Python introduces backscatter, decibels and speckle - Python 3.10+ with pystac-client, planetary-computer, rasterio, numpy, scipy, pandas, geopandas, osmnx, requests and matplotlib - An internet connection: imagery is streamed from the Microsoft Planetary Computer and labels from the Defra Data Services Platform, with no account or API key **Short answer:** to monitor crops with Sentinel-1 in Python, search the Planetary Computer’s `sentinel-1-rtc` collection for one relative orbit across the season. Stream the VV and VH bands onto a common grid, then average the **linear** backscatter over each field to remove speckle, and only then convert to decibels. Plot VV, VH and the **VH/VV ratio** (VH dB − VV dB) per crop. A growing canopy raises VH and the ratio. Oilseed rape, winter wheat and spring crops each leave a distinct curve, and the radar records it every 12 days whatever the weather. England’s arable east is cloudy for much of the year. An optical satellite such as Sentinel-2 often fails to see a field for weeks, and the gap tends to fall in spring, just when crops grow fastest. Radar sees through cloud. In this tutorial, we follow 247 fields between Royston and Cambridge through the 2024 harvest year with free **Sentinel-1** data. We label them with the open **Crop Map of England** and measure how many usable Sentinel-2 views the same fields got. You will learn how to: 1. get field boundaries from **OpenStreetMap** and crop labels from **CROME 2024** through its OGC API; 2. stream a season of **Sentinel-1 RTC** from one orbit and average it per field; 3. measure **speckle**, and see why field-averaging beats filtering for this job; 4. read the **VV, VH and VH/VV ratio** growth signatures of winter wheat, oilseed rape, spring barley and maize; 5. count the **Sentinel-2 cloud gaps** over the same fields and compare NDVI with the radar ratio; 6. test how well the radar time series alone **separates the crops**. All outputs below are real results from [`s1_crop_monitoring.py`](https://www.technicalgis.com/code/sentinel-1-crop-monitoring/), run on 17 October 2026 with Python 3.12, pystac-client 0.9.0, rasterio 1.5.2 and osmnx 2.1.1. A full run takes about a minute on a home broadband connection. ## Why radar for agriculture? Sentinel-1 carries a C-band synthetic aperture radar with a 5.6 cm wavelength. It makes its own illumination, so it images at night, and it passes through cloud. What comes back from a field depends on the soil’s roughness and moisture and on the size, shape, water content and orientation of the plants. - **VV** (vertical send, vertical receive) responds strongly to the soil surface and to vertical stems. A dense, upright cereal canopy _attenuates_ VV, so VV falls as wheat grows tall. - **VH** (vertical send, horizontal receive) comes mostly from **volume scattering**: the pulse bounces around among leaves, branches and pods, which depolarises it. VH rises with biomass, especially for broad-leaved, branching crops such as oilseed rape. - The **VH/VV ratio** cancels some of the effect of soil moisture and incidence angle that both bands share. It tracks canopy development closely, which is why it is often compared with NDVI. If you are new to SAR, start with [Sentinel-1 SAR flood mapping in Python](https://www.technicalgis.com/tutorials/sentinel-1-sar-flood-mapping/), which covers decibels, speckle and orbits in more detail. ## Step 1: field boundaries from OpenStreetMap We need field polygons to average over. In much of rural England, OpenStreetMap volunteers have mapped farmland as `landuse=farmland`. Some polygons are single fields; others cover several. We fetch them with OSMnx, keep those of 3–100 ha, and shrink each by 20 m. That drops hedges, tracks and the mixed pixels at the field edge. `s1_crop_monitoring.py` ```python BBOX = (-0.12, 52.08, 0.02, 52.17) # lon/lat: arable South Cambridgeshire, between Royston and Cambridge CRS = "EPSG:32630" # UTM zone 30N: the native grid of these Sentinel-1 and Sentinel-2 scenes INSET_M = 20 # shrink each field by 20 m to drop hedges, tracks and mixed edge pixels osm = ox.features_from_bbox(BBOX, {"landuse": "farmland"}) fields = osm[osm.geometry.type.isin(["Polygon", "MultiPolygon"])].reset_index().to_crs(CRS) aoi_utm = gpd.GeoSeries([box(*BBOX)], crs=4326).to_crs(CRS).iloc[0] fields = fields[fields.within(aoi_utm)].copy() fields["area_ha"] = fields.area / 1e4 fields["core"] = fields.buffer(-INSET_M) fields = fields[(fields["area_ha"] >= 3) & (fields["area_ha"] <= 100) & ~fields["core"].is_empty].reset_index(drop=True) ``` The script saves the result to a GeoPackage and retries a second Overpass server if the first is busy, so you only download the polygons once. `Output — OpenStreetMap farmland polygons` ```text 413 polygons of 3-100 ha inside the study area, median 10.5 ha, total 5,636 ha ``` ## Step 2: crop labels from the Crop Map of England To know which signature belongs to which crop, we need labels. The Rural Payments Agency publishes the **Crop Map of England (CROME)** every year under the Open Government Licence. It is a grid of hexagons of about 0.4 ha, each with a crop code. The 2024 map is available through an **OGC API – Features** endpoint, which returns GeoJSON in pages of up to 10,000 features: `s1_crop_monitoring.py` ```python CROME = ("https://environment.data.gov.uk/spatialdata/crop-map-of-england-2024/ogc/features/v1/" "collections/Crop_Map_of_England_2024_Cambridgeshire/items") CROPS = {"AC66": "Winter wheat", "AC67": "Winter oilseed rape", "AC63": "Winter barley", "AC01": "Spring barley", "AC17": "Maize", "AC03": "Sugar beet", "LG20": "Winter field beans", "PG01": "Grass", "TG01": "Temporary grass", "FA01": "Fallow"} MIN_HEX, MIN_PURITY = 10, 0.8 # a field needs >= 10 CROME hexagons, and >= 80% of them must agree feats, url = [], CROME params = {"f": "json", "limit": 10000, "bbox": ",".join(map(str, BBOX))} while url: # OGC API Features: follow the "next" links page by page r = requests.get(url, params=params, timeout=300) r.raise_for_status() page = r.json() feats += page["features"] url, params = next((lk["href"] for lk in page.get("links", []) if lk["rel"] == "next"), None), None hexes = gpd.GeoDataFrame.from_features(feats, crs=4326).to_crs(CRS) hexes["geometry"] = hexes.centroid joined = gpd.sjoin(hexes, fields.set_geometry("core")[["core"]], predicate="within") stats = joined.groupby("index_right")["lucode"].agg(n_hex="size", lucode=lambda s: s.mode().iloc[0], purity=lambda s: s.value_counts(normalize=True).iloc[0]) fields = fields.join(stats) fields["crop"] = fields["lucode"].map(CROPS).fillna("Other") labelled = fields[(fields["n_hex"] >= MIN_HEX) & (fields["purity"] >= MIN_PURITY)].reset_index(drop=True) ``` We give each OSM polygon the most common CROME code among the hexagons inside it. We keep it only if at least 80% of the hexagons agree. That drops polygons that span several fields with different crops. `Output — CROME 2024 labels` ```text 24,734 CROME hexagons; 413 polygons contain hexagons 247 fields are >= 80% one crop (the rest are mixed, or the OSM polygon spans several fields) fields median_ha crop Winter wheat 129 12.2 Spring barley 35 13.0 Winter oilseed rape 21 14.7 Winter barley 18 15.5 Grass 13 9.8 Winter field beans 8 8.5 Maize 8 15.2 Sugar beet 6 13.1 Fallow 4 15.0 Other 4 7.0 Temporary grass 1 7.0 ``` ## Step 3: a season of Sentinel-1 RTC, averaged per field The Planetary Computer’s `sentinel-1-rtc` collection already has terrain-corrected backscatter (γ⁰) as Cloud Optimised GeoTIFFs. Over this area, four relative orbits pass: two ascending and two descending. Each sees the fields from a different angle, and backscatter depends on the angle, so we **keep one orbit**. Orbit 59 passes at about 18:00 UTC, heading north. The 2024 harvest year starts in September 2023, when winter crops are sown: `s1_crop_monitoring.py` ```python SEASON = "2023-09-01/2024-09-30" # the 2024 harvest year: winter crops are sown in autumn 2023 ORBIT = 59 # one relative orbit (ascending, ~18:00 UTC): identical geometry every 12 days catalog = pystac_client.Client.open(STAC_URL, modifier=planetary_computer.sign_inplace) search = catalog.search(collections=["sentinel-1-rtc"], bbox=BBOX, datetime=SEASON, query={"sat:relative_orbit": {"eq": ORBIT}}) s1_items = sorted(search.items(), key=lambda i: i.datetime) ``` `Output — Sentinel-1 STAC search` ```text 33 scenes from relative orbit 59 (ascending), 04 Sep 2023 to 22 Sep 2024 platforms: {'SENTINEL-1A': 33} example: S1A_IW_GRDH_1SDV_20230904T174210_20230904T174235_050181_060A2D_rtc | polarisations ['VV', 'VH'] ``` We rasterise the shrunken field cores onto a 10 m grid in UTM zone 30N. For each scene, we stream VV and VH onto that grid with a `WarpedVRT` and sum the pixels per field with `scipy.ndimage.sum`. The key detail is that we average **linear power**, never decibels: the mean of logarithms is not the logarithm of the mean. `s1_crop_monitoring.py` ```python labels = rasterio.features.rasterize(zip(labelled["core"], labelled["fid"]), out_shape=(HEIGHT, WIDTH), transform=TRANSFORM, fill=0, dtype="int32") FIDS = labelled["fid"].to_numpy() npix = ndimage.sum(np.ones_like(labels), labels, FIDS) def field_means(img: np.ndarray) -> np.ndarray: """Mean of the valid (non-NaN) pixels in each field core. Use it on linear power, never on dB.""" ok = np.isfinite(img) s = ndimage.sum(np.where(ok, img, 0), labels, FIDS) n = ndimage.sum(ok, labels, FIDS) with np.errstate(invalid="ignore", divide="ignore"): return np.where(n >= 0.8 * npix, s / n, np.nan) # need 80% of the field core covered def s1_scene(item) -> dict: out = {} for pol in ("vv", "vh"): g0 = read_warped(item.assets[pol].href) # WarpedVRT onto the 10 m grid, average resampling g0[g0 <= 0] = np.nan out[pol] = field_means(g0) return {"date": item.datetime, **out} with ThreadPoolExecutor(max_workers=8) as pool: scenes = list(pool.map(s1_scene, s1_items)) VV = np.array([s["vv"] for s in scenes]) # dates x fields, linear power VH = np.array([s["vh"] for s in scenes]) ``` The full script also caches each scene to a small `.npz` file and keeps three full images for the maps. `Output — analysis grid and reading` ```text 1000 x 1040 pixels of 10 m (10.0 x 10.4 km), EPSG:32630 field cores: median 941 pixels per field (min 343, max 5119) read and field-averaged 33 scenes x 2 polarisations in 12 s ``` Reading 66 rasters takes 12 seconds because rasterio fetches only the part of each Cloud Optimised GeoTIFF that covers our 10 km box. ![Two maps of a 10 by 10 km area of South Cambridgeshire. Left: a colourful Sentinel-1 VH composite where fields appear as patches of red, green, blue, magenta and white. Right: 247 field polygons coloured by crop over a grey radar background, mostly winter wheat in brown, with spring barley in blue, oilseed rape in green, winter barley in orange and smaller numbers of maize, beans and sugar beet.](https://www.technicalgis.com/_astro/01-vh-composite-crop-map.D3O_r2LX_ZElUlV.webp) _Left: VH backscatter on three dates as red (15 November 2023), green (26 March 2024) and blue (18 June 2024). Fields with different crops change differently through the year, so they take different colours. Right: the 247 single-crop fields, labelled from CROME 2024._ The colour composite already separates crops. Oilseed rape is bright in VH all spring, so it shows as white or pale green. Spring crops are bare in November and March, and their canopy grows by June, so they look blue. Wheat looks dark red to brown. ## Step 4: speckle, and why we average per field Each Sentinel-1 pixel is speckled: random interference between scatterers makes neighbouring pixels in the same field differ by several decibels. In the [flood tutorial](https://www.technicalgis.com/tutorials/sentinel-1-sar-flood-mapping/) we used a Lee filter, because we needed a map of every pixel. For crop monitoring we want one number per field, so we can simply average all the pixels in it. How good is that average? A split-half test tells us: put each pixel at random into half A or half B, average each half, and compare. `s1_crop_monitoring.py` ```python inside = labels > 0 pix_db = db(vh_img[inside]) lab_in = labels[inside] within_std = pd.Series(pix_db).groupby(lab_in).std() rng = np.random.default_rng(0) half = rng.random(inside.sum()) < 0.5 # split each field's pixels at random into two halves lin = vh_img[inside] mean_a = pd.Series(lin[half]).groupby(lab_in[half]).mean() mean_b = pd.Series(lin[~half]).groupby(lab_in[~half]).mean() half_diff = (db(mean_a) - db(mean_b)).dropna() ``` `Output — speckle` ```text VH on 2024-03-26: median within-field pixel std 1.61 dB split-half test: std of (half A - half B) = 0.114 dB, so a whole-field mean is good to about 0.057 dB ``` Each half has a speckle error of about 0.114 / √2 ≈ 0.08 dB. The whole field has twice as many pixels, so its error is about 0.057 dB, nearly 30 times smaller than the 1.6 dB scatter of single pixels. That is far smaller than the 5–8 dB changes that crops cause, so **no speckle filter is needed** once you average per field. ![Three panels. Left: a noisy, grainy VH image of a 3 by 3 km area. Middle: the same area with each field filled with its single mean value, showing smooth uniform colours. Right: histograms of single-pixel VH for wheat and oilseed rape, which overlap widely between about −22 and −10 dB, with tick marks below for the field means, which form two tight, separate clusters near −18 and −13 dB.](https://www.technicalgis.com/_astro/02-speckle-field-averaging.R5KMXWQI_Z1mFOkq.webp) _Speckle on 26 March 2024. Single pixels of wheat and oilseed rape overlap widely (right, outlines), but the field means (tick marks) form two tight, separate groups._ ## Step 5: crop growth signatures in VV, VH and VH/VV With a clean time series for each field, we take the median field of each crop on each date and plot the interquartile range around it. The ratio in decibels is just a subtraction: `s1_crop_monitoring.py` ```python vv_db, vh_db = db(VV), db(VH) ratio_db = vh_db - vv_db # VH/VV ratio in dB = VH(dB) - VV(dB) long = pd.DataFrame({ "date": np.repeat(S1_DATES, len(FIDS)), "fid": np.tile(FIDS, len(S1_DATES)), "vv_db": vv_db.ravel(), "vh_db": vh_db.ravel(), "ratio_db": ratio_db.ravel()}) long = long.merge(labelled[["fid", "id", "crop", "area_ha"]].rename(columns={"id": "osm_id"}), on="fid") long.round(3).to_csv(OUT / "field_backscatter_timeseries.csv", index=False) med = long.groupby(["crop", "date"])[["vv_db", "vh_db", "ratio_db"]].median() ``` ![Three stacked line charts from September 2023 to September 2024 for winter wheat, oilseed rape, spring barley and maize. VV: all crops sit near −6 to −8 dB in winter; from April, wheat drops to about −15.7 dB in mid-May and spring barley to −15 dB in June, while oilseed rape rises to about −5.5 dB in May and June. VH: oilseed rape climbs steadily from −18 dB in September to −10 dB in May and June, while the others stay between −15 and −19 dB. VH/VV ratio: wheat rises from about −10 dB in January to −3 dB in mid-May; spring barley rises later, in May and June; maize rises last, in July and August.](https://www.technicalgis.com/_astro/03-crop-signatures-vv-vh-ratio.j6tVHBZz_Z2jl8e0.webp) _Median field and interquartile range per crop for VV, VH and the VH/VV ratio, from 33 dates of orbit 59. The bands are narrow: fields of the same crop behave alike._ `Output — crop signatures (median of field means, dB)` ```text crop fields VH Nov-Dec VH Mar VH May-Jun VV May-Jun VH/VV May-Jun VH/VV Sep Winter wheat 129 -15.8 -16.7 -18.3 -14.7 -3.7 -8.0 Winter oilseed rape 21 -13.7 -13.7 -10.5 -6.4 -4.1 -8.8 Spring barley 35 -15.1 -17.3 -18.1 -12.8 -5.2 -8.3 Maize 8 -14.3 -16.5 -18.1 -10.2 -8.0 -6.8 Winter barley 18 -16.0 -16.5 -16.2 -11.8 -4.3 -8.2 Sugar beet 6 -13.9 -18.4 -16.5 -9.7 -6.6 -6.5 Winter field beans 8 -15.0 -17.3 -13.9 -8.4 -5.4 -8.3 Grass 13 -16.3 -16.4 -16.2 -11.1 -5.0 -5.7 largest all-field VV jumps: 10 Sep 2024 +3.8 dB, 22 Oct 2023 +2.4 dB, 13 May 2024 -2.3 dB Winter wheat: VH/VV highest -3.1 dB on 13 May 2024, lowest -10.3 dB on 26 Jan 2024; VV lowest -15.7 dB on 13 May 2024 Winter oilseed rape: VH/VV highest -3.6 dB on 13 May 2024, lowest -9.1 dB on 22 Sep 2024; VV lowest -10.2 dB on 04 Sep 2023 ``` How to read the curves: - **Winter oilseed rape** is the easiest crop to recognise. It is drilled in late August and builds a leafy rosette before winter, so its VH is already 2 dB above wheat in November and December (−13.7 against −15.8 dB). In spring it grows into a dense, branching, metre-high canopy of stems, flowers and then pods: a perfect volume scatterer. VH reaches about −10 dB in May and June, 8 dB above wheat. VV stays high too (−6.4 dB). After harvest in July both drop sharply. - **Winter wheat** looks like bare soil until March. From April, its stems extend and the vertical, water-filled canopy attenuates VV: VV falls from about −6 to **−15.7 dB in mid-May**, while VH changes little. The VH/VV ratio therefore climbs from −10.3 dB in late January to **−3.1 dB on 13 May**. As the crop ripens and dries in July, the ratio falls back, and stubble and cultivation after harvest bring VV back up. - **Spring barley** follows the same cereal pattern about a month later: it is sown in March, and its ratio peaks in June. - **Maize** is sown in May and grows through the summer, so its ratio rises last, in July and August. The VV jumps that move all crops together on the same date, such as **+3.8 dB on 10 September 2024**, are not crop growth. Rain wets the soil, which raises its dielectric constant and its backscatter. This is the main nuisance in SAR crop monitoring, and it is why the ratio, which partly cancels it, is popular. Look for these “every field jumps at once” dates and check them against rainfall records before you interpret a single-date change. ## Step 6: how often could Sentinel-2 see the same fields? Now the optical comparison. We search for **every** Sentinel-2 L2A acquisition over the area in the same season, whatever its cloud cover. For each, we read the scene classification layer (SCL) at 10 m and count the field pixels classed as vegetation or bare soil (classes 4 and 5). Cloud, cloud shadow, cirrus and unclassified pixels count as unusable. Where at least 20% of the field pixels are clear, we also read red and near-infrared and compute NDVI per field, using only clear pixels. `s1_crop_monitoring.py` ```python CLEAR_SCL = [4, 5] # SCL vegetation, not vegetated. Cloud, shadow, cirrus, unclassified etc. are unusable def s2_scene(item) -> dict: scl = read_warped(item.assets["SCL"].href, Resampling.nearest, nodata=0, dtype="uint8") clear = np.isin(scl, CLEAR_SCL) covered = float((scl[inside] > 0).mean()) n_clear = ndimage.sum(clear, labels, FIDS) out = {"covered": np.array(covered), "clear_frac": np.array(float(clear[inside].mean())), "ndvi": np.full(len(FIDS), np.nan)} if out["clear_frac"] >= 0.2: # only read the 10 m bands when there is something to see off = -1000 if float(item.properties["s2:processing_baseline"]) >= 4.0 else 0 red = read_warped(item.assets["B04"].href, nodata=0) + off nir = read_warped(item.assets["B08"].href, nodata=0) + off with np.errstate(invalid="ignore", divide="ignore"): ndvi = np.where(clear, (nir - red) / (nir + red), np.nan) s = ndimage.sum(np.nan_to_num(ndvi), labels, FIDS) with np.errstate(invalid="ignore", divide="ignore"): out["ndvi"] = np.where(n_clear >= 0.8 * npix, s / n_clear, np.nan) return {"date": item.datetime, **out} ``` The `−1000` offset matters: Sentinel-2 products from processing baseline 04.00 (January 2022) onwards add 1000 to every reflectance value, and NDVI is wrong if you forget to remove it. `Output — Sentinel-2 over the same fields` ```text 155 Sentinel-2 L2A acquisitions of tile 30UXC (any cloud cover), 03 Sep 2023 to 29 Sep 2024 read SCL (and red/NIR where useful) for 155 acquisitions in 8 s; 155 cover the fields fully acquisitions with >= 80% of field pixels clear: 17 of 155 (11%) acquisitions where at least one field is clear: 39 clear NDVI observations per field over 13 months: median 25, min 17, max 31 (Sentinel-1: 33 on every field) longest gap between clear views of the same field: median 43 days (range 30-90); clear views per field in March-April 2024: median 2 longest gap between >= 80%-clear Sentinel-2 dates: 90 days (12 Feb 2024 to 12 May 2024); longest Sentinel-1 gap: 12 days S1 scenes (orbit 59) S2 acquisitions S2 >= 80% clear Sep 2023 3 12 2 Oct 2023 2 12 2 Nov 2023 3 12 1 Dec 2023 2 12 1 Jan 2024 3 13 3 Feb 2024 2 11 1 Mar 2024 3 13 0 Apr 2024 2 12 0 May 2024 3 12 2 Jun 2024 3 12 2 Jul 2024 2 13 2 Aug 2024 3 12 1 Sep 2024 2 9 0 ``` Sentinel-2A and 2B together visited the area 155 times, about every 2–3 days, but only **17 acquisitions (11%)** saw at least 80% of the field area clear of cloud. There was **not one such day in March or April 2024**, a famously wet spring. The typical field went 43 days without a single clear view, and some went 90 days. Over the same period, Sentinel-1 delivered a clean measurement of every field every 12 days. ![Top: a timeline with 33 evenly spaced blue ticks for Sentinel-1 and 155 ticks for Sentinel-2, most of them grey, some orange and few green. Bottom: bars of the percentage of field pixels clear for each Sentinel-2 date; only 17 bars exceed the 80% line, with a gap from mid-February to mid-May 2024 with no bar above 80%.](https://www.technicalgis.com/_astro/04-s1-vs-s2-availability.BHjtHfE6_Hfynx.webp) _Every Sentinel-1 date from one orbit (blue) against every Sentinel-2 date, coloured by how much of the field area was clear. Green bars are the only near-cloud-free views; none fell in March or April 2024._ ## Step 7: NDVI and the VH/VV ratio side by side On the dates when Sentinel-2 _can_ see, how does NDVI compare with the radar? We pair each crop’s median NDVI with the nearest radar date, within three days, and compute a correlation: `s1_crop_monitoring.py` ```python for crop in ["Winter wheat", "Winter oilseed rape"]: a, b = ndvi_med.loc[crop], med.loc[crop, "ratio_db"] pairs = pd.merge_asof(a.rename("ndvi").reset_index().sort_values("date"), b.rename("ratio").reset_index().sort_values("date"), on="date", direction="nearest", tolerance=pd.Timedelta("3D")).dropna() r = np.corrcoef(pairs["ndvi"], pairs["ratio"])[0, 1] ``` `Output — NDVI versus VH/VV ratio` ```text Winter wheat: 39 NDVI dates with >= 5 clear fields; 16 within 3 days of a radar date; Pearson r = 0.82 NDVI outliers (> 0.25 from the 5-date running median): 07 May 2024 (0.15) Winter oilseed rape: 34 NDVI dates with >= 5 clear fields; 13 within 3 days of a radar date; Pearson r = 0.48 NDVI outliers (> 0.25 from the 5-date running median): 25 Oct 2023 (0.09), 07 May 2024 (0.21) ``` ![Two charts. Top, winter wheat: the radar VH/VV ratio line rises from about −10 dB in January to −3 dB in May and falls after July; NDVI squares rise from 0.2 in autumn to about 0.85 in May and June, then fall to 0.2 after harvest, closely following the ratio, with gaps in February to April. Bottom, oilseed rape: NDVI is already 0.7–0.8 in autumn and winter while the ratio is still low; both peak in May.](https://www.technicalgis.com/_astro/05-ndvi-vs-vhvv-ratio.BPShdCno_tGm7L.webp) _Median VH/VV ratio (line, every 12 days) and median NDVI (squares, only when clear) for winter wheat and oilseed rape fields. NDVI misses most of the spring; the radar does not._ For **wheat**, the ratio follows NDVI closely (r = 0.82): both track the green canopy from tillering to ripening. For **oilseed rape**, the agreement is weaker (r = 0.48). Rape is green and leafy from October, so NDVI is high all winter, but the radar ratio rises only when the plants build stems and pods in spring. The two sensors measure different things: NDVI sees chlorophyll, radar sees structure and water. They complement each other rather than one replacing the other. The outliers are also worth noticing. On 7 May 2024, NDVI over wheat fell to 0.15 between neighbours of about 0.85. Haze or thin cloud that the SCL missed is the likely cause. Cloud masks are not perfect, and radar has no such failure mode. ## Step 8: can radar alone tell the crops apart? As a final test, we try to classify each field from its radar curves alone. We stack the 33 VV and 33 VH values of each field into a 66-number vector and standardise it. Then we assign each field to the crop with the nearest average curve (centroid), computing the centroids **without** the field being tested (leave-one-out). Crops with fewer than 8 fields are left out. `s1_crop_monitoring.py` ```python X = np.hstack([vv_db.T, vh_db.T])[sel] # fields x (dates x 2) feature matrix X = (X - X.mean(axis=0)) / X.std(axis=0) y = labelled.loc[sel, "crop"].to_numpy() pred = [] for i in range(len(y)): # leave one field out, classify it by the nearest class centroid keep = np.arange(len(y)) != i cents = {c: X[keep][y[keep] == c].mean(axis=0) for c in CLASSES} pred.append(min(cents, key=lambda c: np.linalg.norm(X[i] - cents[c]))) ``` `Output — leave-one-out classification` ```text 219 fields, 66 features (33 dates x VV, VH); overall accuracy 87% Winter wheat recall 86% (129 fields) Winter oilseed rape recall 100% (21 fields) Winter barley recall 78% (18 fields) Spring barley recall 89% (35 fields) Maize recall 75% (8 fields) Winter field beans recall 100% (8 fields) ``` ![Confusion matrix of six crops. The diagonal is dark: 111 winter wheat, 21 oilseed rape, 14 winter barley, 31 spring barley, 6 maize and 8 beans correct. The largest off-diagonal cell is 12 winter wheat fields predicted as spring barley.](https://www.technicalgis.com/_astro/06-confusion-matrix.CmVjGEjp_1DywRs.webp) _Leave-one-out nearest-centroid classification from the Sentinel-1 time series alone, against CROME 2024 labels._ Even this very simple method, with no machine-learning library, matches CROME on 87% of fields and on every oilseed rape and bean field. The main confusion is 12 wheat fields predicted as spring barley: both are cereals with a similar canopy, and late-drilled wheat behaves like a spring crop. A random forest or a temporal neural network on several orbits would do better. That is roughly how CROME itself is made. ## Limitations and how to improve them - **One orbit, one angle.** We used only orbit 59. Adding the other ascending orbit (132) and the two descending ones (8 and 81) roughly triples the number of dates, but each orbit needs its own reference curves, or a normalisation for incidence angle. - **Soil moisture.** Rain moves VV and VH on all fields at once. The ratio helps, but in winter, when there is little canopy, it still follows the weather. Compare with a rainfall series, or use a change-detection approach against neighbouring fields. - **Field boundaries.** OSM `landuse=farmland` polygons often cover several fields. We dropped the mixed ones, which is why only 247 of 413 survived. For production work, use the Rural Payments Agency’s **Land Parcels**, or segment fields from imagery. The [Segment Anything tutorial](https://www.technicalgis.com/tutorials/segment-anything-satellite-imagery/) shows one way. - **Labels from a model.** CROME is a classification, not a survey. Some “errors” in Step 8 may be CROME’s. - **Crop calendars vary.** Signatures shift with sowing date, variety and weather. Curves from one year and region transfer imperfectly to another. ## Frequently asked questions ### Can Sentinel-1 be used for crop monitoring? Yes. Sentinel-1 measures how crops scatter microwaves, which depends on their structure, biomass and water content. With a revisit of 6–12 days that is never interrupted by cloud, it gives a reliable time series for every field, which is ideal for crop type mapping, growth stage monitoring and detecting harvest, ploughing or mowing. It is used operationally in national crop maps, including the Crop Map of England. ### What does the VH/VV ratio show for crops? The VH/VV ratio, VH (dB) minus VV (dB), rises as the canopy grows, because volume scattering from leaves and stems raises VH and a vertical cereal canopy weakens VV. In this tutorial, the median winter wheat field went from −10.3 dB in January to −3.1 dB in mid-May. The ratio is less affected by soil moisture than either band on its own and correlates with NDVI for cereals (r = 0.82 for wheat here). ### Why is VH higher for oilseed rape than for wheat? Oilseed rape grows into a tall, branching canopy of stems, leaves, flowers and pods with many randomly oriented elements. That depolarises the radar pulse and gives strong cross-polarised (VH) backscatter. Wheat stems are mostly vertical and regularly spaced, which attenuates VV rather than adding VH. In May and June 2024, rape fields were about 8 dB brighter in VH than wheat fields (−10.5 against −18.3 dB). ### Do I need to filter speckle for crop monitoring? Not if you average over fields. Average the linear backscatter of all the pixels in a field, then convert the mean to decibels. With about 900 pixels per field, the error from speckle is about 0.06 dB, far below crop effects of several decibels. Use a speckle filter, such as Lee or refined Lee, when you need pixel-level maps instead. ### Should I use ascending or descending orbits? Either, but do not mix them in one time series without care. Ascending and descending passes, and different relative orbits, view the field from different directions and incidence angles, which shifts the backscatter by a decibel or more. Morning (descending) passes over Europe can also see dew on leaves. Start with one orbit, and add others as separate series. ### Is Sentinel-1 better than Sentinel-2 for agriculture? Neither is better; they measure different things. Sentinel-2 NDVI measures greenness directly and is easier to interpret, but in this example it got a clear view of most fields on only 17 of 155 dates. Sentinel-1 never misses a date but is affected by soil moisture and is harder to read. Most operational crop maps use both. ## Next steps - Learn the radar basics, including the Lee filter and decibels, in [Sentinel-1 SAR flood mapping in Python](https://www.technicalgis.com/tutorials/sentinel-1-sar-flood-mapping/). - Delineate field boundaries yourself with [Segment Anything on satellite and aerial imagery](https://www.technicalgis.com/tutorials/segment-anything-satellite-imagery/). - Learn STAC and Sentinel-2 cloud masking in [Measuring Lake Mead’s shrinking shoreline with Sentinel-2 and STAC](https://www.technicalgis.com/tutorials/sentinel-2-stac-lake-mead/) and [Monitoring reservoir water quality from space with Sentinel-2](https://www.technicalgis.com/tutorials/reservoir-water-quality-sentinel-2/). - Add the other three orbits, detect the harvest date of each field from the sudden VH/VV drop, or replace the nearest-centroid classifier with a random forest. - Compare the radar signatures with an optical view of crop growth in [NDVI time series in Python with Sentinel-2 and odc-stac](https://www.technicalgis.com/tutorials/ndvi-time-series-odc-stac/). ## Data and licences - **Sentinel-1 RTC**: contains modified Copernicus Sentinel data 2023–2024, processed by ESA; radiometric terrain correction by Catalyst/Microsoft. Licensed CC BY 4.0. Accessed from the [Microsoft Planetary Computer](https://planetarycomputer.microsoft.com/dataset/sentinel-1-rtc) on 17 October 2026. - **Sentinel-2 L2A**: contains modified Copernicus Sentinel data 2023–2024. Accessed from the [Microsoft Planetary Computer](https://planetarycomputer.microsoft.com/dataset/sentinel-2-l2a). - **Crop Map of England (CROME) 2024**: © Rural Payments Agency, published by Defra under the [Open Government Licence v3.0](https://www.nationalarchives.gov.uk/doc/open-government-licence/version/3/). Accessed through the [Defra Data Services Platform](https://environment.data.gov.uk/) OGC API. - **Field boundaries**: © OpenStreetMap contributors, available under the [Open Database Licence](https://www.openstreetmap.org/copyright). - Code: [`s1_crop_monitoring.py`](https://www.technicalgis.com/code/sentinel-1-crop-monitoring/), tested with Python 3.12.0, pystac-client 0.9.0, planetary-computer 1.0.0, rasterio 1.5.2, geopandas 1.2.0, osmnx 2.1.1 and pandas 3.0.6. ## Key terms - **Backscatter (γ⁰)**: The share of the radar pulse returned towards the satellite, normalised for the illuminated area and usually shown in decibels. - **VV and VH**: Polarisations: VV is sent and received vertically; VH is sent vertically and received horizontally. VH comes mostly from volume scattering in leaves and stems. - **VH/VV ratio**: Cross-polarised over co-polarised backscatter. In decibels it is VH(dB) − VV(dB). It rises as a crop canopy grows and is less sensitive to soil moisture than either band alone. - **Speckle**: The grainy texture of every coherent radar image, caused by interference between many scatterers in each pixel. Averaging many pixels or dates reduces it. - **CROME**: The Crop Map of England: an annual machine-learning classification of crop type for hexagons of about 0.4 ha, published openly by the Rural Payments Agency. - **NDVI**: Normalised difference vegetation index, (NIR − red) / (NIR + red), from optical imagery such as Sentinel-2. It needs a cloud-free view. ## Data for this tutorial - [Microsoft Planetary Computer](https://www.technicalgis.com/data-sources/#src-microsoft-planetary-computer) · Sentinel-1 RTC and Sentinel-2 L2A as Cloud Optimised GeoTIFFs via STAC - [OpenStreetMap](https://www.technicalgis.com/data-sources/#src-openstreetmap) · landuse=farmland polygons used as field boundaries - [Copernicus Data Space Ecosystem](https://www.technicalgis.com/data-sources/#src-copernicus-data-space-ecosystem) · Sentinel-1 GRD and Sentinel-2 from ESA Licences and formats for each source are listed in the [GIS data sources catalogue](https://www.technicalgis.com/data-sources/). --- Source: https://www.technicalgis.com/tutorials/sentinel-1-crop-monitoring/