Digital Twins for Agrifood: from Satellites to Smart Irrigation
Full text
10/6/2025 1 Digital Twins for Agrifood: from Satellites to Smart Irrigation 2-hour tutorial Prof. Adriano Camps* With contributions from Dr. Hog Wang, Prof. Mercé Vall·llossera, Prof. Carlos López, Zhongmin Ma, Gerard Portal, Miriam Pablos *CommSensLab-UPC, Dept. of Signal Theory and Communications, Universitat Politècnica de Catalunya Institut d’Estudis Espacials de Catalunya ASPIRE Visiting International Professor, UAE University E-mail: adr[email protected] 8/10/2025/ © A. Camps, UPC 2025 1/90 Motivation: Why Digital Twins for Agriculture? 8/10/2025/ © A. Camps, UPC 2025 2/90
10/6/2025 2 What is a Digital Twin? 8/10/2025/ © A. Camps, UPC 2025 3/90 Digital Twins in Agriculture Plantation with green crops growing in agricultural farm [pexels.com] 8/10/2025/ © A. Camps, UPC 2025 4/90
10/6/2025 3 Case Study: AI4WATER Project Digital Twin (i) 8/10/2025/ © A. Camps, UPC 2025 5/90 Case Study: AI4WATER Project Digital Twin (ii) 8/10/2025/ © A. Camps, UPC 2025 6/90
10/6/2025 4 Tutorial Outline 8/10/2025/ © A. Camps, UPC 2025 7/90 Plants need water to develop, and only a portion of the total biomass is converted into crops You can't produce "more with less" Traditional Management of Irrigation Water (i) 8/10/2025/ © A. Camps, UPC 2025 8/90 Adapted from“Experiencia en la Mejora de la Eficiencia en el Uso del Agua en Agricultura de Regadío: Tecnología, Información e Implicación de los usuarios”, Vicente Bodas, Albacete, 29/2-1/3, 2024
10/6/2025 5 •FAO56 model supported by Remote Sensing ET = Transpiration + Evaporation ET = Ks · Kcb · ETo +Ke ·ETo Satellite: Temporal Evolution Agroclimatology SiAR network[1] / RuralCat[2] Water balance in the root zone Water balance in the surface (evaporation) Adapted from“Evapotranspiración y balance de agua en suelo en cultivos leñosos”, Juan Manuel Sánchez & Jaime Campoy, Albacete, 29/2-1/3, 2024 Readily Available Water (RAW) Total Available Water (TAW) ETo Tarazona de La Mancha (Albacete) https://www.fao.org/3/x0490E/x0490e0c.htm#t ranspiration%20component%20(kcb%20eto) Kcb [1] https://servicio.mapa.gob.es/websiar/NecesidadesHidricas.aspx [2] https://ruralcat.gencat.cat/web/guest/agrometeo.estacions Ks Traditional Management of Irrigation Water (ii) 8/10/2025/ © A. Camps, UPC 2025 9/90 𝐶𝑊𝑆𝐼 = 1 − 𝐸𝑇 𝐸𝑇 0 < CWSI < 1 Hydric stress coefficient real ET potential ET CWSI=0 CWSI=1 Difference between actual ET and potential ET is a measure of the degree of water stress of the plant. Traditional Management of Irrigation Water (iii) AquaCrop - The FAO Crop Water Productivity Model [https://www.fao.org/aquacrop] 8/10/2025/ © A. Camps, UPC 2025 10/90 Adapted from“Evapotranspiración y balance de agua en suelo en cultivos leñosos”, Juan Manuel Sánchez & Jaime Campoy, Albacete, 29/2-1/3, 2024
10/6/2025 6 The Role of Remote Sensing (i) 8/10/2025/ © A. Camps, UPC 2025 11/90 The Role of Remote Sensing (ii) NDVI crop mapping. Credits: poco_bw [istockphoto.com] 8/10/2025/ © A. Camps, UPC 2025 12/90
10/6/2025 7 Active: Passive: Microwaves Optical RADAR LIDAR Microwave Radiometers Optical Radiometers [http://web.physik.uni-rostock.de/ cluster/students/fp3/lidar_en.html] [http://www.srh.noaa.gov/srh/sod/radar/ radinfo/radinfo.html]] [http://lcogt.net/spacebook/black-body-radiation] [http://dusty.physics.uiowa.edu/~goree/ teaching/thermal.html] [http://www.smosbec.icm.csic.es/south_america_seen_by_smos] GNSS-R GNSS-RO Using Signals of Opportunity Types of Remote Sensors 8/10/2025/ © A. Camps, UPC 2025 13/90 Key satellites for Agricultural Monitoring 8/10/2025/ © A. Camps, UPC 2025 14/90
10/6/2025 8 Synthetic Aperture Radar - SAR (i): Sentinel-1 Acquisition number: |S hh +S vv |,|S hv +S vh |,|S hh -S vv | dB=δ p 3 - Entropy (H) 1 2 3 4 5 6 7 8 Acquisition number: dB=δ p 3 - Alpha Anisotropy (A) 1 2 3 4 5 6 7 8 •SAR Polarimetry (PolSAR) •Multi-temporal data: object & change detection VV VH HH HV [https://hyp3-docs.asf.alaska.edu/hyp3-docs/guides/introduction_to_sar/] Rough surface Volume Double bounce V-pol H-pol 8/10/2025/ © A. Camps, UPC 2025 15/90 SAR Polarimetry (PolSAR) & Multi-temporal data for crop monitoring Entropy FQ9 ASC FQ9 ASC 123456 1 2 345 61 2 345 6 Synthetic Aperture Radar - SAR (ii): Sentinel-1 8/10/2025/ © A. Camps, UPC 2025 16/90
10/6/2025 9 [https://land.copernicus.eu/global/products/] Sentinel-1 ASCAT Daily 1 km Soil Moisture Index (SSM) •It is not Soil Moisture, it is an “index” from [0, 1] •Measures reflectivity changes •SAR is affected by speckle noise average from 10 m to 1000 m 6 combined passes: 14-19/6/2023 Single pass: 19/6/2023 Synthetic Aperture Radar - SAR (iii): Sentinel-1 8/10/2025/ © A. Camps, UPC 2025 17/90 NDVI (Normalized Difference Vegetation Index): •It does not measure the soil moisture •It does measure the plant “health” •It is sensitive to the outer layer •Many applications for laptop and cell phone 𝑁𝐷𝑉𝐼 = 𝜌𝑁𝐼𝑅 − 𝜌𝑅𝑒𝑑 𝜌𝑁𝐼𝑅 + 𝜌𝑅𝑒𝑑 [https://physicsopenlab.org/wp-content/uploads/2017/01/veg.gif] [https://bikeshbade.com.np/media/uploads/2020/05/07/ ndvichart_eD4HXBw.png] [credits ESA] Sentinel-3 OLCI (300 m) Sentinel-2 MSI (20 m) 18 Optical Sensors: Sentinel-2 and Sentinel-3 Sentinel-2 8/10/2025/ © A. Camps, UPC 2025 18/90
10/6/2025 16 Remote Sensing: Data Quality Issues 8/10/2025/ © A. Camps, UPC 2025 31/90 Evolution of RFI @ L-band over time Fusion with Ground Data 8/10/2025/ © A. Camps, UPC 2025 32/90
10/6/2025 17 Case Study: Soil Moisture Maps vs. Probe readings 8/10/2025/ © A. Camps, UPC 2025 33/90 Open Tools for Remote Sensing Data & Digital Twins [https://browser.dataspace.copernicus.eu/] [https://earthengine.google.com/] [https://bec.icm.csic.es/] 8/10/2025/ © A. Camps, UPC 2025 34/90
10/6/2025 18 Summary: Remote Sensing in Digital Twins for Agriculture [Picture from Unsplash] 8/10/2025/ © A. Camps, UPC 2025 35/90 The Role of Meteorological Data (i) 8/10/2025/ © A. Camps, UPC 2025 36/90
10/6/2025 19 The Role of Meteorological Data (ii) [Picture from Unsplash] 8/10/2025/ © A. Camps, UPC 2025 37/90 Meteorological Data Sources ECMWF, NOAA, national and regional meteorological agencies May require specific APIs for each 8/10/2025/ © A. Camps, UPC 2025 38/90
10/6/2025 20 Key Meteorological Variables 8/10/2025/ © A. Camps, UPC 2025 39/90 Example API Call for Automated Weather Data Retrieval: 1. Client sends a POST request with JSON data to create a new user 2. API Gateway validates the request with the authentication service 3. Authentication Service confirms the token is valid 4. Business Logic Service processes the validated request 5. Database stores the data and returns confirmation 6. The response flows back through the same path with status information 8/10/2025/ © A. Camps, UPC 2025 40/90
10/6/2025 21 Integration with EO 8/10/2025/ © A. Camps, UPC 2025 41/90 Uncertainty Issues [Picture from Unsplash] 8/10/2025/ © A. Camps, UPC 2025 42/90
10/6/2025 22 Summary: Meteorological Data in Digital Twins for Agriculture [Picture from Unsplash] 8/10/2025/ © A. Camps, UPC 2025 43/90 The Role of Ground Observations (i) 8/10/2025/ © A. Camps, UPC 2025 44/90
10/6/2025 23 The Role of Ground Observations (ii) 45 Bell-lloc d’Urgell La Fuliola AI4WATER in situ sensors: •Blue: meteorological stations •Red: soil moisture/temperature/electric conductivity probes •Yellow: probes not operational any more Secà al 2023 Irrigat al 2023 8/10/2025/ © A. Camps, UPC 2025 45/90 Types of Ground Data 8/10/2025/ © A. Camps, UPC 2025 46/90
10/6/2025 24 Soil Probes (Sentek/Irrimax) 20 x IRRIMAX probes: SM, T, EC every 10 cm from 5 to 115 cm 8/10/2025/ © A. Camps, UPC 2025 47/90 ATMOS-41 Station 2 x ZENTRA ATMOS 41 meteorological stations: rays, ETo, atmospheric pressure, vapor pressure, precipitation, solar radiation, air temperature… 8/10/2025/ © A. Camps, UPC 2025 48/90
10/6/2025 25 Example measurements 8/10/2025/ © A. Camps, UPC 2025 49/90 Example of moisture profiles (rainfed field 2023, La Fuliola) 8/10/2025/ © A. Camps, UPC 2025 50/90
10/6/2025 32 Summary: Data Pipeline [Picture from Unsplash] 8/10/2025/ © A. Camps, UPC 2025 63/90 Virtual Layer - the “core” of the Digital Twin (i) 8/10/2025/ © A. Camps, UPC 2025 64/90
10/6/2025 33 Virtual Layer - the “core” of the Digital Twin (ii): Why Machine Learning ? •Capture Non-Linear Interactions in Agro-Environmental Systems •No need to “implement” FAO56 models... ML “learns” them from the data 8/10/2025/ © A. Camps, UPC 2025 65/90 Time Series Forecasting Methods 8/10/2025/ © A. Camps, UPC 2025 66/90
10/6/2025 34 Why LSTM ? 8/10/2025/ © A. Camps, UPC 2025 67/90 LSTM Architecture [1] Ct-1 Ct ht-1 ht itot ft [1] Hochreiter, S. and Schmidhuber, J. , Long Short-Term Memory. Neural Computation, 9(8), 1735-1780,1997. 8/10/2025/ © A. Camps, UPC 2025 68/90
10/6/2025 35 Training Setup 8/10/2025/ © A. Camps, UPC 2025 69/90 Training Curves Preventing Overfitting 8/10/2025/ © A. Camps, UPC 2025 70/90
10/6/2025 36 Validation Metrics 8/10/2025/ © A. Camps, UPC 2025 71/90 Layers Input variables Learning rate Batch size epochs lookback Forecast length Number of layers dropout hidden size Train/test split 0_10cm_60m '0_10cm soil moisture’, 'temperature_2m’, 'relative_humidity_2m’, 'precipitation’, ‘cloud_cover’, ‘et0_fao_evapotranspiration’, 'wind_speed_10m’, ‘wind_gusts_10m’, 'vapour_pressure_deficit' 0.0001 128 100 (early stop at 58) 7 14 1 0.1 64 81%/19% 10_40cm_60m '0_10cm soil moisture’, 'soil_moisture_10_to_40cm’, 'temperature_2m’, 'relative_humidity_2m’, 'precipitation’, 0.0001 128 100 (early stop at 20) 1 14 1 0.1 64 81%/19% 40_100cm_60m '0_10cm soil moisture’, 'soil_moisture_10_to_40cm’, 'soil_moisture_40_to_100cm’, 'relative_humidity_2m’, 'precipitation’ 0.0001 128 100 (early stop at 28) 7 14 1 0.1 64 81%/19% Training and validation @ 60 m spatial resolution (i) Parameter settings 8/10/2025/ © A. Camps, UPC 2025 72/90
10/6/2025 37 Training performance 0-10 cm @ 60 m 10-40 cm @60 m 40-100 cm @ 60 m Layers MSE RMSE MAE 0 -10 cm @ 60 m 0.0009 0.0296 0.0233 10 - 40 cm @ 60 m 0.0012 0.0356 0.0296 40 – 100 cm @ 60m 0.0018 0.0430 0.0357 Evaluation metrics Training and validation @ 60 m spatial resolution (ii) 8/10/2025/ © A. Camps, UPC 2025 73/90 Multi-layer Predictions 8/10/2025/ © A. Camps, UPC 2025 74/90
10/6/2025 38 Dataset Resolution 8/10/2025/ © A. Camps, UPC 2025 75/90 Interpretability importance Sample SHAP values of NN to downscale SMOS soil moisture maps (0-10 cm) push prediction higher push prediction lower no impact -- importance ++ Each dot represents one sample or observation in the dataset 8/10/2025/ © A. Camps, UPC 2025 76/90
10/6/2025 39 Example Predictions 8/10/2025/ © A. Camps, UPC 2025 77/90 Soil moisture profiles forecast up to 14 days from present Scalability 8/10/2025/ © A. Camps, UPC 2025 78/90
10/6/2025 40 Training and validation @ 1 km spatial resolution (i) Parameter settings Layers Input variables Learning rate Batch size epochs lookback Forecast length Number of layers dropout hidden size Train/test split 0_10cm_1km '0_10cm soil moisture’, 'temperature_2m’, 'relative_humidity_2m’, 'precipitation’, 'evapotranspiration’, 'vapour_pressure_deficit' 0.0001 64 10 1 14 3 0.3 128 70%/30% 10_40cm_1km '0_10cm soil moisture’, 'soil_moisture_10_to_40cm’, 'temperature_2m’, 'relative_humidity_2m’, 'precipitation’, 'evapotranspiration’, 'vapour_pressure_deficit' 0.0001 64 10 1 14 3 0.3 128 70%/30% 40_100cm_1km '0_10cm soil moisture’, 'soil_moisture_10_to_40cm’, 'soil_moisture_40_to_100cm’, 'relative_humidity_2m’, 'precipitation’, 'evapotranspiration’, 'vapour_pressure_deficit' 0.0001 64 10 1 14 3 0.3 128 70%/30% 8/10/2025/ © A. Camps, UPC 2025 79/90 Layers MSE RMSE MAE 0-10 cm @ 1 km 0.0016 0.0402 0.0320 10-40 cm @ 1 km 0.0002 0.0143 0.1003 40-100 cm @ 1 km 0.0003 0.0172 0.0128 Evaluation metrics Training performance 0-10 cm @ 1 km 10-40 cm @ 1 km 40-100 cm @ 1 km Training and validation @ 60 m spatial resolution (ii) 8/10/2025/ © A. Camps, UPC 2025 80/90
10/6/2025 41 Forecasting Soil Moisture - 14-day SM Predictions for Multi-layer profiles Actual Rain Rate from Meteo Data 8/10/2025/ © A. Camps, UPC 2025 81/90 Soil moisture profiles forecast up to 14 days from present Summary: Virtual Layer - the “core” of the Digital Twin 8/10/2025/ © A. Camps, UPC 2025 82/90