NSF Standard Research Proposal — NSF 24-514

AgroMind: Federated Deep Learning for Precision Irrigation Scheduling in Semi-Arid Mediterranean Climates Under Climate Change Stress

Harnessing the Data Revolution (HDR)  ·  CISE/IIS
Dra. Carmen Vidal-Roca, PhD
Univ. Politécnica de Valencia (UPV)
Dr. A. Fuentes · Dra. P. Sharma (Purdue)
September 15, 2026
36 months (Jan 2027 – Dec 2029)
$498,750 total direct + F&A
91
Projected Score / 100
Intellectual Merit93
Broader Impacts90
Innovation89
Feasibility87
Fundability
HIGH
Top 10% predicted
1
Project Summary
NSF Required · 1 page

Overview

Water scarcity represents one of the most critical threats to global food security. Citrus and almond cultivation in semi-arid Mediterranean climates accounts for over 4.2 million hectares worldwide, yet conventional irrigation management wastes an estimated 30–40% of applied water through temporal and spatial mismatch. This project introduces AgroMind, a federated deep learning framework that integrates multimodal data streams — Sentinel-2 satellite imagery, distributed IoT soil-moisture sensors, and regional climate forecasts — to generate real-time, farm-scale irrigation recommendations with unprecedented accuracy and generalizability.

Intellectual Merit

AgroMind advances knowledge at the intersection of federated machine learning, precision agriculture, and climate adaptation science. The central hypothesis — that privacy-preserving federated learning across heterogeneous farm networks yields superior, generalizable irrigation models compared to siloed local approaches — has not been rigorously tested at scale. This work will (1) pioneer a novel multimodal fusion architecture that synchronizes asynchronous satellite and in-situ time series; (2) establish formal theoretical bounds on federated convergence under non-IID agricultural data distributions; and (3) produce a publicly available, richly labeled dataset of 10,000+ farm-years across three climate zones, enabling a decade of downstream research.

Broader Impacts

The societal implications of AgroMind extend from individual farm economics to regional water policy. A demonstrated 25% reduction in irrigation water use across 50 pilot farms would conserve an estimated 2.1 million cubic meters annually — equivalent to the annual water supply of 14,000 households. The project will train four doctoral students (two from underrepresented groups) through an explicitly designed co-mentorship structure. Educational materials will be integrated into UPV's Agronomy Engineering curriculum, reaching 400+ students annually. All models, code, and datasets will be released under open licenses, ensuring global reproducibility and adoption.


2
Research Objectives & Specific Aims
Adapted NSF-style Aims
Long-term goal: Develop the scientific foundations and validated tools needed to make AI-driven precision irrigation a standard practice in semi-arid agriculture globally, reducing agricultural water waste by ≥25% without yield penalty by 2035.

The immediate objective of this proposal is to design, validate, and openly disseminate AgroMind — a federated deep learning system for real-time irrigation scheduling — across 50 citrus and almond farms in the Valencia region of Spain. We will test the central hypothesis that multimodal federated learning across privacy-preserving farm nodes will achieve irrigation recommendations with ≥85% accuracy and yield ≥25% water savings versus conventional scheduling, even under climate-change-accelerated precipitation variability.

AIM 1
Multimodal Federated Architecture
Design and implement an asynchronous federated learning architecture that fuses Sentinel-2 NDVI/NDWI time series (10m resolution, 5-day revisit), 48-node IoT soil-moisture arrays (0–30cm, 30–60cm depths), and ECMWF 7-day ensemble forecasts. Develop novel non-IID aggregation weights accounting for soil taxonomy heterogeneity across farm nodes.
Deliverable: Open-source FL framework (PyTorch + Flower) with documented API
AIM 2
Field Validation (2 Growing Seasons)
Validate AgroMind recommendations in randomized controlled trials across 50 farms over the 2027 and 2028 growing seasons. Primary metric: irrigation volume reduction vs. farmer-managed controls. Secondary: yield (kg/ha), water-use efficiency (kg/m³), and energy consumption. IRB equivalent (AEMPS) approval obtained prior to Year 1 fieldwork.
Deliverable: Peer-reviewed validation paper; pre-registered trial on OSF.io
AIM 3
Open Agricultural Dataset
Curate and publish a FAIR-compliant labeled dataset comprising 10,000+ farm-observation instances: matched satellite tiles (10m), soil sensor readings (15-min intervals), irrigation logs, weather station records, and annotated yield data. Dataset will cover 3 climate zones (Valencia, Murcia, and Purdue's partner sites in California Central Valley).
Deliverable: Dataset on Zenodo/HuggingFace; data paper in Scientific Data
AIM 4
Climate-Stress Generalization
Evaluate AgroMind's performance under projected 2030–2040 climate scenarios (CMIP6 SSP2-4.5 and SSP5-8.5) using dynamically downscaled regional climate model outputs. Develop adaptive fine-tuning protocols enabling rapid model recalibration (<3 weeks) for novel farm contexts with minimal labeled data (few-shot learning).
Deliverable: Climate-resilience stress-test report; adaptation protocol guide

3
Significance & Innovation

Significance. Agriculture accounts for 70% of global freshwater withdrawals (FAO, 2023). In Spain alone, the agricultural sector consumes 80% of total water abstracted, yet up to 40% is wasted through temporal mismatches between irrigation application and actual crop demand (Ministerio de Agricultura, 2024). Climate change is intensifying this challenge: precipitation variability in the Mediterranean basin has increased by 23% over the past three decades and is projected to intensify under all CMIP6 scenarios. Critically, no existing decision support system addresses the privacy–generalizability tradeoff: farmers will not share raw field data, yet models trained on single-farm data do not generalize. AgroMind directly resolves this barrier.

Gaps in current knowledge. Existing precision irrigation tools (Cropland Data Layer, AquaCrop, NetIrrigation) rely on centralized data or coarse-resolution crop models. None leverage the combinatorial signal available from simultaneous fusion of high-resolution satellite phenology, real-time soil moisture dynamics, and probabilistic weather forecasting. Federated learning has been demonstrated in medical imaging and mobile keyboards; its application to heterogeneous, asynchronous agricultural IoT networks with privacy constraints remains an open, understudied problem.

Core Innovation: AgroMind is the first system to (1) apply federated learning to multi-scale agricultural sensor fusion; (2) formally characterize convergence guarantees under non-IID farm-heterogeneity; and (3) embed climate-scenario stress-testing as a first-class validation phase. This is not an incremental extension of prior work — it reframes precision irrigation from a single-farm optimization problem to a privacy-preserving collective intelligence system.

Intellectual novelty across three dimensions:

  • Algorithmic: Asynchronous federated aggregation with topology-aware weighting (novel contribution beyond FedAvg/FedProx)
  • Systems: Edge-cloud hybrid architecture enabling inference at field level with 4G/LoRaWAN connectivity
  • Domain: First formal specification of agricultural non-IID distributions and their impact on FL convergence
  • Data: Largest labeled multimodal agricultural federated dataset in the public domain

4
Preliminary Data & Feasibility

The proposed research builds on a 2024–2025 pilot conducted across 12 citrus farms in L'Horta Sud (Valencia), demonstrating proof-of-concept for the core federated architecture:

Pilot Irrigation Savings 18% ↓ water use
Across 12 pilot farms (24 ha total), AgroMind v0.1 reduced irrigation volumes by 18.3% compared to farmer-managed controls, with no statistically significant yield difference (p = 0.72, Wilcoxon rank-sum test, n=12 pairs). This exceeds the 15% target for Phase 1 and provides strong preliminary evidence for the 25% target in the full trial.
Model Accuracy 82% F1-score
The pilot federated model achieved 82% F1-score on irrigation-need binary classification (irrigate within 48h / defer) using leave-one-farm-out cross-validation. Centralized baseline: 79% (demonstrating FL outperforms local models). Target for funded project: ≥85% with 50-farm training corpus and full Sentinel-2 integration.

Publications establishing team competence: Vidal-Roca et al. (2025) "Multimodal Deep Learning for Water Stress Detection in Citrus" — Computers and Electronics in Agriculture, 214:107821; Fuentes & Vidal-Roca (2024) "Soil Sensor Network Deployment for Precision Irrigation in Semiarid Conditions" — Agricultural Water Management, 301:108342; Sharma et al. (2025) "Federated Learning for Smart Farming: Opportunities and Challenges" — IEEE IoT Journal, 12(4):3801.

Institutional resources confirmed: UPV HPC cluster (128 NVIDIA A100 GPUs, 2 PB storage); agreements with Conselleria d'Agricultura de la Comunitat Valenciana for farm access; Purdue ACRE field station for California validation; 48 IoT sensor nodes already deployed and operational.


5
Research Approach & Methodology

Overall design. A three-year mixed-methods study combining computational model development (Years 1–3), randomized field trials (Years 1–3), and dataset curation/release (Years 2–3). A pre-registration of hypotheses and analysis plans will be filed on OSF.io prior to data collection.

Year 1 — Architecture & Deployment (Jan–Dec 2027). PI Vidal-Roca will lead development of the AgroMind federated learning backbone. Key technical milestones: (i) multimodal encoder pre-training on 5 years of historical Sentinel-2 imagery for Valencia region using masked image modeling (MAE-style, ViT-B/16 backbone); (ii) IoT-to-cloud ingestion pipeline (MQTT over LoRaWAN → Apache Kafka → Delta Lake); (iii) federated aggregation server (FedProx baseline + novel topology-weighted FedAgrо variant); (iv) IoT sensor expansion from 48 to 200 nodes across 50 farms. Co-IP Fuentes will oversee field sensor deployment; Dra. Sharma will coordinate Purdue satellite data pipeline.

Year 2 — First Field Season Validation (Jan–Dec 2028). Randomized crossover trial: 50 farms divided into 25 AgroMind-managed and 25 farmer-managed (control), with crossover at Season 2. Primary endpoint: irrigation water volume (m³/ha). Secondary: yield (kg/ha citrus; kg/ha almond), WUE, energy cost per ha. Statistical power: 80% at α=0.05 to detect 20% water reduction (25 pairs; based on pilot variance σ=4.2%). IRB-equivalent animal/environment clearance: N/A (no human subjects, no vertebrates); AEMPS environmental review completed.

Year 3 — Climate Stress-Test & Open Release (Jan–Dec 2029). Apply CMIP6 downscaled climate scenarios (SSP2-4.5, SSP5-8.5) using CORDEX-EUR-11 regional model outputs to simulate 2030–2040 precipitation and temperature regimes. Evaluate AgroMind performance degradation; develop and test rapid recalibration protocol (few-shot fine-tuning, <3 weeks per new region). Complete dataset curation, documentation, and open release. Submit final validation manuscript.

Potential problems and alternative approaches.

  • Risk: Satellite cloud cover (>60% cloud fraction) in winter months. Mitigation: SAR (Sentinel-1 C-band) fallback for soil moisture estimation; Sentinel-2 gap-fill via linear temporal interpolation validated in pilot.
  • Risk: Federated convergence failure under high farm heterogeneity. Mitigation: FedProto personalization layer; theoretical analysis of convergence bounds in Aim 1.
  • Risk: Farmer participation dropout (<50 farms). Mitigation: Signed MoU with 72 farms (40% buffer); economic incentive of free IoT hardware ($2,800/farm value).
  • Risk: Sentinel-2 API disruption. Mitigation: 3-year historical tile archive maintained on UPV storage; Earth Engine backup pipeline.

6
Timeline & Milestones
Activity / Milestone Q1 '27 Q2 '27 Q3 '27 Q4 '27 Q1 '28 Q2 '28 Q3 '28 Q4 '28 Q1 '29 Q2 '29 Q3 '29 Q4 '29
Aim 1: FL Architecture Dev
IoT Sensor Expansion (50 farms)
Aim 2: Field Trial — Season 1
Aim 2: Field Trial — Season 2 (crossover)
Aim 3: Dataset Curation
Aim 4: Climate Stress Testing
Open Dataset Release (Zenodo)
Publications (target: 4 peer-reviewed)
Broader Impacts: PhD Cohort Training

Technical Development   Field Validation   Data / Dissemination   Key Milestone


7
Broader Impacts
NSF Equal Weight
🎓
Doctoral Training
4 PhD students (2 from underrepresented groups; 1 Latinx, 1 first-generation) trained through structured co-mentorship with 3 advisors across 2 institutions. Annual symposium with industry partners.
📖
Curriculum Integration
AgroMind case study and dataset integrated into UPV Agronomy Engineering (400+ students/yr). Hackathon co-organized with Purdue reaching 200+ undergraduate students.
🌍
Open Science
100% open code (MIT license, GitHub), open data (CC-BY 4.0, Zenodo), and open access publications. Data Management Plan pre-registered on DMPTool.
💧
Water Conservation
Estimated 2.1M m³/yr water savings at pilot scale (50 farms). Policy brief delivered to Conselleria d'Agricultura for regional irrigation policy update.
🏭
Commercialization Pathway
UPV Technology Transfer Office engaged. LOI signed with AgriSense S.L. for spin-off formation contingent on Year 2 results. Estimated 50–100 FTE jobs in 5 years post-award.
🌐
International Impact
Cross-institutional collaboration with Purdue extends findings to California Central Valley (comparable climate), amplifying impact across the US agricultural sector at no additional federal cost.

8
Team Qualifications

Dr. Carmen Vidal-Roca (PI, 20% effort Y1–3) is Associate Professor of Computational Agronomy at UPV. She holds a PhD in Agronomy (UPV, 2015) and a postdoctoral fellowship at ETH Zürich (2016–2018) in precision agriculture systems. Her 47 peer-reviewed publications (h-index: 22) include foundational work on remote sensing for citrus stress detection and IoT-enabled irrigation control. She has held two prior competitive grants (MICINN PID2021, BBVA Foundation) and supervised 6 completed PhD theses.

Dr. Alejandro Fuentes (Co-IP, 15% effort Y1–3) is Professor of Soil Sciences at UPV with 25 years of field experience in Mediterranean arid zone hydrology. His expertise ensures experimental rigor in soil sensor placement, calibration, and interpretation — directly de-risking the IoT component of AgroMind.

Dra. Priya Sharma (International Collaborator, Purdue, 10% effort Y2–3) is Assistant Professor of Agricultural Informatics at Purdue. Her recent IEEE IoT Journal paper on federated learning for smart farming (35 citations, 2025) establishes her as an ideal collaborator for the FL architecture and California validation components. She brings independent NSF CAREER award funding (NSF 2512834) that partially supports her effort.


9
Budget & Justification (Summary)
3-Year Total
Category Year 1 Year 2 Year 3 Total (Direct)
Personnel (PI + Co-IP + PhD students) $72,400 $74,600 $76,900 $223,900
Fringe benefits (28%) $20,272 $20,888 $21,532 $62,692
Equipment (IoT sensors, edge nodes) $42,000 $8,000 $2,000 $52,000
Travel (conferences + farm visits) $8,500 $10,200 $9,800 $28,500
Materials & Supplies (cloud compute, reagents) $18,000 $14,000 $11,000 $43,000
Other (publication fees, participant incentives) $4,800 $6,200 $5,000 $16,000
Indirect Costs / F&A (52% MTDC, UPV negotiated rate) $24,100 $24,900 $23,658 $72,658
TOTAL REQUESTED $190,072 $158,788 $149,890 $498,750

Justification highlights: Personnel costs reflect UPV FY2027 salary scales at 20% PI effort, 15% Co-IP effort, and 100% FTE for 4 PhD students (2 funded in Y1, 2 additional in Y2). Equipment ($42K Y1) covers 152 additional IoT soil sensor nodes at $276/node (direct purchase, lowest-cost 3-quote procurement). Cloud compute ($12K/yr) covers GPU hours for FL training on Azure Research Credits supplemented by UPV HPC. Travel ($10K/yr average) funds 2 international conference presentations (NSF PI meetings + IEEE/ACM venues) and quarterly farm visits. All F&A calculated at UPV's federally negotiated 52% MTDC rate (agreement dated March 2026).


10
NSF Review Criteria — Self-Assessment
Intellectual Merit Score: 93/100
Potential to advance knowledge: High — AgroMind addresses a fundamental gap in federated ML theory (non-IID agricultural distributions) and produces the first large-scale open dataset for the domain. The theoretical convergence analysis (Aim 1) constitutes an independent scientific contribution beyond the applied system.

Team qualifications: Strong — PI's h-index 22 with domain-specific track record; complementary Co-IP expertise; international collaboration adds California validation at no marginal cost to NSF.
Broader Impacts Score: 90/100
Societal benefit: Quantified and concrete — 2.1M m³/yr water conservation, policy brief to regional government, spin-off with 50–100 job projection.

Education & diversity: 4 PhD students (2 from underrepresented groups), curriculum integration reaching 400+ students/yr, hackathon for 200+ undergraduates.

Open science: Full open release of code, data, and publications — maximizes multiplier effect of federal investment.
Innovation Score: 89/100
Three distinct innovation axes: (1) first FL application to heterogeneous agricultural IoT; (2) novel topology-weighted FedAgro aggregation algorithm with formal convergence proof; (3) climate-stress generalization as a first-class evaluation protocol. Not incremental — reframes the precision irrigation paradigm from single-farm to collective intelligence.
Feasibility / Resources Score: 87/100
Pilot data (18% water savings, 82% F1-score) directly de-risks the core approach. UPV HPC resources confirmed. 72 farmer MoUs signed (40% buffer on 50-farm target). Potential weakness: international collaboration adds coordination complexity; mitigated by weekly video syncs and co-location visits budgeted.

11
Compliance & Formatting Checklist
  • Project Description ≤15 pages (PAPPG 24-1 §II.D.2.b)
  • Project Summary: Overview, Intellectual Merit, Broader Impacts — separate paragraphs ✓
  • Arial 11pt font, 1-inch margins throughout ✓
  • References Cited: separate section, not counted in page limit ✓
  • Results from Prior NSF Support: not applicable (first submission) ✓
  • Data Management Plan: 2 pages, filed separately ✓
  • Biosketch (PI + Co-IP + Collaborator): NSF-format, 3 pages each ✓
  • Budget: automated via NSF Research.gov system ✓
  • International collaboration letter (Dra. Sharma / Purdue): pending institution sign-off — due Aug 30, 2026
  • Facilities, Equipment & Resources: 2 pages ✓