Inspiration
Agriculture runs on uncertainty. Farmers commit to seeds and inputs months before they know what the season will bring. Lenders underwrite crop loans without knowing what is actually in the ground. Cropora was built to close that information gap using eleven years of satellite data that already exists and is largely underutilized at the field level.
What it does
Cropora is a field intelligence platform for Iowa cropland built on three public satellite datasets spanning 2013 to 2023.
It does four things: analyzes crop-specific NDVI phenology for corn and soybean across the growing season; classifies every Iowa cropland pixel into regular rotation, monoculture, or irregular using eleven years of USDA CDL data; trains two Random Forest models for crop type prediction (78% accuracy using NDVI features alone, 96% using CDL history); and projects the 2024 crop type for every field with a pixel-level confidence score.
All of this feeds into a Field Resilience Score, a composite 0-100 metric combining rotation consistency, crop diversity, and recent stability, giving lenders and agronomists a single actionable number before planting season begins.
How we built it
Three Python notebooks in Colab, one per analytical task, with a Streamlit dashboard for the final interface. CDL rasters were downloaded via the USDA CropScape API for all eleven years. NDVI composites came from NASA AppEEARS (MOD13Q1.061) and soil moisture from SMAP L4 (SPL4SMGP.008). All datasets were reprojected and spatially aligned using rasterio. The dashboard uses Streamlit and Plotly with an interactive Iowa region map, a what-if rotation threshold slider, and a field intelligence lookup with plain-English lending recommendations per region.
Challenges ran into
The eleven-year CDL stack at native 30m resolution produced an array nearly 4.7GB in memory, crashing Colab repeatedly. Downsampling to 300m using rasterio reduced the stack to 47MB while preserving the rotation signal at agronomically meaningful resolution. The CropScape API also required bounding boxes in EPSG:5070 rather than WGS84, a silent failure mode that cost significant debugging time.
Accomplishments
96% accuracy on the CDL history model with a nearly diagonal confusion matrix, and a predicted 2023 crop map visually indistinguishable from CDL ground truth. The Field Resilience Score synthesizes three independent satellite data streams into one number that a non-technical lender or insurer can actually act on.
What was learned
Geospatial preprocessing takes longer than the modeling. The 18-point accuracy gap between the NDVI-only model and the CDL history model was the most interesting finding: agronomic memory, what a field grew last year and the year before, is the dominant signal for predicting what it will grow next. Single-season satellite imagery is informative but incomplete.
What's next for Cropora - Field Intelligence Platform
Expanding to the full CONUS corn belt, incorporating a multi-year SMAP baseline for true anomaly detection, and publishing the Field Resilience Score as an annually updated county-level risk index timed to release each spring before planting season.
Log in or sign up for Devpost to join the conversation.