Sentinel-1 SAR and Sentinel-2 optical land-cover classification with a prototype decision-support dashboard for the monsoon-driven Cauvery Delta

Introduction Persistent monsoon cloud cover makes optical satellite monitoring unreliable over deltaic farmland, and documented reference data for checking classification accuracy are scarce. We tested how well Sentinel-1 SAR and Sentinel-2 optical features separate Rice, Vegetation and Built-up land in the Cauvery Delta, Tamil Nadu, India (10.85°N–11.15°N, 79.05°E–79.45°E), over the June 2024–May 2025 agricultural year. Methods We drew a simple random sample of 500 points and labelled each by visual interpretation of high-resolution imagery, checked against seasonal Sentinel-2 composites. After excluding points outside the three classes, 449 remained (149 Rice, 256 Vegetation, 44 Built-up); a second interpreter labelled the first 100 points independently. At each point, Google Earth Engine supplied annual NDVI, annual Sentinel-1 VV, VH and VV/VH values, and a 29-date Sentinel-1 time series condensed into 12 summary statistics. We compared logistic regression, Random Forest and a fully connected neural network (HydraSAR-DNN) using 5-fold cross-validation, repeated three times, with folds built either at random or from 1 km spatial blocks. Results Under 1 km block cross-validation, NDVI on its own reached only 51.6–55.8% accuracy, below the 57.0% majority-class baseline. Adding annual SAR backscatter lifted accuracy to 62.7–67.0%, and Sentinel-1 time-series summaries with NDVI lifted it further to 70.7–72.4%. The best model was logistic regression (72.4%, Cohen’s κ 0.475; McNemar p = 0.001 against the annual-feature network). Spatial blocking shifted accuracy by no more than 2.4 percentage points, and the neural network never beat the best simpler classifier. The two interpreters agreed on 68.4% of points (κ 0.52), disagreeing mostly over Rice and Vegetation, which is also where the classifiers went wrong. Discussion In this setting, the reliability of the reference labels, not model complexity, sets the ceiling on accuracy. A prototype dashboard shows the classifier output next to an ET₀-based irrigation calculator, a rule-based pest-risk indicator and trilingual SMS templates; these modules run on placeholder inputs and have not been field-tested. The framework is intended to support SDG 2 (Zero Hunger), SDG 6 (Clean Water and Sanitation) and SDG 13 (Climate Action), although field trials are needed before any such contribution can be claimed. The reference sample, extracted features and code are openly released.