Data-driven Agriculture: Enhancing Crop Prediction with NPK Sensor Integration in Semi-arid Karnataka, India

Abstract

Background: Agriculture supports over 2.5 billion livelihoods worldwide, yet soil fertility management on Indian smallholder farms remains largely empirical. Tumakuru district’s red soil compound this problem with low organic matter and high spatial variability in N, P and K. Conventional lab testing, costs ₹350-450 and takes 3-7 days, putting it out of reach for most smallholders. Further, existing studies rarely validate sensors against lab references, report seasonal accuracy to actionable fertilization guidance. To address these issues, a low-cost edge-computing system integrated with soil sensors and machine learning is proposed. Methods: An NPK sensor and DHT22 sensor were interfaced with a Raspberry Pi 3B+ and deployed at 18 plots in Tumakuru district across kharif (June-October 2023) and rabi (November 2022-March 2023) seasons. Sensor readings were validated against laboratory reference methods (Kjeldahl for N; Olsen for P; ammonium acetate for K; n = 20 composite samples). A field-collected dataset of 3,600 instances was classified by Decision Tree and K-Nearest Neighbour (KNN) classifiers under 10-fold stratified cross-validation. A novel Confidence-Weighted Majority Voting (CWMV) ensemble was evaluated by McNemar’s test. Result: Sensor laboratory correlations were strong (N: r = 0.91; P: r = 0.87; K: r = 0.89; RMSE≤9.3%). Cross-validated accuracy was 96.7±0.4% (Decision Tree) and 96.5±0.5% (KNN). The CWMV ensemble achieved 97.3% test-set accuracy with statistically significant improvement. Seasonal accuracy varied by 0.4 percentage points. Hardware cost was ₹4,200 (≈US$50; ₹0.19 per prediction; 99.95% cheaper than laboratory testing) and energy consumption was 0.101 Wh per 1,000 inference cycles.