Harnessing Artificial Intelligence in Modern Agriculture: Advancements in Smart Farming and Precision Technologies

Abstract

 AgriSaarthi combines MobileNetV2 powered leaf disease detection and granite-3-8b-instruct RAG advisory, with test-set accuracy (weighted F1: 0.948) of 38 disease classes in the New Plant Diseases Dataset (87,000 images, 26 crops). The system obtains per-class F1 of over 0.97 on tomato (9 diseases such as late blight, leaf mold, yellow leaf curl), grape (4 disease such as black rot, leaf blight), and apple (4 diseases such as scab, cedar apple rust) which are the best categories since most of them are trained. This system integrates MobileNetV2 based convolutional model with a diagnostic accuracy of 95.2\% and an average inference latency of 45ms and the retrieval grounding enhanced the factual accuracy of advisory output by 18.5\% and lowered the frequency of hallucinations by 66.7\% compared to combined language models. This is factually more accurate by 18.5 percentage points over granite-3-8b-instruct without RAG (94.7\% vs 76.2\%) and with less hallucinations by 66.7 percentage points (4.0\% vs 12.0\%) on 500 agricultural queries rated by domain experts. The trilingual (English, Hindi, and Marathi) support of native systems avoids the loss of context and preserved meaning which is characteristic of pipeline systems. Grad-CAM heatmap and structured source attribution develop farmer confidence. Evidence contrary to expectations, in a 12-week field study in Pune and Nashik districts (N=50 smallholder farmers, 0.5-3 acres, 72\% of speakers via Marathi), Wilcoxon-significant confidence gains of the intervention over text-only baselines were found ($Z=-5.47$, $p<0.001$, $r=0.77$ effect). This provides a feasible roadmap of explainable multimodal agricultural AI to serve resource constrained rural areas.