Scholarlink

<p>Traditional soil testing methods, while reliable, are often time-consuming and costly for large-scale applications. This study proposes a machine learning–based approach to predict essential soil nutrients in Karnataka, India, using Landsat-8/9 satellite imagery and soil health card data. We developed the KSHSID dataset with 168,000 location-specific instances combining soil and satellite features across districts. To enhance prediction, both global geolocation and local neighborhood context were incorporated. Three models–K-Nearest Neighbors (KNN), Multi-Layer Perceptron (MLP), and Light Gradient Boosting Machine (LightGBM)–were evaluated. LightGBM outperformed others, showing the benefits of integrating remote sensing with contextual data. The model achieved strong regression performance with <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\textrm{R}^{2}\)</EquationSource> </InlineEquation> scores of 0.77 (N), 0.61 (P), 0.72 (K), 0.62 (B), 0.75 (Fe), 0.63 (Zn), 0.70 (Cu), 0.65 (S), 0.76 (Mn), 0.71 (OC), 0.79 (pH), and 0.98 (EC). For classification, it produced high F1-scores: 0.85 (N), 0.82 (P), 0.83 (K), 0.84 (B), 0.88 (Fe), 0.85 (Zn), 0.95 (Cu), 0.89 (S), 0.91 (Mn), 0.80 (OC), 0.77 (pH), and 0.99 (EC). Based on N, P, K, and OC predictions, crop-fertilizer recommendations were generated for 119 crop varieties. These results demonstrate the potential of satellite and soil-health integration to reduce dependence on lab testing and support precision farming.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Smart Agriculture: Spatial Context-Based Nutrient Prediction and Crop-Fertilizer Recommendations Using Remote Sensing and Soil Health Data

  • 期刊论文

  • 发表日期：

    2026年3月25日

查看全文
  • Chirag Gupta,
  • Nagamma Patil

摘要

Traditional soil testing methods, while reliable, are often time-consuming and costly for large-scale applications. This study proposes a machine learning–based approach to predict essential soil nutrients in Karnataka, India, using Landsat-8/9 satellite imagery and soil health card data. We developed the KSHSID dataset with 168,000 location-specific instances combining soil and satellite features across districts. To enhance prediction, both global geolocation and local neighborhood context were incorporated. Three models–K-Nearest Neighbors (KNN), Multi-Layer Perceptron (MLP), and Light Gradient Boosting Machine (LightGBM)–were evaluated. LightGBM outperformed others, showing the benefits of integrating remote sensing with contextual data. The model achieved strong regression performance with \(\textrm{R}^{2}\) scores of 0.77 (N), 0.61 (P), 0.72 (K), 0.62 (B), 0.75 (Fe), 0.63 (Zn), 0.70 (Cu), 0.65 (S), 0.76 (Mn), 0.71 (OC), 0.79 (pH), and 0.98 (EC). For classification, it produced high F1-scores: 0.85 (N), 0.82 (P), 0.83 (K), 0.84 (B), 0.88 (Fe), 0.85 (Zn), 0.95 (Cu), 0.89 (S), 0.91 (Mn), 0.80 (OC), 0.77 (pH), and 0.99 (EC). Based on N, P, K, and OC predictions, crop-fertilizer recommendations were generated for 119 crop varieties. These results demonstrate the potential of satellite and soil-health integration to reduce dependence on lab testing and support precision farming.