Scholarlink

<p>Agroforestry systems hold promise for diversifying cereal-dominated rotations in Punjab by integrating high-value timber species with staple crops. Yet conventional rule-based GIS approaches often yield overly generalized suitability maps, limiting their utility for site-specific decision-making. Here, we develop a data-driven framework using deep neural networks (DNNs) to predict habitat suitability for ten key agroforestry tree species across Punjab. We combined 19 WorldClim bioclimatic layers and seven ISRIC SoilGrids edaphic variables into a 26-predictor raster stack at 1&#xa0;km resolution. A feed-forward DNN (input: 26 units; hidden layers: 32 and 16 ReLU (rectified linear unit) units with 30% and 20% dropout; output: sigmoid) was trained on presence–pseudo-absence data (80/20 split with 20% validation) and evaluated via ROC AUC, achieving an average AUC of 0.844 (range: 0.786–0.929). Spatial projections in this study highlighted high suitability areas for Populus deltoides along riparian corridors and identifying marginal areas where drought-tolerant species may be preferred. Integrating SHapley Additive exPlanations (SHAP) revealed that precipitation and temperature seasonality, along with soil pH and organic carbon, were the gnificant sidrivers of species-specific suitability. Non-metric multidimensional scaling combined with hierarchical clustering further grouped species into ecological guilds, guiding assemblage-level selection. Our explainable DNN approach delivers high-resolution, explanable suitability maps, offering a transparent analysis for climate-resilient agroforestry expansion in Punjab.</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

Modeling agroforestry tree species suitability in Punjab using deep neural networks and SHAP analysis

  • 期刊论文

  • 发表日期：

    2026年5月27日

查看全文
  • Neelesh Yadav,
  • Mahesh Manchanda,
  • Prateek Srivastava

摘要

Agroforestry systems hold promise for diversifying cereal-dominated rotations in Punjab by integrating high-value timber species with staple crops. Yet conventional rule-based GIS approaches often yield overly generalized suitability maps, limiting their utility for site-specific decision-making. Here, we develop a data-driven framework using deep neural networks (DNNs) to predict habitat suitability for ten key agroforestry tree species across Punjab. We combined 19 WorldClim bioclimatic layers and seven ISRIC SoilGrids edaphic variables into a 26-predictor raster stack at 1 km resolution. A feed-forward DNN (input: 26 units; hidden layers: 32 and 16 ReLU (rectified linear unit) units with 30% and 20% dropout; output: sigmoid) was trained on presence–pseudo-absence data (80/20 split with 20% validation) and evaluated via ROC AUC, achieving an average AUC of 0.844 (range: 0.786–0.929). Spatial projections in this study highlighted high suitability areas for Populus deltoides along riparian corridors and identifying marginal areas where drought-tolerant species may be preferred. Integrating SHapley Additive exPlanations (SHAP) revealed that precipitation and temperature seasonality, along with soil pH and organic carbon, were the gnificant sidrivers of species-specific suitability. Non-metric multidimensional scaling combined with hierarchical clustering further grouped species into ecological guilds, guiding assemblage-level selection. Our explainable DNN approach delivers high-resolution, explanable suitability maps, offering a transparent analysis for climate-resilient agroforestry expansion in Punjab.