Scaling agroforestry land suitability analysis in Odisha, India: a machine learning approach

Rajkumar Singh, Shalini Gakhar, Pulakesh Das, A. Jaya Prakash, Mukunda Behera, Chandrashekhar Biradar, Sujoy Mudi, Venkatesh Kolluru, Atul Dogra, Shiv Kumar Agrawal, Shiv Kumar Dhyani. (1/6/2026). Scaling agroforestry land suitability analysis in Odisha, India: a machine learning approach. Frontiers in Remote Sensing, 7.

Abstract

Agroforestry practices are one of the major pillars of Natural Resources Management (NRM), offering substantial benefits by improving environmental conditions, socio-economy, soil health, biodiversity, and climate resilience. Despite multi-dimensional benefits, robust data for regional planning and effective implementation are lacking, particularly in combined with existing land management practices. Further, analyses are often conducted using a set of reference datasets without validating the scalability. This study aimed to address this gap by leveraging a pre-trained machine learning (ML)-based Multi-Criteria Evaluation (MCE) model to assess agroforestry land suitability in Odisha, India, and validating the outcome with reference data from an independent region. Multi-temporal Sentinel-2 data were used to generate Land Use Land Cover (LULC) and cropping intensity maps, where Random Forest (RF) achieved above 94% classification accuracy, outperforming Support Vector Machine (SVM; above 93%). In comparison to traditional expert-defined weights, RF model-derived variable importance was integrated with fuzzy-MCE approach for agroforestry site suitability analysis. A diverse data array, such as topography, soil parameters, climate conditions, and socioeconomic factors, were employed, wherein the proximity variables contributed ∼70% of total weight. Independent validation using field data from another region with similar agricultural practices and socio-economic characteristics showed high mean suitability (>0.87; range 0.71–0.95). Moreover, the generated Receiver Operating Characteristic (ROC) analysis indicated an Area Under the Curve (AUC) of 0.89, exhibits strong model performance and high capability to site suitable agroforestry sites. Intervention-specific analysis indicated that >94% of double- and single-cropped lands, ∼96% of settlement areas, and >90% of permanent fallow and wastelands were found highly suitable for (i) bund and boundary plantations and intercropping, (ii) home gardens, and (iii) block/bulk plantation agroforestry practices, respectively. Further, we deployed the site suitability layer through a web platform. The developed WebGIS portal enables open data access, spatial querying and intervention planning, providing a practical decision-support tool for state-level agroforestry planning and implementation.