A Methodological Proposal for SHAP-Explained Cocoa Yield Prediction Using Sentinel-2 Multispectral Time Series and IoT Sensors: A Neutrosophic Digital Twin Component for EUDR Traceability in Ecuador

Resumen

This paper presents a methodological proposal for an explainable machine learning pipeline that integrates Sentinel-2 multispectral time series (NDVI/EVI with SCL-based cloud quality control), IoT agroclimate sensors, and INAMHI official climate data as inputs to Random Forest and XGBoost cocoa yield prediction models, with SHAP (SHapley Additive exPlanations) as the primary interpretability framework. The pipeline is proposed as Module 4 (M4) of a Multisource Geospatial Digital Twin whose architecture was formally selected in the companion paper [14] using N-AHP-TOPSIS (edge-cloud, CC = 0.7625). To make incomplete and conflicting traceability evidence explicit, the proposal adds a single-valued neutrosophic evidence layer in which each lot is described by independent confirmation, indeterminacy, and contradiction components . The indeterminacy component records satellite gaps, incomplete productive records, and disagreement between data sources instead of absorbing them into a scalar compliance score. Design targets are  ≥ 0.70 and RMSE ≤ 100 kg/ha for yield prediction, ≥ 95% lot coverage with valid GeoJSON polygons for EUDR Due Diligence Statement generation, and ≥ 60% reduction in traceability evidence preparation time. A Sentinel-2 QA/QC scheme based on pix_validos_pct per lot per date explicitly handles Guayas coastal cloudiness, converting data availability gaps into quantified, audit-traceable quality metrics. Expert validation yields a global V-Aiken coefficient of 0.865. Empirical validation through a pilot with UNOCACE (Yaguachi, Guayas, 1,000+ producers) is the planned next phase.