A Multi-Source Spatiotemporal Framework for Vegetation Anomaly Detection in Solar Photovoltaic Fields Using Hierarchical Labels and Hybrid Deep Learning

Abstract

Moroccan installations of solar photovoltaic panels experience operational difficulties due to shading and vegetation-related soiling, which reduce energy output by 15–30%. Most monitoring systems depend upon a single vegetation index, which can reduce the accuracy of detecting even moderate anomalies. This paper presents a novel integration of multi-criteria hierarchical labeling with dual-branch deep learning for enhanced vegetation anomaly detection. We combined MODIS (2000–2015) and Sentinel-2 (2015–2025) images and NASA POWER weather records to study a 25-year vegetation record using multi-source satellite data in 5 of Morocco’s ecologically diverse zones. We introduced a three-class hierarchical labeling scheme (normal, moderate, severe) for dynamic vegetation models based on combined vegetation indices (NDVI, EVI, NDWI) and meteorological thresholds. The proposed dual-branch architecture uses independent data streams for unfused data, which include temporal multi-scale CNNs (TMSCNN) for spatiotemporal modeling and bidirectional LSTMs for weather-integrated vegetation data. Systematic ablation studies show improvements from using NDVI (68.98%) to multispectral indices (77.74%), meteorological integration (81.02%), and a final accuracy of 82.34% ± 0.88%. The moderate anomaly class exhibits lower precision (65%), demonstrating the challenge of operationalizing severity-based anomaly classification. This work integrates hierarchical, multi-criteria labeling and hybrid deep learning for solar photovoltaic vegetation monitoring.