A deep learning based novel spectral matching technique for orchards classification using spectrotemporal signatures leveraging multidimensional data format

Abstract

Orchard classification is vital for precision agriculture, enabling crop monitoring and sustainable land management. Despite progress in classification techniques, deep learning based spectral matching for orchard classification remains underexplored. This study introduces a deep learning based novel spectral matching method, the Time-Aware Siamese Network (TASN), to improve orchard classification through Spectrotemporal Signatures (STS) derived from multi-temporal vegetation indices (VI) stored in Multidimensional Data (MDD) format. Focusing on Khairpur, Pakistan, we fused Landsat-8/9 and Sentinel-2 datasets using Google Earth Engine (GEE) to create cloud-free monthly composites. Three VI were computed and different seasonal data cubes were generated to extract STS. Wavelet-based smoothing effectively reduces noise and enhances STS quality. TASN outperformed Spectral Angle Mapper (SAM), achieving 93% accuracy with UNVI based Winter–Spring–Summer (WSS) and Spring–Summer–Autumn (SSA) data cubes and 94% with Winter–Spring–Summer–Autumn (WSSA). Notably, UNVI using TASN slightly surpassed NDVI and EVI across all data cubes. The model’s dual Long Short Term Memory (LSTM) layers effectively distinguished spectrally similar orchards, such as mango and banana, that challenge the SAM approach. These outcomes highlight the potential of deep learning architecture for spectral matching, providing a robust framework for improving orchard classification and advancing agricultural applications.