A Hybrid Spectral–Spatial Deep Learning Framework with Harris Hawk Optimization and Support Vector Machine for Accurate Arecanut Plantation Mapping Using Sentinel-2 Imagery

Sumithra C. V, Manjula T. R.

Spectral similarity with other perennial vegetation and heterogeneous agricultural landscapes still make accurate identification of arecanut plantations from medium resolution satellite imagery a challenge. In this paper, a Hybrid Spectral–Spatial DeepLabV3+ with Harris Hawk Optimization and Support Vector Machine (HSS-DEEPLAB-HHO-SVM) framework is presented for robust arecanut plantation segmentation and land-cover classification using Sentinel-2 multispectral imagery. The proposed method combines adaptive spectral feature extraction, vegetation indices, optimized DeepLabV3+ based spatial feature learning (Adaptive Atrous Spatial Pyramid Pooling (A-ASPP), Channel-Spatial Attention (CSAM) and adaptive spectral-spatial feature fusion) and Harris Hawk Optimization (HHO) for hyperparameter tuning and then utilizes SVM classifier for better decision boundary optimization. The experimental results on the arecanut dataset achieved an Overall Accuracy of 98.54% and a Mean Intersection over Union of 96.93% with a mere 0.20% relative deviation in plantation area estimation from official statistics. Moreover, evaluation on the Indian Pines hyperspectral benchmark showed excellent generalization capability, achieving 99.77% Overall Accuracy. The proposed framework provides an efficient and scalable solution for precision agriculture, crop inventory generation and remote sensing based land-cover mapping.