Sensor-Aware Benchmarking of Deep Convolutional Neural Networks for Land Use and Land Cover Classification Using Landsat-8 and Sentinel-2 Imagery

Land Use and Land Cover (LULC) classification benefits many fields, including environmental studies, urban planning, precision agriculture, and sustainable use of natural resources. The successful application of deep convolutional neural networks (CNNs) has improved satellite image classification accuracy. However, CNNs are highly sensitive to the spatial properties of input images. LULC classification using multispectral satellite images of Sentinel-2 and Landsat-8 was benchmarked for sensor-adaptive performance. In this study, four well-known CNN variants—VGG-16, MobileNet, GoogLeNet, and DenseNet-121—were used to establish the performance of the models for the Guntur district of Andhra Pradesh, India. The geographical area is approximately 2,443 km² and is located approximately at a latitude of 16.2992° N and a longitude of 80.4320° E. From the co-registered multi-date multispectral images, about 5000 equally sized patches were generated representing five LULC classes (water, vegetation, agriculture, urban/built-up area, and bare soil). The datasets were generated using the 10 m resolution bands of Sentinel-2 and selected Landsat-8 bands resampled to a common 10 m spatial grid. The models were trained and validated with an 80:20 split and tested on multi-scene Sentinel-2 and Landsat-8 images to evaluate their generalization capability over different acquisition dates and scenes. To ensure a consistent and fair comparison between the two datasets, all four models were trained with the Adam optimizer, a learning rate of 1 × 10⁻⁴, a mini-batch size of 16, eight training epochs, and cross-entropy loss. The results indicated that the best performance was obtained using the DenseNet-121 architecture on Sentinel-2 image data, with an overall accuracy, F1-score, and Cohen’s kappa of 97.22%, 96.50%, and 0.9645, respectively. VGG-16 achieved the best results for the Landsat-8 images resampled to the common 10 m spatial grid, with the highest overall accuracy (96.08%), F1-score (96.69%), and Cohen's kappa (0.9506) among all the models. The lightweight architectural design and competitive classification performance of MobileNet may make it a practical choice for large-scale and resource-constrained applications. These results highlight how the performance of CNN algorithms relies heavily on the spatial characteristics of the sensors and that there is no universally best-performing CNN architecture for satellite images. The benchmark delivers general guidance to practitioners involved in selecting deep learning algorithms for operational LULC mapping and provides insights for the development of efficient sensor-aware remote sensing algorithms.