Integrated Feature Learning and Decision Modeling for Land Use and Land Cover Analysis

Training the neural network using high-quality labeled data is a major challenge, as there are many grey areas in image classification. Hence, experts are available to label the data. Proposed in this research paper is a Heterogeneous Multi-Stream Deep Learning Framework. Its application of six advanced CNNs will remove such complications using their complementary inductive biases. An evaluation of two fusion paradigms was successfully achieved: a decision-level weighted average ensemble and a multi-stream CNN-SVM at both feature levels. The feature-level fusion method was found to be more discriminative than probabilistic averaging in testing across three different datasets: NWPU-RESISC45, UC Merced, and AID. This approach achieves the best results on all datasets, with our method achieving a highest F1-score of 97.24% on the NWPU-RESISC45 benchmark. The performance of the six-stream design, having 8192 features, was slightly affected, dropping to 97.15% because of the curse of dimensionality. The findings support the five-stream CNN-SVM as the best architecture since it easily strikes a balance between feature richness and the complexity of the classifier.