Remote Sensing Image Land Cover Classification Method Based on Improved CNN Algorithm

The powerful feature learning capability of the AlexNet model of convolutional neural networks is gradually applied in land classification, but the model suffers from problems such as inconsistency between sample form and network requirements in its application. Therefore, this study proposes the optimized AlexNe-finetunet model, LCNet-27 model and LCNet-13 model based on AlexNet model to solve the problems of oversized training samples and model training input size, speed up model training and improve classification accuracy. The highest accuracy was 97.76% for the LCNet-27 model and 95.33% for the LCNet-13 model with 5×5 pixel input size, both higher than that of the AlexNet model finetune, 93.6% of the AlexNet model finetune. The LCNet-27 and LCNet-13 models obtained by optimizing the AlexNet model finetune with this study have improved the accuracy and faster classification speed in remote sensing image land classification, solving the problem of contradiction between sample size and model input size of the AlexNet model, which can be well applied in land classification problem. The novelty of this work lies in the innovative use of relative models for different studies for multiple aspects of land classification problems, using the AlexNe-refinement model, the LCNet-27 model, and the LCNet-13 model for a more comprehensive land classification study.