Deep learning based remote sensing image classification: a case study for African land cover mapping
Land cover mapping is an important research topic with wide attention in the remote sensing domain. Traditional classification methods such as SVM and Random Forest have been playing an important role in this field for many years. In this research, we demonstrate our early efforts on applying deep learning based classification method to large-scale land cover mapping. Based on the Stacked Autoencoder (SAE), one of the deep learning models, we build our classification framework for large-scale remote sensing image processing. The main parameters are adjusted and optimized based on our testing samples. Results show that the SAE classifier trained with an entire set of African training samples achieves an overall classification accuracy of 78.99% when assessed by test samples collected independently, which is higher than the accuracy achieved by other three classifiers (76.03%, 77.74%, 77.86% of RF, SVM, ANN respectively) based on the same set of test samples. The mapping results of SAE are obviously better than those obtained from other classifiers as well.
The overall workflow of this study is showed in Figure 1. First, three traditional classifiers – RF, SVM, ANN and one deep learning classifier – SAE are implemented. Next, training and testing samples are prepared and loaded into each classifier separately. Then the main parameters of each classifier are adjusted continuously until we find the best combination of parameters of which the overall accuracy is the highest on our testing samples. By tuning the parameters, we achieve the best model of each classifier and save it for further use. For the process of land cover mapping, firstly nine features (spectral reflectance of Landsat 7/8, NDVI and MNDWI) of each pixel in a scene are extracted and loaded into our four classifiers. Then we use the best model of each classifier obtained previously to predict the label for each pixel and produce the land cover maps.
Figure 1. Overall Workflow
The experiment results of the SAE-based approach are analyzed and compared against RF, SVM and ANN from many different aspects. Each classifier is trained based on our 534,396 training samples and the main parameters of each classifier are adjusted until the overall accuracy is the highest on our 26,282 testing samples. The classification accuracy and statistical significance tests of each method are compared in table 1. The training, predicting and mapping time comparisons between each method are listed in table 2. The parameters’ impact on the performance of SAE is shown in figure 2. Experiment results show that when the network depth is set as 3 and the number of hidden units is set as 48, we can get the highest overall accuracy of 78.99% after 268 iterations. The final land cover mapping results of a representative scene obtained from each classifier are shown in figure 3.
Table 1. Classification Accuracy Comparison
Table 2. Running Time Comparison
Figure 2. Parameters Adjustment
Figure 3. Mapping Results Comparison