Search

This study proposes a light convolutional neural network (LCNN) well-fitted for medium-resolution (30-m) land-cover classification. The LCNN attains high accuracy without overfitting, even with a small number of training samples, and has lower computational costs due to its much lighter design compared to typical convolutional neural networks for high-resolution or hyperspectral image classification tasks. The performance of the LCNN was compared to that of a deep convolutional neural network, support vector machine (SVM), k-nearest neighbors (KNN), and random forest (RF). SVM, KNN, and RF were tested with both patch-based and pixel-based systems. Three 30 km × 30 km test sites of the Level II National Land Cover Database were used for reference maps to embrace a wide range of land-cover types, and a single-date Landsat-8 image was used for each test site. To evaluate the performance of the LCNN according to the sample sizes, we varied the sample size to include 20, 40, 80, 160, and 320 samples per class. The proposed LCNN achieved the highest accuracy in 13 out of 15 cases (i.e., at three test sites with five different sample sizes), and the LCNN with a patch size of three produced the highest overall accuracy of 61.94% from 10 repetitions, followed by SVM (61.51%) and RF (61.15%) with a patch size of three. Also, the statistical significance of the differences between LCNN and the other classifiers was reported. Moreover, by introducing the heterogeneity value (from 0 to 8) representing the complexity of the map, we demonstrated the advantage of patch-based LCNN over pixel-based classifiers, particularly at moderately heterogeneous pixels (from 1 to 4), with respect to accuracy (LCNN is 5.5% and 6.3% more accurate for a training sample size of 20 and 320 samples per class, respectively). Finally, the computation times of the classifiers were calculated, and the LCNN was confirmed to have an advantage in large-area mapping.

Thermal data products derived from remotely sensed data play significant roles as key parameters for biophysical phenomena. However, a trade-off between spatial and spectral resolutions has existed in thermal infrared (TIR) remote sensing systems, with the end product being the limited resolution of the TIR sensor. In order to treat this problem, various disaggregation methods of TIR data, based on the indices from visible and near-infrared (VNIR), have been developed to sharpen the coarser spatial resolution of TIR data. Although these methods were reported to exhibit sufficient performance in each study, preservation of thermal variation in the original TIR data is still difficult, especially in fire areas due to the distortion of the VNIR reflectance by the impact of smoke. To solve this issue, this study proposes an efficient and improved disaggregation algorithm of TIR imagery on wildfire areas using guided shortwave infrared (SWIR) band imagery via a guided image filter (GF). Radiometric characteristics of SWIR wavelengths could preserve spatially high frequency temperature components in flaming combustion, and the GF preserved thermal variation of the original TIR data in the disaggregated result. The proposed algorithm was evaluated using Landsat-8 operational land imager (OLI) and thermal infrared sensor (TIRS) images on wildfire areas, and compared with other algorithms based on a vegetation index (VI) originating from VNIR. In quantitative analysis, the proposed disaggregation method yielded the best values of root mean square error (RMSE), mean absolute error (MAE), correlation coefficient (CC), erreur relative globale adimensionelle de synthèse (ERGAS), and universal image quality index (UIQI). Furthermore, unlike in other methods, the disaggregated temperature map in the proposed method reflected the thermal variation of wildfire in visual analysis. The experimental results showed that the proposed algorithm was successfully applied to the TIR data, especially to wildfire areas in terms of quantitative and visual assessments.