Transfer Learning Models and Datasets for a Reliable Emergency Landing Field Identification
| Main Author: | Andreas Klos |
|---|---|
| Format: | info dataset Journal |
| Bahasa: | eng |
| Terbitan: |
, 2020
|
| Subjects: | |
| Online Access: |
https://zenodo.org/record/4117572 |
Daftar Isi:
- The file data.tar.gz compromises three HDF5 datasets. This file has been split into 100 files. The files can be merged, decompressed and unpacked with the following commands: cat data* > data.tar.gz tar -xzf data.tar.gz Afterwards, the three files: train_test_data_ss8_supervised_new.hdf5, train_test_data_ss16_supervised_new.hdf5, train_test_data_ss32_supervised_new.hdf5 are ready to get processed. Internal structure of the datasets: Search Window (SW) 8 m^2: HDF5 "train_test_data_ss8_supervised_new.hdf5" { GROUP "/" { GROUP "test" { DATASET "fus_data" { DATATYPE H5T_IEEE_F32LE DATASPACE SIMPLE { ( 76288, 40, 40, 8 ) / ( 76382, 40, 40, 8 ) } } DATASET "labels" { DATATYPE H5T_STD_I64LE DATASPACE SIMPLE { ( 76382, 1 ) / ( 76382, 1 ) } } } GROUP "train" { DATASET "fus_data" { DATATYPE H5T_IEEE_F32LE DATASPACE SIMPLE { ( 380928, 40, 40, 8 ) / ( 380998, 40, 40, 8 ) } } DATASET "labels" { DATATYPE H5T_STD_I64LE DATASPACE SIMPLE { ( 380998, 1 ) / ( 380998, 1 ) } } } }} SW 16 m^2: HDF5 "train_test_data_ss16_supervised_new.hdf5" { GROUP "/" { GROUP "test" { DATASET "fus_data" { DATATYPE H5T_IEEE_F32LE DATASPACE SIMPLE { ( 17024, 80, 80, 8 ) / ( 17054, 80, 80, 8 ) } } DATASET "labels" { DATATYPE H5T_STD_I64LE DATASPACE SIMPLE { ( 17054, 1 ) / ( 17054, 1 ) } } } GROUP "train" { DATASET "fus_data" { DATATYPE H5T_IEEE_F32LE DATASPACE SIMPLE { ( 84992, 80, 80, 8 ) / ( 85068, 80, 80, 8 ) } } DATASET "labels" { DATATYPE H5T_STD_I64LE DATASPACE SIMPLE { ( 85068, 1 ) / ( 85068, 1 ) } } } }} SW 32 m^2: HDF5 "train_test_data_ss32_supervised_new.hdf5" { GROUP "/" { GROUP "test" { DATASET "fus_data" { DATATYPE H5T_IEEE_F32LE DATASPACE SIMPLE { ( 3328, 160, 160, 8 ) / ( 3359, 160, 160, 8 ) } } DATASET "labels" { DATATYPE H5T_STD_I64LE DATASPACE SIMPLE { ( 3359, 1 ) / ( 3359, 1 ) } } } GROUP "train" { DATASET "fus_data" { DATATYPE H5T_IEEE_F32LE DATASPACE SIMPLE { ( 16768, 160, 160, 8 ) / ( 16793, 160, 160, 8 ) } } DATASET "labels" { DATATYPE H5T_STD_I64LE DATASPACE SIMPLE { ( 16793, 1 ) / ( 16793, 1 ) } } } }} The sample count of the various generated dataset is as follows: SW 8 m^2: {train: 380,928 with {0: 190,464, 1: 190,464}, test: 76,288 with {0: 38,152, 1: 38,136}} SW 16 m^2: {train: 84,992 with {0: 42,498, 1: 42,494}, test: 17,024 with {0: 8,516, 1: 8,508}} SW 32 m^2: {train: 16,768 with {0: 8,424, 1: 8,344}, test: 3,328 with {0: 1,672, 1: 1,656}} Each sample is composed as follows: RGB = sample[:,:,:3]; Theoretically: [0, 1] per color channel NIR = sample[:,:,3]; Theoretically: [0, 1] Slope = sample[:,:,4]; Theoretically: [0, 90] Roughness = sample[:,:,5]; Theoretically: [0, 78.78] NDVI = sample[:,:,6]; Theoretically: [-1, 1] DOM = sample[:,:,7]; Theoretically: [0, 429.90] ==================================================================================================== The following three files compromise the model and optimizer state variable of our PyTorch models trained on the aforementioned datasets: best_alexnet_final.pth, best_resnet18_final.pth, best_wide_resnet50_2_final.pth Below find a more precise description of each model: best_resnet18_final.pth Model: ResNet-18 Dataset: SW 8 Input: RGB-NIR-Slope -> R: [0,224,224], G: [1,224,224], B: [2,224,224], NIR: [3,224,224], Slope: [4,224,224] best_wide_resnet50_2_final.pth Model: Wide-ResNet-50-2 Dataset: SW 16 Input: NDVI-Slope -> NDVI: [0,224,224], Slope: [1,224,224] best_alexnet_final.pth Model: AlexNet Dataset: SW 32 Input: RGB-Slope -> R: [0,224,224], G: [1,224,224], B: [2,224,224], Slope: [3,224,224] Each model is capable of performing a binary classification, distinguishing between landable and unlandable samples