The implementation of 3DFGC and other models: The environment is pytorch1.0 and python3.6, the code is not tested with other environments. All models are trained on a single NVIDIA GeForce GTX 1060 GPU with 6 GB RAM.
The experimental data: The data was captured from the GaoFen 2 (GF2) satellite. For a detailed introduction of GF2 remote sensing data, please refer to: http://www.cresda.com/EN/ The GF2 images have four bands, red, green, blue and near infrared, with 4 m ground resolution. The two data sets were captured in 2015 and 2017, respectively. The data of 2015 has a pixel size of 1417×2652 and consists of four images captured in June, July, August and September. The data of 2017 has a pixel size of 2102×1163 and contains seven images captured in June, July, August, September, October, November and December. Handcrafted shape files with land cover information are used as a reference for classification.
| RGB | Category |
|---|---|
| (248, 248, 0) | sorghum |
| (0, 0, 248) | rice |
| (248, 0, 0) | tree |
| (0, 248, 0) | corn |
| (0, 0, 0) | background |
Train dataset: The patch-size we used is 256×256, and the center point of each patch is given at: data/2015data/trainDataset2015.txt. The format is: row value/ column value; it is also visualized at: data/2015data/2015labelRGBTrainsetVisualize.tif. Test dataset: The rest area.
| RGB | Category |
|---|---|
| (255, 246, 143) | grass |
| (0, 0, 255) | rice |
| (255, 0, 0) | tree |
| (0, 255, 0) | corn |
| (0, 0, 0) | background |
The settings of this dataset are similar to 2015. See path: data/2017data/ The paper named "Learning discriminative spatiotemporal features for precise crop classification from multi-temporal satellite images", which has been published in International journal of remote sensing.(Ji, S., Zhang, Z., Zhang, C., Wei, S., Lu, M., & Duan, Y. (2020). Learning discriminative spatiotemporal features for precise crop classification from multi-temporal satellite images. International Journal of Remote Sensing, 41(8), 3162-3174.)