Repository associated with the paper: Multi-Temporal Polarimetric SAR (MT-PolSAR) Change Detection For Crop Monitoring And Crop Type Classification
A detailed description to reproduce the work presented in the paper can be found here. In summary you can reproduce the results of the paper running the scripts in the following order:
- Python_scripts/B1_array2D_of_coherency_matrices.py,
- Python_scripts/E_MT_Datacube.py,
- Python_scripts/H2a_Change_mat_datacube_other_viz_abs.py,
- Python_scripts/I4_master.py,
- Colab_1: AgriSAR_train_NNs.ipynb,
- Colab_2: AgriSAR_test_NNs.ipynb,
- Python_scripts/M_Scale_CM_cube (figs 10&11).py,
- Colab_3: AgriSAR_Prediction_maps.ipynb.
The interpretation of multidimensional Synthetic Aperture Radar (SAR) data often requires expert knowledge for simultaneous consideration of several time series of polarimetric features to visualise and understand the physical changes of a target and the temporal evolution between the SAR signal and a target on the ground. Multitemporal polarimetric SAR (MTPolSAR) change detection has been introduced in the literature in [1] and [2] in an effort to solve this by characterising the changes over time. However, the obtained results either only exploit intensity of changes or the resulting changed scattering mechanisms are not guaranteed to represent physical changes of the target. This paper presents a variation in the change detector used in [2] based on the difference of covariance matrices that characterise the polarimetric information of a resolution cell, allowing for an intuitive representation and characterisation of physical changes of a target and its dynamics. We show the results of this method for monitoring growth stages of rice crops and we present a novel application of the method in which the capabilities for image classification are investigated applying it to crop type mapping from MT-PolSAR data. We compare its performance with a neural network-based classifier that uses time series of PolSAR features derived from target covariance matrix decomposition as input. Experimental results show that the classification performance of the method presented here and the baseline are comparable, with differences between the two methods in the overall balanced accuracy and the F1-macro metrics of around 2% and 3%, respectively. The method presented here achieves similar classification performances than a traditional PolSAR data classifier while providing additional advantages in terms of interpretability and insights about the physical changes of a target over time.
Typical rice change matrix. Left: Change matrix. Top right: Main rice growth stages. Bottom right: RGB
interpretation of added and removed scattering mechanisms. The added and removed SMs between two stages correspond to
their intersecting squares in the upper and lower triangular part, respectively
