FULL POLARIMETRIC TIME SERIES IMAGE ANALYSIS FOR CROP TYPE MAPPING

: Crop information and quality are not only fundamental for experts using spatial decision support systems but also have many applications in irrigation management, economic analysis for import or export, food safety, and achieving sustainable agriculture. Remote sensing is a cheap and fast way of reaching this goal. Full polarimetric SAR unlike optical sensors is an all-weather system providing geometrical and physical properties of the earth’s surface events. Due to the dynamic changes in crop properties through their phenological stages, crop type mapping has been challenging. As a result, accurate, reliable, and cost-effective crop type mapping using minimum data and processing has been the goal of the remote sensing and precision agriculture community. In this study, a new method based on time series analysis of full polarimetric SAR data combined with radar indices, polarimetric decompositions followed by the three αs extracted from H/A/α decomposition, and unsupervised H/α/Wishart classification bands as features generated from only 5 dates of RADARSAT CONSTELLATION MISSION 2 data were used for classification of crops. Applying random forest and cat boost algorithm as classifiers an accuracy of 87.4% and 75% was respectively achieved. indicating that both algorithms have promising results. Although the random forest algorithm had better results, the cat boost algorithm had less noise in each field and more homogenous farms were detected.


INTRODUCTION
Due to the rapidly growing acreage of crops, the abruptly increasing population of the world, new global legislations, and the necessity of minimizing the environmental footprint precision agriculturally based actions through remote sensing must be taken to overcome these problems (Jafarbiglu and Pourreza, 2022).One of these actions is crop-type mapping using Optical, SAR, or a fusion of both data sources.Optical data e.g., LIDAR or hyperspectral images impose high acquisition and computational costs (Moradi et al, 2021).Whereas, SAR observations compared to optical sensors are all-weather mapping systems which is the main advantage of SAR images (Yekkehkhany et al, 2014).Making it suitable for Nordic, humid, and mostly cloudy countries e.g., Canada and Greenland.Synthetic aperture radar (SAR) in its different acquisition modes, provides different features regarding the object's physical, geometrical, and structural properties.As a result, it is useful in Remote Sensing and Photogrammetric research including crop type mapping (Adrian et al., 2021;Saadat et al, 2019;Tufail et al, 2022), change detection (Habibollahi, 2022;Saha, 2020), flood monitoring (Carreño Conde and De Mata Muñoz, 2019), Earthquake damage monitoring (Hasanlou, 2021), precision agriculture, oceanography (Zheng, 2019), etc. based on the number of acquired channels SAR images are classified as single, dual, and quad (full) polarizations which contain all the VV and HH known as co-polarizations and VH and HV known as cross-polarizations.As the number of channels rises, so will the number of features and resulting properties, therefore increasing the accuracy.Due to the difference in dielectric properties and the structure of the different crop types, a distinct variation is seen for properties such as the size, shape, and orientation distribution of the scatterers (crops) through the growing season due to the development of crops making it possible to classify different crops in SAR images (Skriver et al, 1999;Yekkehkhany et al, 2014).Therefore, requiring multitemporal image acquisition and time series analysis in important or all phenological stages of different crops.One of the important products derived from the Full Polarimetric SAR observations is called the scattering matrix (T3) which contains a very important source of information about the terrestrial targets (Van Zyl et al, 1990).Thus, used in many physical studies.on the other hand, using the co and crosspolarized bands different indices can be generated through mathematical operations such as Radar Vegetation Index (RVI) (Kim et al, 2011), Radar Forest Degradation Index (RFDI) (Mitchard et al, 2012), Canopy Structure Index (CSI) (Sims and Gamon, 2003), co and cross ratios i.e., HH/VV, VV/VH, and HH/HV, etc.In addition to the mentioned methods for increasing the number of features, it is possible to use different decompositions such as H/ A/  proposed by Cloude and Pottier (Cloude and Pottier, 1997), Pauli, Freeman-Durden (Freeman and Durden, 1993), and Cloude.Adding these features to the scattering matrix increases the possibility of more accurate classification of complex features like crops.
Crop type mapping using SAR images has been assessed in different studies (Yekkehkhany et al, 2014) used JPL-NASA UAVSAR data and SVM with RBF kernel as the classifier, archiving an overall accuracy (OA) of 85.52%.(Mahdianpari et al, 2019) achieved an OA of 84.30% on RADARSAT Constellation Mission (RCM) data on full polarimetric images using a random forest classifier.(Busquier et al, 2019) used TANDEMX data with coherency bands for co and cross-polarized bands at 6 dates achieving an OA of 89.01%, (Gella et al, 2021) used twDTW on longterm Terra SARX data achieving an OA of 80.6%.In this research, we proposed a method based on different indices, and decompositions and added an unsupervised classified map as a new channel to generate reliable features for accurate crop type mapping using only 5 dates of RCM data.Classification is conducted using Random Forest and Cat Boost algorithms.Random Forest is implemented in two stages: 1. learners' creation, and 2. outputs combination (Moradi et al, 2021) and has been used in many agricultural studies e.g., (Sedaghat et al, 2022) used it for surface soil moisture estimation, (Shahrayini and Noroozi 2021) used it for soil salinity and alkalinity mapping and (Shahrayini et al, 2022) used it for multi-depth soil carbon modeling.Cat Boost is an algorithm for gradient boosting on decision trees that were the winner of the Kenya crop type mapping challenge in 2020.Moreover, it has been used in the estimation of aboveground biomass (Luo et al, 2021;Zhang et al, 2020) and fire prediction (Zhou et al, 2021) but not for crop type mapping using polarimetric data.Evaluation is carried out in terms of overall accuracy and kappa coefficient.The classification was done in python using Sklearn, Numpy, Rasterio, and cv2 libraries in python 3.8.8.

MATERIALS AND METHOD
RADARSAT 2 is a C-band Canadian Space Agency Earth observation satellite launched on 14 December 2007.It acquires data in full polarimetric mode and different spatial resolutions for different products.The data used in this paper is for Carman, Manitoba, Canada in 2020 from RADARSAT 2 in 5 dates as shown in Table 1.All images are in simple look complex (SLC) mode with 9 meters ground resolution and 20-kilometer swath width.To classify crops using the proposed supervised method ground truth (GT) data is required, which was acquired from Google Earth Engine's (GEE) AAFC product generated from Sentinel-1, 2, and Landsat missions with 30meter spatial resolution in 2020.However, due to the difference in the resolution of GT and RCM data, resampling was conducted using the bilinear resampling method.
The proposed framework in this paper consists of four main steps: (a) data preprocessing, (b) processing, (c) classification, and (d) evaluation.Several polarimetric features are extracted from preprocessed data i.e., the coherency or covariance matrix (T3).Span (total backscattered power), decompositions including H/A/α, Freeman-Durden, Paulie, and Cloude followed by radar indices including RVI, RFDI, CSI, Biomass index (BMI) also known as average like-polarized and Volume Scattering index (VSI) the is mostly caused by a canopy and defined using average cross-polarized and BMI (Pope et al. 1994).All these bands are extracted and used as features for each time step resulting in 26 features for each time step and 130 total features.Finally, random forest (RF) and cat boost classification algorithms were applied.Their results were compared to find the best model for the classification of different crops using the generated features and RCM time series images.The complete method is shown in Figure 1.Preprocessing includes 1. Calibration, and 2. Applying a Radar speckle filter to reduce noise using a Gamma filter with 9×9 kernel on radar indices.Processing steps include 1. Polarimetric coherency matrix generation, speckle filter to reduce noise using Refined Lee filter 9×9 on T3 matrix, decompositions, and features generation, 2. Geometric correction using range doppler train correction method and SRTM DEM (30m).the complete pre-processing and processing were repeated for each time step in SNAP software.The formulas for generated indices and Span are shown in Table 2. where  ℎℎ is the corrected radar backscatter coefficient.
Some of these features are shown in Figures 2 and 3.   Finally, 60% of these images were divided to train for training the classifiers and the remaining 40% were considered as test data for evaluating the algorithm.To train RF efficiently these settings were applied: 300 trees, 100 maximum features, and a minimum sample split of 2 at each node.As for the cat boost classification algorithm learning rate of 0.1, random strength of 0.1, depth of 12, and a multiclass loss function with 150 iterations were considered.

RESULTS
The proposed method is applied to the extracted multitemporal polarimetric bands.Based on the GT (shown in Figure 5), several types of crop classes are considered including Soybeans, Sunflower, Wheat, Canola, Corn, Rye, Oats, Beans, Broadleaf, Potato, and Peas.Also, classes of mixed wood, barren land, and shrub land were included in the ground truth.Classified maps using RF and cat boost are shown in Figure 8. Validation was carried out using the accuracy metrics of OA and Kappa which are illustrated in Table 3.In this paper, the use of canopy-related radar features, span, and different decompositions derived from a series of RADARSAT2 SLC images in summer 2020 was assessed for crop type mapping.The reviewed papers either used a coarser, longer time series, images with higher spatial resolution e.g., TANDEMX spotlight mode with 1m resolution, or a combination of all the above-mentioned data and reached a maximum accuracy of less than 87%.On the other hand, studies on RADARSAT2 reached a maximum OA of 84.30% which is less than our proposed method which reached an OA of 87%.This method was trained on only 60% of the data and only five dates during the summer.Also, it is not only sensitive to the crop types but also sensitive to other land cover classes such as barren land, urban, mixed wood, shrubland, and field boundaries making it suitable for future purposes of agricultural cadaster, planning, and monitoring.RF is a fast, low memoryconsuming classifier and can be trained on low data unlike recent deep learning approaches that are data hungry, slow and require station systems making them inefficient for Polarimetric studies because these datasets aren't easily or freely available, especially in higher resolutions, longer time series or larger scales with their relative ground truths.In conclusion, this method is accurate, fast, and reliable.
Researchers are suggested to test other canopy-related features and test their impact on classification accuracy.

Figure 1 .
Figure 1.Flowchart of the proposed method.

Figure 7 .
Figure 7. Ground truth with Legend

Table 1 .
Data acquisition time.

Table 2 .
Formulas for generated indices and Span