Satellite-Based Assessment of Hailstorm Affected Potato Crop for Insurance Purpose

32 Assessing the extent of hailstorm affected crop is one of the thrust areas for quantifying 33 mid-season adversaries under crop insurance values chain. This study evaluated the pre- 34 and post-hailstorm responses on spectral bands and vegetation indices derived from 35 Sentinel-2 data for assessing the severity class of the affected potato crop. The potato 36 crop was mapped using pre-event satellite data with overall accuracy of 88% ( k =0.82). 37 Pair-wise Games-Howell t-test showed significant differences among the post-hailstorm 38 potato severity classes in Red, Near Infrared & Short-wave Infra-red (SWIR) bands and 39 Normalized vegetation indices. Percentage change (from pre- to post-event) in band 40 reflectance and vegetation indices showed a better sensitivity in differentiating damage 41 severity. Differential behaviour of SWIR-1 (Band-11) and SWIR-2 (Band-12) were 42 observed within severely affected potato crop under dry and wet soil conditions. Decision 43 matrix based on percentage change in Normalized difference Vegetation Index ( D NDVI) 44 and Normalized difference Tillage Index ( D NDVI) could able to capture the damage 45 severity classes with an overall accuracy of 86.7%. Higher proportion of affected area 46 were found to be associated with larger percentage of Potato yield reduction based on 47 measured yield data at Insurance unit level. The proposed methodology could be adopted 48 for operational assessment of the impact of hailstorm events on crops. 49


7
"Unaffected" (<20% damage of the canopy foliage), "Moderately-affected" (20-50% damage 157 of the canopy foliage) and "Severely-affected" (> 50% damage of the canopy foliage). The 158 locations of the field data points of the different classes of potato crop affected by the hailstorm 159 are presented in Fig. 2b. Soil moisture variations were also observed in "Severely affected" 160 class as "Dry" and "Wet". Total 54 points were collected over "Unaffected" class, whereas 48 161 and 61 points were collected over "Moderately-affected" and "Severely-affected" categories 162 respectively. Each field data points were converted to a polygon by considering minimum 3x3 163 homogeneous pixels and further used for statistics generation. Out of the total points collected 164 over the study area, nearly 75% of the points were used for developing the methodology and 165 the remaining points were used for validation.

Mapping potato crop 177
Training classes were generated using pre-event ground data (during 14-19 Feb, 2019) and 178 spectral signatures were generated using six bands of sentinel-2 corresponding to 19 th February. 179 Training classes comprised of potato, rice, and other crops (chilli, vegetables and scrubs). The  The band combinations used to generate these indices along with their sensitivity towards 199 different biophysical properties are mentioned in Table-2.  200   Insert Table 2   201 The percentage change in VIs (VI) from pre-to post-event is calculated using equation 2: 202 Where, Yield2019 is GP averaged potato yield (ton ha -1 ) in 2019 and YieldAverage is historical 215 five-year average yield (ton ha -1 ) of GP. Percent yield deviation data were divided into five 216 yield reduction classes and compared with ∆B and ∆VI. 217

Hailstorm and damage to potato crop 219
Hooghly and West Medinipur districts of West Bengal state were exposed to hailstorm during 220 25-28 February, 2019 accompanied with moderate to heavy rainfall causing significant damage 221 to potato crop from falling hails and water stagnation. The daily India Meteorological 222 Department (IMD) gridded rainfall data showed high intensity rainfall over the Hooghly and 223 West Medinipur districts (Fig. 3). The cumulative rainfall between 25-28 February was found 224 to be more than 100 mm in the parts of districts. 225 Insert Figure 3 226 As per the ground truth data collected, there were two prominent standing crops over 227 the two districts i.e., potato and rice. The rice fields were found to be unaffected by the 228 hailstorm-rainfall as they were in the early tillering stage and grown in flooded condition. On 229 the other hand, hails had caused considerable damage to the above ground succulent foliage of 230 the potato crop by breaking/ lacerating it and exposing the below canopy soil. The ridge and 231 furrow structure of the potato field were also disturbed and the potato tubers were exposed 232 partly. The water stagnation due to heavy rainfall further disrupt the soil aeration, causing 233 yellowing of the leaf, rotting of the potato tuber and forced-harvesting in some places. 234 Fig. 4 showed varying degree severity of damage of the potato crop due to the hailstorm 235 event. The unaffected crops were found have high in leaf greenness and leaf moisture, high 236 ground cover (>80%) and less exposure to the soil. Whereas, moderately affected crops were 237 relatively low in leaf greenness and leaf moisture, canopy cover was found to be moderate (50-238 80%). The severely affected crop appeared to be yellowish or dried with less canopy cover 239 (<50%), soil is completely exposed showing the ridge-furrow structure of the potato field.   with the degree of severity. The mean reflectance of NIR band in "unaffected" crop was 30% 263 with standard deviation (SD) of 3%, while for "moderately affected" and "severely affected" 264 crop it was found to be 25±2.5% and 20±4.6% respectively. Marginal response was also 265 observed over B4 (Red), B11 (SWIR1) and B12 (SWIR2) bands. 266 Insert Figure 6 267 To investigate further, four bands (B4, B8, B11 and B12) were selected and the data distribution 268 of these bands during pre-and post-event over the different severity classes were presented in 269 violin-plots (Fig 7). red-reflectance was not found to be significantly different between "moderately affected" and 292 "severely affected" crop. On the other hand, the NIR (B8) region of the spectral band is sensitive 293 to the leaf internal or mesophyll structure. The defoliation causes destruction of the leaf internal 294 structure depending on the severity of the damage. Hence, statistically significant differences 295 were observed between the mean of post-event NIR-reflectance between "unaffected" and 296 "moderately affected", "unaffected" and "severely affected", "moderately affected" and 297 13 "severely affected" crop. The SWIR-1 (B11) and SWIR-2 (B12) bands were sensitive to the 298 surface wetness (Bidgoli et al. 2020). The surface wetness is attributed both by the leaf and soil 299 moisture. The defoliation caused by the hailstorm substantially reduces the leaf wetness, but 300 the associated rainfall led to the increase in soil moisture. Hence, the combined effect of both 301 has been captured by the SWIR bands. Further, SWIR2 band (2.1 µm) is also sensitive to the 302 fractional vegetation cover as it is close to the cellulose absorption band (Quemada and 303 Daughtry 2016). The mean of post-event SWIR1-reflectance showed significant difference 304 between "unaffected" and "severely affected", "moderately affected" and "severely affected" 305 crop. No significant difference of post-event SWIR1-refelctance was observed between 306 "unaffected" and "moderately affected" crop. On the other hand, post-event SWIR2-refletance 307 was found to be significantly different for "unaffected" and "moderately affected" crop only. It 308 is important to mention here that the dispersion of post-event SWIRs reflectance is very high

Response on vegetation indices 320
Converting reflectance of different bands into a normalized index is an effective approach for 321 improving the sensitivity towards assessing the target features (Xue and Su 2017). Hence, we 322 generated four normalized indices i.e. NDVI, NDWI, LSWI and NDTI using the selected four 323 bands as mentioned in the section 3.3. The details of the band combinations are mentioned in 324 Table 2. The variations of the above-mentioned indices during pre-and post-event conditions 325 over the different severity classes are presented in box-plots (Fig 8). It is mentioned in the 326 section 3.3 that all the selected bands (Red, NIR, SWIR-1 and SWIR-2) showed no significant 327 differences between the damage severity classes at pre-event condition (Fig. 7) signifying 328 homogeneous potato crop before the occurrence of the hailstorm. Likewise, the indices derived 329 from these four selected bands did not show any significant differences between the damage 330 severity classes at pre-event condition (Fig. 8). But distinct variations of data distribution of all 331 the four indices over the damage severity classes are observed at post-event condition. As a 332 result, mean of all the four indices showed statistically significant differences between the 333 severity classes during post-event condition (Fig. 8). It is apt to mention here that among all the 334 band-reflectance only NIR (i.e. B8) showed such sensitivity towards separating the severity 335 classes. Hence, there has been a substantial improvement of the sensitivity towards separating   Table 3. The variance analysis of these 344 changes was done and the F-values along with its probability of occurrence by chance are 345 15 mentioned respectively. Post-hoc Games-Howell tests were performed for separation of the 346 mean percentage changes and statistically significant differences between the severity classes 347 are mentioned in Table 3. Out of the four selected band-reflectance, only  ̅̅̅̅̅̅̅ showed high 348 sensitivity and could able to separate different damage severity classes.  ̅̅̅̅̅̅̅ and  ̅̅̅̅̅̅̅̅̅̅ 349 could separate the "unaffected" and "moderately affected" class significantly. Whereas, 350  2 ̅̅̅̅̅̅̅̅̅̅̅ could able to separate "moderately affected" and severely affected" classes.  Table 3 361

6 Variabilities of SWIR-reflectance over the severity classes 362
As mentioned in section 3.3, there were high variabilities/ dispersions of the post-event SWIR-363 reflectance over the "severely affected" potato crop as evident by the shape of the violin plot in 364 Fig. 7. Further, the differential response of  1 ̅̅̅̅̅̅̅̅̅̅̅ and  2 ̅̅̅̅̅̅̅̅̅̅̅ over the different severity 365 classes were observed in Table 3. To explain the high variability of SWIR-reflectance, all the 366 field observations over the "severely affected" potato crop were segregated based on the surface 367 soil wetness condition i.e. "Severely affected (dry soil)" & "Severely affected (wet soil)". 368 Further, all the data points of ΔSWIR1 and ΔSWIR2 over the different damage severity classes 369 were put in a scatterplot and shown in Fig.9. The data points pertain to different severity classes 370 were found to form distinct clusters. As the damage severity increases, the severity-isolines of 371 the clusters (shown as dotted line in Fig. 9) were found to be frame-shifted. The slopes of the 372 isolines remained nearly invariant but the offsets were found to be significantly different. The 373 data point over the "unaffected" potato crop were found to be clustered near to the origin, in 374 the first and second quadrant of the plot within 10 to -10 of ΔSWIR1 and ΔSWIR2. On the 375 other hand, data points over the "moderately affected" crop were found to cluster with -10 to -376 20 of ΔSWIR1 and 10-20 of ΔSWIR2. The data points over the severely affected crop were 377 found to be widely spread over the first, second and the fourth quadrants of the Fig. 9. The data 378 points pertain to "severely affected (wet soil)" were typically found in the fourth quadrant of 379 the plot. Hence, negative values of ΔSWIR2 were found over the "severely affected (wet soil)". 380 As discussed earlier, the hailstorm affected potato crop in two ways. In first case, the 381 aboveground succulent vegetation got damaged by hail without appreciable increase in the 382 background wetness. In the second case, there was appreciable increase in soil wetness in 383 addition to the foliar damage. High soil wetness condition was predominantly observed over 384 the "severely affected" crop and may lead to the tuber rot or force harvesting. These 385 observations of high soil wetness condition were mainly found in the lower ridges of the study 386 area with limited soil drainage condition. The ΔSWIR1 is primarily sensitive to the surface 387 wetness, hence there had been mainly negative changes of ΔSWIR1 due to the hailstorm 388 damage. The ΔSWIR2 is sensitive to fractional vegetation cover (exposed soil surface) and 389 surface wetness as well. In case of "severely affected (dry soil)" categories, there had been 390 significant decrease in the fractional vegetation cover and it exposed of the underlying fine 391 textured dry soil. Thus, it had increased the post-event SWIR2 reflectance and causing positive 392 change in ΔSWIR2. In case of "severely affected (wet soil)" condition the effect of soil wetness 393 on the SWIR2 reflectance superseded the changes (decrease) in fractional vegetation cover. 394 Hence, we found net absorption in SWIR2 and negative change in ΔSWIR2. Such effect is not 395 observed for SWIR1 reflectance as it is primarily sensitive to the surface wetness. This 396 differential behaviour of ΔSWIR1 and ΔSWIR2 in dry fine textured soil is also explained by 397   (Table 3) Table 4  419 Further, scatterplot between ΔNDVI and ΔNDTI over the different damage severity classes is 420 shown in Fig 10

Decision matrix to map the affected area 428
Based on the detailed analysis of pre-and post-event Sentinel-2 data and the observations made 429 thereafter, the following methodology is proposed to assess the potato crop area affected by the 430 hailstorm (Fig. 11). Assessment of hailstorm damage of a crop requires cloud-free pre-event 431 and post-event satellite observations along with field data points of the crop, its stages, growing 432 environment and the intensity of the damage. In the present study, we used 19 th February, 2019 433 (pre-event) and 1 st March, 2019 (post-event) sentinel-2 data to achieve the objectives. The pre-434 event satellite data along with ground truth points were used to map the potato crop and further 435 analysis was done over the potato crop mask only. Two vegetation indices i.e. NDVI and NDTI 436 were derived using relevant band combinations using pre-and post-event observations (Table  437 2 and Fig. 11). The percentage change of these vegetation indices between pre-and post-event 438 i.e. ΔNDVI and ΔNDTI were derived to assess the changes in crop vigour and surface wetness 439 respectively. Based on the response of ΔNDVI and ΔNDTI over the different damage severity 440 classes as mentioned in section 3.7 (Fig. 10), these were sliced into different deviation classes 441 as shown in Fig. 11. These deviation classes of ΔNDVI and ΔNDTI were then combined further 442 19 using decision matrix as mentioned below and also shown in Fig. 12. 443 • If ΔNDVI ≥ -20 and ΔNDTI≥ -20, the potato crop is "unaffected". 444 • If -20 >ΔNDVI ≥ -30 and -20>ΔNDTI≥ -30, the potato crop is "moderately affected". 445 • If ΔNDVI < -30 and ΔNDTI< -30, the potato crop is "severely affected". 446 It is important to mention here that the combination of ΔNDTI < -30% and ΔNDVI > -20% 447 were non-existent in the study area as large change in ΔNDTI is not possible without significant 448 change in vegetation cover i.e. ΔNDVI (Renier et al. 2015). Hence, such categories of classes 449 were not included in the decision matrix. 450 Insert Figure 11 and Figure 12  451 Decision matrix was then implemented over the potato pixels to get the different categories of 452 affected crop over in the study area (Fig 13). Out of the total potato area of 1.21 lakh ha over 453 both the districts combined, nearly 12% of the area was found to be under "severely affected" 454 category and 26% of the area was "moderately affected". The "moderately affected" area was 455 found to have spatial association with the "severely affected area". GP-wise percentage of 456 affected potato area (both severely and moderately) were mapped and presented in Fig.14  Post-hailstorm field observations (not included to generate criteria for decision matrix) were 463 used for accuracy assessment of the affected area map (Table 5). The "unaffected" potato crop 464 was well classified as evident from high producer's (92.7%) and users (90%) accuracy. The 465 20 accuracy was found to decrease slightly for other two classes due to omission / commission 466 errors. The producer accuracies were found to be 75.2% and 88.2 % for "moderately affected" 467 and "severely affected" classes respectively. On the other hand, the user's accuracy of 468 "moderately affected" and "severely affected" classes were found to be 80.1% and 77.3% 469 respectively. The overall accuracy was found to be 86.7 % with kappa coefficient of 0.81. 470 Insert Table 4 471

Hailstorm affected area vis-à-vis potato yield reduction 472
To assess the match between the hailstorm affected area and yield reduction of the potato crop, 473 we calculated the GP-wise yield deviation from normal (Y) using equation 3 as discussed in 474 section 2.5. GP-wise potato yield deviation of the study year (2019) is presented in Fig 15(a). 475 The normal (long-term average) potato yield of the study area (Hooghly and West Medinipur) 476 districts were found to be nearly 20 tones/ha. Large yield deviation was observed due to the Hooghly district were reported large reduction of potato yield. To assess the match between the 481 satellite derived affected potato areas and the reported yield reduction from long term average, 482 the affected (moderate and severely) areas were classified into five classes (≤10%, 10-20%, 20-483 40%, 40-60% and >60%) and the yield reduction at gram panchayat were also made five classes 484 (<20%, 20-40%, 40-60%, 60-80% and >80%). Under each class of the affected area, the 485 distribution of the GPs having different yield reduction classes were presented in Fig 15b. It 486 was observed that the GPs with more than 60% affected area showed >80% or 60-80% yield 487 reduction. The proportion of high yield reduction classes were found to be reduced as the 488 proportion of affected area decrease. The GPs with <10% affected area was found to be 489 21 dominated by the yield reduction class of <20%. In nutshell, the yield reduction of potato crop 490 was corroborating well with the % of damage area at GP level. The result could have been 491 improved further by the well distributed sampling procedure to address the local variations. 492 Insert Figure 15  The NIR-reflectance was found to be highly sensitive to the changes in the canopy structure 503 and surface wetness due to hailstorm. Red and SWIR bands were also showed sensitivity 504 towards it. To accommodate the response of multiple bands towards damage of the crop, four 505 different normalized vegetation indices (NDVI, NDWI, LSWI and NDTI) were derived using 506 combinations of Red, NIR, SWIR1 and SWIR2 bands. All these indices showed high sensitivity 507 and could able to separate different damage severity classes of potato crop. Based on the least 508 co-linearity among these indices, NDVI and NDTI were selected to map the affected area. 509 Decision matrix was prepared using the percentage change (pre-and post-event) of NDVI and 510 NDTI over the different damage severity classes and further used it to map the potato crop area 511 into "unaffected", "moderately affected" and "severely affected" by hailstorm. Overall 512 accuracy of the affected area map was found to be 86.7%. GP-wise yield reduction of potato 513 22 crop based on the CCE data were also found to be corroborating with the % of the area affected 514 due to the hailstorm. Geospatial map of GP level affected potato crop area was also prepared to 515 facilitate informed decision making. The study has thus established as scientific basis to 516 objectively assess potato crop area affected due to hailstorm. Such value-added products would 517 be very helpful in relief management and crop insurance value chain. Future study may be 518 extended towards assessment of quantitative impact of hailstorm on the yield of potato crop. 519 Acknowledgements 520 Authors thankful to Director, NRSC, Hyderabad, India and Deputy Director, Remote Sensing 521 Applications, NRSC for providing constant encouragement and facility to carry out the work. 522

Conflicts of interest: Authors have not reported any conflict of interest. 525
Availability of data and material: The satellite data that support the findings of this study 526 are openly available at https://www.copernicus.eu/en/access-data. Other datasets are provided 527 in the manuscript. 528