Interactive comment on “ A conceptual model of organochlorine fate from a combined analysis of spatial and mid / long-term trends of surface and ground water contamination in tropical areas ( FWI ) ” by

Then, we test 3 values of C5bdegrad in a wide range surrounding the one of CLD (Cdegrad): C5bdegrad = 0, C5bdegrad= Cdegrad, C5bdegrad =10 x Cdegrad. Notice these values are highly speculative since there is no experimental value of C5bdegrad and that Cdegrad is the result of optimization process in our paper. Results are reported in the Figure "test" which shows the evolution of the 5bCLD lixiviation and of the ratio for the 3 tested values. The figure 1 shows similar dynamics of ratio evolution or of lixiviation evolution. The difference between the simulations remains weak, notably because the tested values are about 10 and 100 times lower than lixiviation rate (T5bCLD equals 0.1242 here while Cdegrad equals 0.0014)). Consequently, introducing a degradation coefficient does not alter here our first conclusions. Running optimization process with this new term we find T5bCLD = 0.1242 ; Cdegrad = 0.0014 ; C5bdegrad = 0.0010 Our assumptions are consistent with estimations of Dolfing et al. (2012) showing that the solubility is higher for transformation products of CLD.


a, c, e) and 5bCLD / CLD ratio (b, d, f) for surface water (square) and groundwater (star), according to banana cultivated areas and hydrological sectors (a and b), soils (c and d) adapted from Colmet Daage (1965), and geology (e and f) adapted from
. Large squares are relative to sample points having more than ten sampling dates and small squares having fewer than ten sampling dates" 26) Figure 4: y-axis: The numbers are difficult to interpret. Try lg or non-logarithmised numbers instead of ln.
Response: As specified in material and methods, data were log transformed for all analysis. For Figure 4 we worked with log-transformed data. We propose to complete the caption Response: This is now explain in the text accordingly to the response to your comment N°16 Tables: 30) Table 2: Table header needs to be rephrased to improve understandability. Then, it is possible to test 3 values of C 5bdegrad in a wide range surrounding the one of CLD (C degrad ): C 5bdegrad = 0, C 5bdegrad= C degrad , C 5bdegrad =10 x C degrad . Notice these values are highly speculative since there is no experimental C 5bdegrad value and that C degrad is the result of optimization process in our paper. Results are reported in the Figure above which shows the evolution of the 5bCLD lixiviation and of the ratio for the 3 tested values. The figure shows similar dynamics of ratio evolution or of lixiviation evolution. The difference between the simulations remains weak, notably because the tested values are about 10 and 100 times lower than lixiviation rate (T 5bCLD equals 0.1242 here while C degrad equals 0.0014)). Consequently, introducing a degradation coefficient does not alter here our first conclusions. Running optimization process with this new term, we find T 5bCLD = 0.1242, C degrad = 0.0014 and C 5bdegrad = 0.0010. Our assumptions are also consistent with estimations of Dolfing et al. (2012) showing that the solubility is higher for transformation products of CLD.
Change in manuscript: So, to account for the reviewer comment, we propose -1) To complete the current model adding a degradation term for 5bCLD.
-2) Given the lack of knowledge and the uncertainty about degradation rate, we propose to add the following comment in the text L390 "It should be noticed that the degradation values remained uncertain as we did not have any references for comparison. In our case, the optimization process yielded a far lower degradation rate compared to the lixiviation rate (Table 2). Consequently, the model will be less sensitive to changes in the degradation rate than in the lixiviation rate, which is the key parameter for determining the ratio in water. Additionally, there was uncertainty when comparing degradation rates for 5bCLD and CLD. The optimization process yielded degradation rates for 5bCLD and CLD of the same order of magnitude.

Additional simulations showed that setting C5bdegrad ten times higher than Cdegrad instead of zero reduced the 5bCLD / CLD ratio by 10 percent without changing the dynamic of the ratio and of 5bCLD lixiviation (not shown). Given that CLD transformation products are likely to be more mobile in the environment than their parent compound (Dolfing et al. 2012), we assumed that our model gave sufficient bases for interpreting our results.".
Do you have access to CLD/5bCLD ratio in soil to have an idea of the initial signature over time to decipher soil degradation process to those associated to surface and groundwater flowpaths?
Response: Reference of ratio in soils are in the paper of Clostre et al (2015).The median value of 0.011 in nitisols and 0.017 in andosols were used in our paper to constrain our model (see section 3.4). This does not help to speculate about ratios in water since they depend on lixiviation rates of CLD and 5bCLD. In our article, data from Cabidoche et al (2009) were used to assess CLD lixiviation rate (T CLD ) for andosols and nitisols. The 5bCLD lixiviation rate (T 5bCLD ) stemming from the optimization process appears higher than T CLD . This result is consistent with Devault et al (2016) who conclude for a higher mobility for 5bCLD than for CLD. Whatever, it is unlikely that CLD was leached while 5bCLD accumulated in soil profile due to the highest mobility of transformation products (Dolfing et al., 2012).
Change in manuscript: We propose to add the following sentence section 3.4 L406: "… continuously without a plateau.

This result was consistent with Devault et al (2016) who concluded on higher mobility of 5bCLD compared to CLD, and more generally with the results of (Dolfing et al., 2012), who showed that transformation products had higher mobility than CLD."
Second, residence time is used to explain the spatial variability of the ratio compounds/TPs. To support the discussion, the authors should provide existing reported information/simulation of these residence times: -to discuss spatially contrasted compounds/TPs ratio delivery by soil to ground water Response: please see our response to the comment "19)" of the First referee to address the question of degradation of the TP itself (especially for long residence time) Response: longer residence time does not mean that the TP degradation is higher. In fact degradation occurs in the soil, whereas residence time in the aquifer refers to transfers in depth (below soil cover, through the unsaturated and saturated zones) where the degradation (as well as the retention) is considered as null. Groundwater residence time is generally superior to several years (up to several decades -see Gourcy et al., 2009 for instance) that is widely superior to the residence time of the infiltrated water in the soil cover (several days or months).
Third, if the sampling effort, statistical analyses and conceptual development provided a coherent approach for groundwater (slow flowpath), I have many questions on the surface water component.
Response: Global comment about flowpath, as a state of the art for our following responses to the several comments related to that aspect. This global comment justifies our approach and shows in what way the integration of the surface water component will not affect our conclusions. First, volcanic soils in Caribbean islands have a high infiltration capacity (saturated hydraulic conductivity superior to 60 mm/h (Cattan et al., 2006;Crabit et al, 2016). Then, despite high rainfall intensities and amounts, most of rainfall infiltrates (about 95% at the plot scale according to Cabidoche et al, (2009); more than 90% at the watershed scale according to Charlier et al., 2008; generating either subsurface or deep flows. So leaching is the main process in pesticide transport in surface water.
Second, usually, one reason to study separately pesticide transport by surface runoff is that the pesticide concentration in runoff water may vary highly according to time of pesticide application at the plot scale (Saison et al., 2008) as well as at the watershed scale (Charlier et al., 2009). It is not the case for CLD which have been applied long time ago: boundary conditions relative to pesticide concentration in soil are almost steady. Surely, during application period, agricultural practices may have affect 5bCLD/CLD ratio day by day. However our model aims to simulate the ratio evolution over a long time period. A second reason to consider separately runoff and infiltration water is that pesticide concentration in surface water at the plot scale may differ from infiltrated water. There are few references about this point for CLD. Cabidoche et al (2009) notice that CLD concentration in surface runoff was more than 3-fold lower than in drainage, while runoff volume was 10 times lower than drainage volume. They consequently neglected loads in runoff that represented less than 1/30 of those in drainage at the plot scale.
Given the previous consideration, we then choose to focus here on lixiviation process, which affect the ratio dynamic on the long term. The reviewer ask the question of the effect of eventdriven process (storm event, surface runoff, erosion, application practices) on long term trends and how they can modify CLD concentration in water and the ratio. It is a difficult issue that would require getting spatial distribution of storm event, and their contribution to river pollution. This lack of knowledge probably leads to minor CLD exportation. Indeed, most of the time (even in rainy regions), surface flow in the river is driven by baseflow from aquifer's drainage, originated from water infiltration. Knowing that groundwater concentrations are widely higher than in rivers, concentrations during storm events would lead to generate diluted concentrations in surface waters.
We propose different changes relatively to the reviewer comments. We equally propose to add a § "main assumption about CLD transfer" in discussion section Hydric erosion appears to be due only to bad soil management practices, which concentrate runoff that then forms streams that are able to carry aggregates".Thus, erosion from cultivated soils is probably not a major way of CLD transportation. Moreover, given the high contribution of erosion from river beds and from non-contaminated areas in the upstream zone (due mostly to torrential type flow rivers in FWI), the impact of surface water contamination by sediments is considered as a minor process." -For surface water, it could be relevant to know if the CLD concentrations correspond only to the dissolved phase or if it is a "total" concentration.
Response: the CLD concentration is a total concentration. See the change we suggest for the following comment L153 -Information on the filtration and purification steps are not provided in the M&M section.
Response: there was no purification nor filtration since the suspended matter content of samples was low (less than 250 mg L -1 ). Analyses were performed on raw water. We propose to add the following sentences section 2.2.2 Change in manuscript: L153 "Analyses were carried out on raw sampling water. Thus, the water CLD and 5bCLD contents corresponded to dissolved and particulate fractions. It should be noted that the particulate fraction of the samples was low (< 250 mg L-1) due to sampling conducted mainly during periods of low flow." -Can contaminated sediments in river potentially be remobilized by event and alter trend assessment in surface water?
Response: see our previous response -In the conceptual model, the surface runoff and the surface water to groundwater seem not considered. The choice targeting mainly leaching and not the other off-site transport is never discussed. The authors mentioned "hope for pollution mitigation" based on statistical model, but I wonder how fast flow in river can modify this assessment.
Response: see our previous responses In my opinion, the paper can't be published without strengthening of these points.
Specific remarks: L324 GW, as well as in SW fed by it. And vice et versa ?
Response: whereas infiltration from ditches towards aquifers is a likely process in such regions due to the high permeability of the shallow formations (Charlier, 2007), and even if in some cases, river infiltrations may contribute also to groundwater recharge (Charlier et al., 2011), we consider that the infiltration of surface water is a neglecting process of groundwater contamination at a global scale. In fact, in cultivated areas, surface water is generally widely less contaminated in CLD than groundwaters.
L323 The age of the main geological units was used as an indicator of hydrogeology and notably residence time in the aquifers. Could you provide evidence ? Residence time assessment form others studies ?
Response: please see our response to the comment "19)" of the first referee L57 and L60, 1993 or 1992, I guess banned in 1992 but used until 1993. Please explain.
Response: Yes there was exemption until 1993. We propose: Changes in the manuscript: L60 "…ban in 1992 (there was exemption in FWI until 1993)" L121 "they are intergrades" ?
Response: Intergrades are defined by Colmet-Daage relative to the climatic sequence ferralsols -> vertisols for soils that are "intermediate". Since Colmet-Daage classification is specific, we propose to suppress the last part of the sentence which is unclear "and they are intergrades resulting from the alteration of ferralitic soils" L134 unweathered formations, to several decades for old weathered formations (provide range for "Old") Change in manuscript: "between a few years for recent unweathered formations (<0.5-1My), to several decades for old weathered formations (> 1My)" L139 routine basis with CLD. For (double space before For) Response: OK L150 5bCLD is the main CLD co-and alteration product of CLD: what do you mean by alteration product ? Transformation / degradation product ? Please clarify ?
Response: In fact, 5bCLD can be considered both as a co-product and as a degradation product. Consulting biochemists, the word "alteration" seemed more convenient. We propose the following change: Change in manuscript: "5bCLD is the main alteration product of CLD (the term "alteration" here means that 5b is both a co-product and a degradation product) for which …" L151 "Reference standards for CLD and 5bCLD were purchased" : provide purity degree   L 320 "This statistically confirmed"/ Missing word ?
Response and change L320: We propose to rephrase: "This is statistical confirmation of the result mapped in Figure 3…"  Response: baseflow periods refers to periods without flood flows (or storm flows). Please see also our response to your previous comments on sampling L499 "catchment scale", you used watershed during all the manuscript why changed now? "The residence timeestimated by the water apparent age: not discussed or characterized before? Absract. In this study, we investigated the management of long-term environmental pollution by organic 17 pollutants such as organochlorine pesticides. We set out to identify conditions that are conducive to reducing 18 pollution levels for these persistent molecules and then propose a conceptual model of organochlorine fate in 19 water. Our approach looked at spatio-temporal changes in pollutant contents in surface water (SW) and 20 groundwater (GW) on a large scale, in order to decipher the respective roles of soil, geology, hydrology and past 21 treatment practices. The case of chlordecone (CLD) on the island of Martinique (1,100 km²) was selected given 22 the sampling campaigns carried out since 2007 over more than 150 sites. CLD, its metabolite chlordecone-5b-23 hydro (5bCLD) and the metabolite/parent compound ratio were compared. As regards the spatial variability of 24 water contamination, our results showed that banana cropping areas explained the location of contaminated SW 25 and GW, whereas the combination of soil and geology factors explained the main spatial variability in the 26 5bCLD / CLD ratio. For temporal variability, these conditions defined a high diversity of situations in terms of 27 the duration of pollution, highlighting two groups: water draining old geological formations & ferralsols or 28 vertisols vs. recent geology & andosols. A conceptual leaching model provided some key information to help 29 interpret downward trends in CLD and 5bCLD observed in water. Lastly, a conceptual model of organochlorine 30 fate is proposed to explain the diversity of the 5bCLD/CLD ratio in water. Our conclusions highlight the 31 combined role of soil and groundwater residence time for differentiating between conditions that are more 32 conducive, or not, to the disappearance of CLD from the environment. This paper presents a model that provides 33 an overall perception of organochlorine pesticide fate in the environment. To sum up, in FWI human practices and the physical environment lead to high variability conditions for CLD 93 and 5bCLD that may impact the environment. Our aim was to identify the conditions that are conducive to a 94 decrease in pollution levels, in order to propose a conceptual model of organochlorine fate in water. We focus 95 here on river contamination, which is driven by all the environmental compartments, thus being an integrative 96 survey site for land-use, soil variability, and aquifer contributions. Based on the sampling campaigns in 97 Martinique (FWI) since 2007, we explored river contamination trends over time and the relationships between 98 surface and underground CLD rates in water. Spatial and temporal distributions of contamination were 99 interpreted according to soil and geology mapping, hydrology and past CLD treatment practices. This work will 100 lead on to identifying areas with a low or high impact on water pollution, in order to manage polluted areas more 101 effectively. 102 2 Material and methods 103

Study site 104
Location and climate. The study area covered the volcanic island of Martinique (1,100 km²) in the French West 105 Indies in the Caribbean (Figure 1). The climate is tropical, hot and humid. Annual rainfall is almost a linear 106 function of altitude (0 to 1,500 m ASL) and ranges from 2,500 to 10,000 mm on the east coast, and 1,000 to 107 10,000 mm on the west coast. Data selection for statistical models. For statistical analysis, we discarded data where CLD concentrations were 233 below detection limits (and consequently 5bCLD concentrations too, as 5bCLD concentrations are always lower 234 than CLD), as they would have led to an inappropriate ratio value (ratio of 1 according to the value assessment 235 Supprimé: measurable 7 rule described below). Additionally, although we gathered data from contaminated areas, some of the water 237 samples were contaminated with CLD, but no 5bCLD was detectable. For the statistical analysis, we kept all the 238 data (with and without quantifiable 5bCLD) from sampling points for which at least half the samples had 239 quantifiable 5bCLD contents (≥ 0.03 or 0.01 μg L −1 ). This avoided overestimating the concentration for the 240 sampling point, which would have been the case if we had discarded all the data with no quantifiable 5bCLD. 241 For SW, we selected 963 data items (i.e. water samples analysed for CLD and 5bCLD). This SW data set 242 covered 38 sampling points out of a total of 136. For GW, we selected 123 data items. This GW data set came 243 from 7 sampling points. 244 Data selection for temporal analysis on specific rivers. In order to highlight differences between pesticide 245 trends depending on the sampling point, we chose rivers for which the analysis covered the entire 2009-2014 246 period. This led to the selection of 14 sampling points, all having more than 50 analyses. As stated above, we 247 discarded analyses where CLD and 5bCLD contents were below detection limits. 248

Statistical analysis 249
Models. To ensure that the residue distribution of the analysis of variance (ANOVA) model followed the 250 assumptions of equal variance and normality, we used log transformed (natural log) data. We analysed our SW 251 and GW data sets by a multi-way analysis of variance using the MIXED procedure in SAS software (SAS 252 Institute Inc, 2002). The effects to be taken into account in the models were chosen by comparison of the AIC 253 (Akaike Information Criterion). 254 Model 1 was used on the SW data set to test different effects on the CLD content, the 5bCLD content and the 255 ratio of the 5bCLD content to the CLD content in SW. The soil and geology factors were dependent on each 256 other. For this reason, only combinations of these 2 factors were considered in the model. 257 Eq. (1) = µ + + + + + 258 where Y ijklm is the observation (i.e. ln(5bCLD), ln(CLD) or ln(5bCLD /CLD) ), µ is the general mean, α i is the 259 (soil x geology) type effect, β ij the hydrological sector effect for each (soil x geology) type, γ t is the date effect, 260 D ijk the random effect of the sampling point for each (soil x geology) type and ε ijtkl is the residual error. Indices i, 261 j, t, k, l represent following factors for soil x geology, hydrological sector, date, sampling point and sample 262 replication, respectively. 263 Model 2 was used on the GW data set. Soil and geological factors were closely linked for the GW data set 264 (andosols were always associated with recent geological formations and ferralsols with old geological 265 formations), making it impossible to distinguish the soil effect from geology; likewise for groundwater basins 266 and hydrographic sectors. Consequently, only soil and hydrogaphic sectors were tested for model 2: 267 Eq. (2) ′ = µ′ + ′ + ′ + ′ + ′ + ′ 268 where Y' ijklm is the observation (i.e. ln(5bCLD), ln(CLD) or ln(5bCLD /CLD) ), µ' is the general mean, α' i is the 269 soil type effect, β' ij the hydrological sector effect for each soil, γ' t the date effect, D' ijk the random effect of the 270 sampling point for each soil and ε' ijtkl is the residual error. Indices i, j, t, k, l represent following factors for soil, 271 hydrological sector, date, sampling point and sample replication, respectively 272 where Koc (L kg -1 ) is the partitioning coefficient between the sorbed part on soil organic matter and the 306 dissolved part in water, D (kg dm -3 ) the bulk density, C (g kg -1 ) the soil carbon content, R (dm) the annual 307 amount of rainfall, S the soil surface area (dm²) and d the soil depth (dm). 308 The calculation steps are given below:   Figure 2 shows the relationship between the means of 5bCLD and CLD in rivers at each sampling point. We 330 found that the water 5bCLD content was at least tenfold lower than the water CLD content. However, there was 331 not a unique relationship between 5bCLD and CLD. The frequency distribution of the means of the 5bCLD to 332 CLD ratio in SW and GW clearly showed that a threshold of 0.07 divided the data set into two groups: a low and 333 a high ratio around 0.02 and 0.1, respectively. According to Devault et al. (Devault et al., 2016), these 334 differences cannot stem from the use of different commercial products or different batches of the same product. 335 Indeed, these authors, found no significant statistical difference between the ratio of the commercial products 336 Kepone® and Curlone® used in FWI, no more than they did between samples from different batches of 337 Curlone®. They found a mean ratio in commercial products of 0.00077 ± 0.00027, i.e. ten times lower than our 338 observations in river. 339  Figure 3 shows that the group with the high ratio (>0.07) was mainly located either in the highly 346 contaminated northern areas, or in some parts of the low-contamination areas in southern and western 347

Martinique. 348
We observed overall consistency between the distribution of SW and GW contamination: the higher the CLD 349 content or 5bCLD / CLD ratio for SW, the higher the CLD content or 5bCLD / CLD ratio for GW. However, the 350 west coast displayed some exceptions, since we observed contaminated GW (primarily low contamination) while 351 CLD was not detected in the rivers in the neighbourhood. Similarly, the 5bCLD / CLD ratio for GW belonged to 352 the high value group (>0.07 µg L -1 ), while the 5bCLD / CLD ratio for SW belonged to the low value group, or 353 was not available because of no contamination. 354

Land-use practices: high level of contamination in historical banana areas 356
Globally, for the water CLD content, the SW and GW contamination sites matched with the historical banana 357 areas since 1970, i.e. during CLD application. Surprisingly, SW and GW contamination occurred outside these 358 banana areas. This was mostly with low concentrations under 0.1 µg l -1 and rarely with the higher levels (one 359 point in the South-West for GW, far from the banana area,). Most of these isolated points had a high 5bCLD / 360 CLD ratio, leading the 5bCLD / CLD ratio not to match banana field distribution, suggesting past CLD misuse. 361

Hydrographic sector: a functional relationship between measurement points 362
Introducing hydrographic subsectors made it possible to establish a functional relationship between measurement 363 point data. Notably, this helped to explain why some points close to each other did not have the same 364 contamination level. For example, although sample points of subsector 1 and 2 were very close (see Figure 3a), 365 they did not have the same contamination level. In contrast, all the sample points of subsector 1 had the same 366 contamination level (same for subsector 2). This suggests that the hydrographic sector, i.e. the water flows 367 within the same hydrological unit, mainly determined the contamination level of the sample points, rather than 368 the geographical closeness of those points However, some differences were found on the north-east coast. This 369 was encountered in zone 3, where the contamination levels seemed to be linked to the altitudinal gradient. 370 Contamination increased downwards in coherence with banana field distribution along the coast at the lowest 371 altitudes. The statistical results summarized in Table 1 confirm this interaction between hydrographic sectors and 372 soil /geology for CLD and GW. However, no effect was found for the 5bCLD content and the 5bCLD / CLD 373 ratio. 374

Soil type: a factor explaining some ratio variations in SW 375
Table 1 shows significant differences in GW CLD contamination according to the soil/geology pair: GW on 376 nitisols, which are associated with old formations (older than 1 My), was more contaminated than on andosols 377 associated with recent formations (1My to present). This did not result in any significant difference for SW. 378 However, for SW, we observed significant differences for the 5bCLD / CLD ratio, opposing a low ratio for 379 nitisols to a higher ratio for andosols (Figure 4). We also noted a higher ratio for vertisols. This is statistical 380 confirmation of the result mapped in Figure 3, showing high 5bCLD / CLD ratios on vertisols in southern 381

Geology: a factor explaining ratio variations in SW and GW 383
The age of the main geological units was used as an indicator of hydrogeology, and notably residence time in the 384 aquifers, which is linked to pesticide transfer kinetics in GW, as well as in SW fed by it. Thereby, shorter 385 residence times were observed for aquifers located in more recent and unweathered geological formations. It can 386 be seen in Figure 3 that the highest CLD contents in water matched with recent geological formations in the 387 banana cropping area (northern half of the island). Medium and low CLD contents were observed in other older 388 geological units, or outside banana cropping areas. As regards the 5bCLD / CLD ratio, the highest values were 389 only observed in the most recent units (0.5 My to present), for the most contaminated water bodies in the North 390 Atlantic area (not shown). 391 It is interesting to note that the soil effect depended on geology. Figure 4 illustrates this, presenting the mean 392 ratio for each soil type according to the age of the geological formations. For andosols and ferralsols/andosols, 393 the ratio appeared to be significantly higher for recent geology.  groundwater data sets showed an interesting evolution pattern with, in some cases, a decrease in CLD content 417 associated with an increase in water 5bCLD content. 418

In SW: the pesticide concentration and ratio globally decreased 419
From all the available data, we observed a highly significant downward trend in mean river concentrations for 420 the CLD content, 5bCLD content and the 5bCLD / CLD ratio in water (a slope of -0.008, -0.028, -0.018, 421 respectively). It is interesting to note that the decreasing trend for the 5bCLD content was about threefold higher 422 than for the water CLD content. 423 More specifically, Figure 6 shows the evolution of water CLD content for the 14 rivers with the highest 424 measurement frequency. Globally, the mean Sen trend was -0.008 for the log, meaning that the CLD content was 425 halved after 7.5 years. Although most of the rivers showed a significant decrease in water CLD content, some of 426 them were characterized by a constant level of contamination (Saint Pierre, Pont RN Rouge) and even one by a 427 slight increase (Camping Matouba). Independently, we noted a high variation in the level of contamination. 428 A further analysis of temporal evolution (Figure 7) highlighted a relationship between Sen trends for CLD and 429 the mean water 5bCLD contents (regression p-value =0.06): the lower the water 5bCLD content, the greater the 430 decrease in water CLD content. A similar trend was observed for the 5bCLD / CLD ratio (regression p-431 value=0.05), while the relationship was not significant for mean water CLD content. This indicated that the 432 decrease intensity did not depend on water CLD content. Additionally, Figure 7b shows that the smallest decreases in water CLD content were partly associated with 437 recent (0.1 My to present) geological formations and that the largest decreases were associated with older ones. Lastly, regarding soils, Figure 7a shows that while andosols were distributed over the entire range of Sen trends, 442 ferralsols and vertisols characterized large decreases in water CLD content. 443 To sum up, high water CLD contents decreased with low water 5bCLD contents and low 5bCLD / CLD ratios 444 were encountered for basins situated on old geological formations and mostly ferralsols or vertisols. On andosols 445 and recent geological formations, the water CLD content did not vary over the study period, and the water 446 5bCLD content and 5bCLD / CLD ratio were high. These conditions define a high diversity of situations with 447 regard to the persistence of pollution. 448

Model simulation 449
In order to grasp the complex fate of CLD and 5bCLD, we used the simple model presented in Sect. 2.3.4. It is 450 an iterative leaching model investigating the theoretical fate of CLD and 5bCLD in water, accounting for 451 CLD and 5bCLD lixiviation rates (T CLD and T 5bCLD ), as well as the rate of CLD degradation into 5bCLD (C degrad ). 452 Table 2 gives the results of the optimization processes in order to assess T 5bCLD , C degrad and C 5bdegrad from realistic 453 values of T CLD and the 5bCLD / CLD ratios. Thus, according to Eq. 6 (see Sect. g kg -1 for soil carbon content C, and 4,000 and 2,000 mm for annual rainfall R. We targeted the 5bCLD / CLD 457 ratios of 0.1 and 0.025 in water (cases And1, Nit1 and And2, Nit2, respectively), which corresponded to the 458 median 5bCLD / CLD ratios of SW for the two groups identified in Sect. 3.1. We applied a constraint on the 459 5bCLD / CLD ratios in soil, considering that the ratios should lie between 0.01 and 0.017, referring to the 460 median value encountered for andosols and nitisols, respectively (Clostre et al., 2015). 461 It should be noticed that the degradation values remained uncertain as we did not have any references for 462 comparison. In our case, the optimization process yielded a far lower degradation rate compared to the 463 lixiviation rate (Table 2). Consequently, the model will be less sensitive to changes in the degradation rate than 464 in the lixiviation rate, which is the key parameter for determining the ratio in water. Additionally, there was 465 uncertainty when comparing degradation rates for 5bCLD and CLD. The optimization process yielded 466 degradation rates for 5bCLD and CLD of the same order of magnitude. Additional simulations showed that 467 setting C 5bdegrad ten times higher than C degrad instead of zero reduced the 5bCLD / CLD ratio by 10 percent 468 without changing the dynamic of the ratio and of 5bCLD lixiviation (not shown). Given that CLD transformation 469 products are likely to be more mobile in the environment than their parent compound (Dolfing et al. 2012), we 470 assumed that our model gave sufficient bases for interpreting our results. 471 Figure 8 shows the results of two simulations: simulation And2 corresponds to an andosol situation with high 472 soil retention, and simulation Nit1 to a nitisol situation with low soil retention (Table 2). It should be noted that, 473 according to Eq. (3) and (4), Figure 8 shows the leached quantities of CLD and 5bCLD, not the concentration. 474 However, as the two compounds were lixiviated with the same quantities of water, the shape of the concentration 475 curve and quantity curve did not differ. 476 The simulation results showed that the ratio increased with time over the entire period up to a plateau (see Figure  477 8). A decrease in the ratio was not simulated, although a global trend was noted for our observed data on the 478 these dynamics (linear decrease in log scale, Figure 6). Interestingly, we found that the decrease rate for andosols 487 (simulation And2 - Figure 8) was far lower than for nitisols (simulation Nit1). This matched the andosol 488 situation, where no significant decrease in the river was observed. 489 5bCLD first increased and then decreased at the same time as CLD. This may explain why we found a 5b CLD / 490 CLD ratio increase, whereas a 5bCLD decrease was observed. Our simulations also showed that T 5bCLD must be 491 higher than T CLD otherwise the ratio increased continuously without a plateau. This result was consistent with 492 Devault et al (2016) who concluded on higher mobility of 5bCLD compared to CLD, and more generally with 493 the results of (Dolfing et al., 2012), who showed that transformation products had higher mobility than CLD. 494 Optimization processes also gave a higher value for T 5bCLD (Table 2), given that high ratios are unlikely when 495 T CLD is high (0.15) since it yields a T 5bCLD of 1 (meaning that all 5bCLD is leached). Our results showed high spatial and temporal variability for water CLD content in SW and GW contamination. 501 By relating water CLD content to its metabolite compound, 5bCLD, we highlighted physical conditions relative 502 to soils and geology that may explain its variability in water, but also in the dynamics of pollution trends. We 503 summarized our conclusions in a conceptual scheme presented below. But first, let us specify the interpretation 504 framework. 505

Main assumptions about CLD transfer 506
In our study, we focused on long-term trends for CLD and 5bCLD concentration in water, along with their ratio. 507 We considered that the main process determining pollutant concentrations in water was relative to CLD 508 desorption by water infiltrating the soil. We assumed this hypothesis for different reasons. 509 Firstly, rain water mainly infiltrates. In fact, given the high soil infiltration rate (saturated hydraulic conductivity which prevents clay dispersion and sheet erosion. Hydric erosion appears to be due only to bad soil management 518 practices, which concentrate runoff that then forms streams that are able to carry aggregates". Thus, erosion from 519 cultivated soils is probably not a major way of CLD transportation. Moreover, given the high contribution of 520 erosion from river beds and from non-contaminated areas in the upstream zone (due mostly to torrential type 521 flow of rivers in FWI), the impact of surface water contamination by sediments was considered as a minor 522 process. 523 Lastly, by neglecting transport via surface runoff (since sampling mainly occurred outside storm event periods), 524 we probably underestimated pollutant exportation. Thus, we expected that it should not have a great impact on 525 the long-term dynamics of concentrations and ratios in rivers, which is one of the main topics of our paper. with different CLD contents may release the same quantity of CLD into water. However, our simulations 536 showed (see Figure 8) that over a long time scale, CLD contents in a river will quickly decrease for basins 537 draining soils such as nitisols, due to their low capacity to retain CLD. 538 In this environment, our results were in line with CLD degradation, being visible over a decadal time period 539 despite its strong persistence in the environment. This was hypothesized by observing the distribution of 5bCLD 540 / CLD ratios in water (median of 0.03; 1 st centile of 0.006) with a far higher median and first centile value than 541 in the commercial products Kepone® and Curlone® used in FWI (mean ratio of 0.00077 ± 0.00027, (Devault et 542 al., 2016)). This was consistent with the result obtained by Devault et al. (2016), who found high 5bCLD / CLD 543 ratios in soils and, in particular, larger amounts of 5bCLD than should have been applied using commercial 544 formulations. 545 The water CLD content in SW decreased as well as the water 5bCLD content and the 5bCLD / CLD ratio. Given

Hypothesis relative to leaching processes 557
One of the main questions in this paper was what the 5bCLD / CLD ratio represents. To answer this sensitive 558 issue, we differentiated between three dimensions. A temporal dimension, because the 5bCLD / CLD ratio is 559 assumed to increase over time as degradation progresses. A spatial dimension, since the 5bCLD / CLD ratio may 560 depend on local degradation conditions. A dynamic dimension, since the 5bCLD / CLD ratio may depend on the 561 mobility properties of both molecules, CLD and 5bCLD. 562 The temporal dimension was firstly related to the long application period (from 1970 to 1993 for CLD), given 563 that land-use changes led to different application phases in the 70s and 80s and that land-use changes are 564 correlated with soil contamination levels (Desprats et al., 2004). Secondly, comparing simulation results to 565 measurement time series, the temporal dimension could also be grasped by observing GW, if we consider that 566 the residence time within the aquifer gives a temporal window on the water infiltration conditions (Gourcy et al., 567 2009;Tesoriero et al., 2007). The residence time -estimated by the water apparent age -depends on 568 hydrogeological properties, and thus to the geological context (type of lithology and its weathering level, 569 geometry of the geological deposits, etc.). For example, we observed that high 5bCLD / CLD ratios were mainly 570 located in the waters of northern Martinique, where rivers drain recent geological formations. In that area, 571 unweathered formations favour rapid transfers and thus low GW residence times of several years (Arnaud et al.,572 2017; Gourcy et al., 2009). Thus, in that area, GW is young and probably today mainly composed of waters that 573 percolated in the last decade with a 5bCLD / CLD ratio close to the current 5bCLD / CLD ratio in soil leaching 574 waters. Conversely, the higher groundwater residence times in more weathered geological formations probably 575 characterize older GW (residence time of several decades) where the 5bCLD / CLD ratio may reflect an earlier 576 5bCLD / CLD ratio in soil leaching waters -closer to the ratio in the commercial product -during periods of 577 application or just several years after, leading to lower 5bCLD / CLD ratios in water. 578 The spatial dimension is hard to grasp since some of the variability can be attributed to the spatio-temporal 579 variability of land-use changes over the application period. Considering that soil might be an important factor, 580 the results from Clostre et al. (2015) show that the distribution of the 5bCLD / CLD ratio differs little from one 581 soil to another, with a median value of around 0.011 [0.002 0.077] in andosols and 0.017 [0.007 0.081] in 582 nitisols. This does not mean that degradation does not depend on soil, but it does mean that we cannot assess the 583 effect of soil on degradation. It is interesting to note that the simulations accounting for nitisols and andosols in 584 Table 2 give close values of 0.14% and 0.16% for the degradation rate, respectively. The soil factor could 585 therefore not be considered decisive in explaining spatial degradation intensity. 586 For the dynamic dimension, our theoretical leaching model helped to represent how contamination evolved. On 587 the whole, the simulations accounting roughly for andosol and nitisol conditions tallied well with our 588 observations or with results from the literature: i) a large decrease in CLD was associated with a low 5bCLD / 589 CLD ratio, and ii) nitisol situations were more conducive to a contamination decrease than andosol situations, 590 considering pollution duration as noted by Cabidoche (Cabidoche et al., 2009). 591 Lastly, this discussion shows that the combined role of geology and soils together may explain 5bCLD / CLD 592 ratio levels. In a comprehensive way, we derived a conceptual scheme of water contamination on a regional 593 scale. 594 16 4.4 A conceptual scheme of water contamination on a regional scale 596 We propose a conceptual scheme in Figure 9 to explain differences in 5bCLD / CLD ratios in water. We first 597 assumed that degradation occurs in soils. This process, which is combined with other processes determining 598 CLD and 5bCLD fate in soil, results in a general increase in water 5bCLD content and in the 5bCLD/CLD ratio, 599 which is more or less pronounced depending on the soil. Hydrogeology teaches us that SW today could either be 600 a signal of ancient infiltrations and transfers underground, several decades ago, when 5bCLD/CLD ratios in soils 601 were low (long residence time), or a signal of recent percolations, several years ago, when 5bCLD/CLD ratios in 602 soils were high (short residence time). Thus, soil properties and residence times both contribute to explaining the 603 current impact on water quality in SW. This explanation is consistent with high 5bCLD/CLD ratios in northern 604 Martinique on recent geological formations, and low 5bCLD/CLD ratios elsewhere. For high 5bCLD/CLD ratios 605 in the South on vertisols, we can speculate that the degradation process was greater in this soil type (like soil 2 in 606 Figure 9) because lixiviation is lower in the southern area with a lower rainfall rate. This may explain the higher 607 5bCLD/CLD ratios in SW, as simulated by a previous model, despite a longer residence time in the aquifers. 608 All of these results identify a set of conditions that favour the disappearance of CLD from the environment, 609 namely ferralsols with low retention properties on older geological formations, while others -notably andosols 610 with high retention rates on recent formations -are more risky. 611

Conclusion 612
The aim of this paper was to identify conditions that are conducive to a decrease in organochlorine pollution 613 levels in Martinique (FWI). We adopted an unusual approach that accounted, on the one hand, for the 614 interactions between aquifers and rivers on a watershed scale and, on the other hand, for the fate of CLD and its 615 compound 5bCLD. This approach was fruitful and led to the proposal of a global scheme of water contamination 616 on a regional scale accounting for physical conditions relative to soils and geology. This scheme coherently links 617 the various amounts of chlordecone (CLD) and its metabolite 5bCLD in SW and GW. It explains their variability 618 in water, but also in the dynamics of pollution trends. 619 Our results have several implications for evaluating diffuse pollution of agricultural origin. The spatial analysis 620 of metabolite/parent compounds provided some interesting information for identifying risky areas, or areas 621 where persistent pollutants are more likely decreasing. This also provided some insights into key parameters that 622 control these conditions and environmental vulnerability to agricultural pollution. It led to implications regarding 623 where and how to act to reduce impacts (e.g. choice of crops according to pollution levels, since some plants are 624 less sensitive to contamination than others (Clostre et al., 2015), constraints on water management, such as 625 drinking water and irrigation, choice of priority areas to test decontamination processes, setting up compensation 626 plans according to the risk, etc.). Another implication is to promote continuous long-term observations as 627 opposed to one-off sampling, completing modelling approaches: in our case, long CLD time series revealed a 628 faster decrease than that expected by previous model predictions. Lastly, such a spatial and temporal overview is 629 required on a large scale to help stakeholders manage pollution on a territory scale, accounting for the main 630         Large squares are relative to sample points having more than ten sampling dates and small squares having fewer than ten sampling dates.  Figure 9: CLD fate in soils and residence time combined to explain 5bCLD/CLD ratio levels in SW. For SW draining GW with a long residence time, leaching occurred during the application period with a low 5bCLD/CLD ratio whatever the soil type. For SW water draining GW with a short residence time, leaching occurs nowadays from soil with a higher 5bCLD/CLD ratio depending on soils and reflecting CLD fate in soils. 844 845 846