Testing the Fluorescence Advantage: Solar-Induced Fluorescence as an Early Indicator of Agricultural Drought Stress in Marathwada, Maharashtra

SAKSHI D. MASKE 

Drought detection in India, both for official declaration purposes and for triggering payouts under the Pradhan Mantri Fasal Bima Yojana (PMFBY) crop-insurance scheme, still leans heavily on rainfall-deficit records and the Normalized Difference Vegetation Index (NDVI) — two signals that only move once a crop has already begun to visibly suffer. This study asks whether Solar-Induced Fluorescence (SIF), a satellite-derived proxy for photosynthetic activity, picks up that stress measurably earlier than NDVI does. Using GOSIF v2 fluorescence data alongside cloud-screened MODIS NDVI over the eight districts of Marathwada, Maharashtra, across eight growing seasons (2015–2023, excluding 2021 for data-availability reasons), I calculated the lag between the two indices' post-peak seasonal decline. A twenty-year (2001–2020) rainfall climatology puts only two of those eight years past this study's own drought threshold — 2015 (−21.5%) and 2018 (−18.3%) — with the remaining six sitting within roughly one standard deviation of normal. Under the threshold-crossing method, SIF's decline led NDVI's in seven of the eight years; the eighth, 2018, came out essentially flat and marginally negative (−1.1 days) — a real exception, not a rounding artifact. H3, which predicted that drought conditions would widen this lag, still doesn't hold at the larger sample: the two drought years averaged a shorter lag (7.6 days) than the six normal years (15.0 days). A second, methodologically distinct cross-correlation lag gives a more mixed picture than the threshold-crossing result: the correlation-maximizing lag comes out clearly positive (SIF leading) in four of the eight years, essentially zero in one, and clearly negative — NDVI leading SIF, the opposite direction — in the remaining three (2018, 2022, 2023). A case-resampling bootstrap (2,000 replicates per year) backs this mixed picture up rather than smoothing it over: two of those three years (2022, 2023) show 99%+ of replicates landing below zero, a strong, well-resolved signal in the NDVI-leads direction, while 2018 leans the same way without fully resolving. The two years with the clearest SIF-leads signal (2015, 2017) turn out to be statistically distinguishable from the two clearest NDVI-leads years (2022, 2023) — a real between-year difference, not noise. At the district level, mean SIF and rainfall anomaly correlate significantly across all 64 district-year observations (8 districts × 8 years): Pearson r = 0.567, Spearman ρ = 0.551, both p < 0.0001 — real and still significant, but considerably weaker than the r = 0.837 the original 24-point, 3-year sample produced, which looks in hindsight like an artifact of a small, drought-heavy sample. Moran's I confirms genuine spatial structure in both variables, though less consistently for SIF (significant in 4 of 8 years) than for rainfall (significant in all 8), and the correlation's effective independent sample size sits closer to the number of years than the number of district-year rows. Checking against the one well-documented official drought declaration available (Maharashtra, 31 October 2018) turned up a reversal from the original 3-year study once the underlying district boundary was corrected: in 2018 specifically, NDVI now crosses its 90% decline threshold roughly five days before SIF does, not after — though both indicators still beat the official declaration by seven to eight weeks. That reversal, together with 2018's near-zero lag under both methods, is reported here as a real finding of the expanded sample.