Early Detection of Broomrape in Tomato Farms Using Satellite Imagery and Time-Series Analysis

Conference: SPIE Defense + Commercial Sensing 2025
Session: Autonomous Air and Ground Sensing Systems for Agricultural Optimization and Phenotyping X

Paper 13475-30 | Date & Time: 15 April 2025, 5:30 PM - 7:00 PM EDT | Sun Room 4, Ballroom Level

Authors: Mohammadreza Narimani (Presenter), Alireza Pourreza, Mohsen B. Mesgaran, Ali Moghimi, Hamid Jafarbiglu

Abstract

Tomato broomrape (Orobanche ramosa) is a chlorophyll-deficient parasitic plant that poses a significant threat to tomato production by extracting essential nutrients from its host, potentially reducing yields by up to 80%. Due to its predominantly subterranean lifecycle and the production of up to 200,000 seeds per plant that can remain viable for up to 20 years, early detection of broomrape infestation is crucial for effective management and yield protection. In this study, we developed an end-to-end pipeline leveraging Sentinel-2 satellite imagery and time-series analysis to differentiate infected from non-infected farms in California. We began by defining regions of interest based on farmer-reported infections and selecting a relevant date range, then filtered satellite images to include only those with less than 10% cloud cover. Twelve spectral bands (ranging from visible to shortwave infrared) were downloaded along with key metadata (sun and sensor zenith/azimuth angles) and stored as GeoTIFF files. Subsequently, we computed 20 vegetation indices—including NDVI, EVI, and NDMI—and derived five neural network–based traits (LAI, Cab, CCC, FAPAR, and FCOVER) using a pretrained model calibrated with ground truth and radiative transfer–synthesized data. By tracking the seasonal trend of canopy chlorophyll content (CCC) and leveraging farmer reports, we isolated the transplanting-to-harvest period for each target farm and aligned phenological stages using growing degree days (GDD) derived from local weather data. Vegetation pixels were segmented by masking non-relevant areas (e.g., soil and buildings) and then used as input for an LSTM model that, with 18,874 pixels over 48 GDD stages, was trained using a 65% training, 15% validation, and 30% test split. After 100 epochs, the model achieved a training accuracy of 88% and test performance of 87% accuracy, 86% precision, 92% recall, and an F1 score of 89%. Permutation feature importance analysis highlighted NDMI, CCC, FAPAR, and CHL-RED-EDGE as key indicators, consistent with literature suggesting that broomrape infestation reduces water and chlorophyll content, thereby affecting overall vegetation health.

This blog will be completed soon...