Farmers are exposed to numerous stressors that can negatively impact their mental health, leading to conditions such as depression. However, most studies examining depression risk in farmers are limited by small sample sizes, narrow geographic coverage, and a focus predominantly on male farmers and general agricultural contexts. To complement these traditional studies, big data and machine learning (ML) can advantageously be harnessed. While ML algorithms have shown high accuracy in identifying depression predictors in mental health research, no study has yet applied ML in farmers. We aimed to identify key predictors of depression among the entire French farmer workforce across professional categories, activities, and sexes using ML (XGBoost). A secondary analysis of large-scale administrative health databases (TRACTOR project) was conducted. Potential predictors (n = 128 for farm managers and 123 for farmworkers) included a broad range of sociodemographic, health, lifestyle, and occupational variables. The predictor's importance was determined using Shapley's additive explanation. There were 83,592 depression cases among 1,088,561 farm managers and 149,285 depression cases among 5,831,302 farmworkers. Models performed well, with F1 scores ranging from 0.65 to 0.94. We noted differences, even though several predictors were common across populations, activities, and/or sexes. The top predictors of depression included working year, age, sex, experience, job security, income, and preexisting health conditions. The working year, which reflects the cumulative impact of external factors (e.g., harsh weather) on farmers' mental health, emerged as the most important predictor. These findings highlight the potential of ML applied to realworld data for identifying modifiable predictors, thus enhancing early detection and prevention strategies. By differentiating predictors across farming groups, our results suggest that tailored mental health interventions could be developed to better address the unique needs of various farming populations. These insights could inform the development of clinical tools (e.g., risk calculators) to assist in clinical decision-making.