FSense: Non-Destructive Fruit Moisture Estimation with mmWave Radar using Transfer Learning

Wireless sensing offers a non-destructive method to estimate fruit moisture content, which is essential for assessing internal fruit quality. While fruit quality is often judged by external appearance, this can be misleading, as fruits may look perfect externally but be spoiled internally. Therefore, accurate measurement methods are needed to assess the internal conditions of fruits. Current techniques for measuring fruit moisture, crucial for evaluating quality, are often destructive or expensive, limiting their practicality for everyday use. This paper presents FSense, a novel, non-destructive, and cost-effective system for estimating fruit moisture utilizing millimeter-wave (mmWave) technology. To preprocess radar antenna signals, we introduce a twin-domain model based on the discrete wavelet transform to remove fruit-structure dependencies. Additionally, we propose a Knowledge-Based Transfer Learning Model (KB-TLM) to address variability among fruits with thin and thick pericarps and to leverage transfer learning to apply source-domain knowledge to target domains. This model effectively predicts moisture content across various fruits by employing customized convolutional and generative adversarial networks, which extract stable, moisture-dependent features independent of fruit type. Over one month, extensive experiments were conducted on 240 fruits across four categories, namely apple, pear, peach, and watermelon, in two representative environments, including an empty laboratory and a busy library. Under the original random split setting, FSense achieves an RMSE as low as 0.219. More importantly, under the stricter leave-one-fruit-out evaluation designed to avoid leakage from repeated longitudinal measurements of the same fruit, FSense maintains robust moisture estimation, with RMSE values of 0.276, 0.319, 0.348, and 0.392 for apple, pear, peach, and watermelon, respectively. Additionally, leaving one batch out and leaving one environment out in evaluations further confirm the generalization capability of FSense under fruit identity, batch, and environmental shifts. These results show that commodity 76–81 GHz mmWave radar provides a repeatable moisture-sensitive effective response that can be calibrated for non-destructive fruit moisture estimation, supporting potential applications in fruit preservation, quality monitoring, and post-harvest management.