Using Optimized Three-Band Spectral Indices and a Machine Learning Model to Assess Squash Characteristics under Moisture and Potassium Deficiency Stress

Abstract

Moisture and potassium deficiency are two of the main limiting variables for squash crop performance in many water-stressed places worldwide. If major output decreases are to be avoided, it is critical to detect signs of crop stress as early as possible in the growth cycle. Proximal remote sensing can be a reliable technique for offering a rapid and precise instrument and localized management tool. This study tested the ability of proximal hyperspectral remotely sensed data to predict squash traits in two successive seasons (spring and fall) with varying moisture and potassium rates. Spectral data were collected from drip-irrigated squash that had been treated to varied rates of irrigation and potassium fertilization over both investigated seasons. To forecast potassium-use efficiency (KUE), chlorophyll meter (Chlm), water-use efficiency (WUE), and seed yield (SY) of squash, different commonly used and newly-introduced spectral index values for three bands (3D-SRIs), as well as a Decision Tree (DT) model, were evaluated. The results revealed that the newly constructed three-band SRIs based on the wavelengths of the visible (VIS), near-infrared (NIR), and red-edge regions were sensitive enough to measure the four tested parameters of squash in this study. For instance, NDI558,646,708 presented the highest R2 of 0.75 for KUE, NDI744,746,738 presented the highest R2 of 0.65 for Chlm, and NDI670,628,392 presented the highest R2 of 0.64 for SY of squash. The results further demonstrated that the principal component analysis (PCA) demonstrated the ability to distinguish moisture stress from potassium deficiency stress at the flowering stage onwards. Combining 3D-SRIs, DT-based bands (DT-b), and the aggregate of all spectral characteristics (ASF) with DT models would be an effective strategy for estimating four observed parameters with appropriate accuracy. For example, the model’s approximately 30 spectral characteristics were extremely important for predicting KUE. Its outputs with R2 were, for the training and validation datasets, 0.967 (RMSE = 0.175) and 0.818 (RMSE = 0.284), respectively. For measuring Chlm, the DT-DT-b-20 model demonstrated the best. In the training and validation datasets, the R2 value was 0.993 (RMSE = 0.522) and 0.692 (RMSE = 2.321), respectively. The overall outcomes showed that proximal-reflectance-sensing-based 3D-SRIs and DT models based on 3D-SRIs, DT-b, and ASF could be used to evaluate the four tested parameters of squash under different levels of irrigation regimes and potassium fertilizer.