Estimation of P deficiency status of rice in initial stage by canopy RGB images using deep learning approach.

Phosphorus (P) deficiency is one of the major yield-limiting factors in rice cultivation. However, determining its symptoms from appearance requires skill and experience. This study aimed to develop a model estimating P deficiency status of rice in early stage using handheld camera for rice canopies with machine learning (ML) approach. Three P fertilizer treatments including those without P application were established with three replications in 306 farmers' fields in the Central Highlands of Madagascar. Canopy RGB images were taken above 1 m from the soil surface at 40 days after transplanting. Then, aboveground biomass was determined by sampling rice plants from the 0.5 m x 0.5 m square area. The aboveground P concentration was determined by dry-ashing and molybden-blue method. The P deficiency index (PDI) was determined by calculating the ratio (R) of the P uptake without P application to that with the highest value among the other P treatments in each field in each replication. Then, PDI (%) was interpreted as (1 - R) *100. After image processing, the RGB values were extracted using a masking method and utilized to compute color indices (CIs) and morphological indices (MIs), such as perimeter area ratio (PAR) and vegetation fraction (VF). ML algorithms (input variables: CIs and MIs) and deep learning (DL) algorithm (input variables: CIs, MIs and RGB image) were used to estimate PDI. PDI ranged from 0.01% to 99.40%, with average of 22.76%. Random forest algorithm showed the highest estimation accuracy with R2 = 0.78 among several ML algorithms, and DL algorithm improved the accuracy with R2 =0.97. Lower VF, higher leaf PAR, and higher dark green color index were suggested to be important indicators for high PDI, that were all relevant to the typical outlooks of P-deficient rice canopy. This approach can be a valuable tool for real-time monitoring and assessing crop nutritional status in diverse environmental conditions. Given the finite nature of phosphorus ore and increasing environmental concerns from excessive P use in the agricultural systems, this simple diagnosis approach should facilitate rice farmers optimizing P use toward sustainable intensification.